[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"portal-settings:figure:zh":3,"public-menus:all":45,"post:render-less-infer-more-is-ai-breaking-the-traditional-gaming-hardware-upgrade-cycle:zh":531,"related:post:render-less-infer-more-is-ai-breaking-the-traditional-gaming-hardware-upgrade-cycle:zh:1":3135},{"statusCode":4,"data":5,"message":44},200,{"tenantId":6,"lang":7,"defaultLang":8,"siteUrl":9,"contactEmail":10,"brandName":11,"logoUrl":12,"siteName":11,"siteDescription":13,"ogImage":14,"robotsIndex":15,"socialLinks":14,"reservedSlugs":14,"seoPolicy":16},"figure","zh","en","https:\u002F\u002Ffigure.rocks","info@stajic.de","Figure Rocks","\u002Ffavicon-32x32.png","",null,true,{"branding":17,"relatedContent":18,"crossDomainLinks":19},{"logoUrl":12},{"enabled":15},[20,23,26,29,32,35,38,41],{"url":21,"label":22,"isActive":15,"showInFooter":15,"includeInSameAs":15},"https:\u002F\u002Floving.rocks","loving.rocks",{"url":24,"label":25,"isActive":15,"showInFooter":15,"includeInSameAs":15},"https:\u002F\u002Fstajic.de","stajic.de",{"url":27,"label":28,"isActive":15,"showInFooter":15,"includeInSameAs":15},"https:\u002F\u002Fbazify.com","bazify.com",{"url":30,"label":31,"isActive":15,"showInFooter":15,"includeInSameAs":15},"https:\u002F\u002Fbazify.de","bazify.de",{"url":33,"label":34,"isActive":15,"showInFooter":15,"includeInSameAs":15},"https:\u002F\u002F2mesta.com","2mesta.com",{"url":36,"label":37,"isActive":15,"showInFooter":15,"includeInSameAs":15},"https:\u002F\u002F2mesta.de","2mesta.de",{"url":39,"label":40,"isActive":15,"showInFooter":15,"includeInSameAs":15},"https:\u002F\u002Fbazify.at","bazify.at",{"url":42,"label":43,"isActive":15,"showInFooter":15,"includeInSameAs":15},"https:\u002F\u002Fweb-hoch3.de","web-hoch3.de","Portal settings resolved",[46],{"id":47,"name":48,"location":49,"isActive":15,"isDefault":15,"items":50},2,"Main Menu","sidebar",[51,65,189,282,358,424],{"id":52,"title":53,"url":61,"target":62,"icon":63,"isActive":15,"productId":14,"categoryId":14,"shopCategoryId":14,"articleId":14,"pageId":14,"portfolioId":14,"children":64},"1",{"de":54,"en":55,"es":56,"fr":57,"it":55,"ru":58,"sr":59,"zh":60},"Startseite","Home","Inicio","Accueil","Главная","Početna","首页","\u002F","_self","i-lucide-home",[],{"id":66,"title":67,"url":76,"target":62,"icon":77,"isActive":15,"productId":14,"categoryId":14,"shopCategoryId":14,"articleId":14,"pageId":14,"portfolioId":14,"children":78},"3",{"de":68,"en":69,"es":70,"fr":71,"it":72,"ru":73,"sr":74,"zh":75},"Spiele","Games","Juegos","Jeux","Giochi","Игры","Igre","游戏","\u002Fgames","i-lucide-gamepad-2",[79,92,107,121,135,148,162,176],{"id":80,"title":81,"url":76,"target":62,"icon":90,"isActive":15,"productId":14,"categoryId":14,"shopCategoryId":14,"articleId":14,"pageId":14,"portfolioId":14,"children":91},"3-1",{"de":82,"en":83,"es":84,"fr":85,"it":86,"ru":87,"sr":88,"zh":89},"Spiele-Hub","Games Hub","Centro de juegos","Hub de jeux","Hub Giochi","Игровой центр","Centar za igre","游戏中心","i-lucide-layout-grid",[],{"id":93,"title":94,"url":103,"target":62,"icon":104,"isActive":15,"type":105,"productId":14,"categoryId":14,"shopCategoryId":14,"articleId":14,"pageId":14,"portfolioId":14,"children":106},"1-1",{"de":95,"en":96,"es":97,"fr":98,"it":99,"ru":100,"sr":101,"zh":102},"Neu & Angesagt","New & Trending","Novedades & Tendencias","Nouveautés & Tendances","Novità & Tendenze","Новинки & тренды","Novo & Popularno","新品 & 热门","\u002Fnews\u002Fgaming-news","i-lucide-sparkles","custom",[],{"id":108,"title":109,"url":118,"target":62,"icon":119,"isActive":15,"productId":14,"categoryId":14,"shopCategoryId":14,"articleId":14,"pageId":14,"portfolioId":14,"children":120},"1-2",{"de":110,"en":111,"es":112,"fr":113,"it":114,"ru":115,"sr":116,"zh":117},"Angebote","Deals","Ofertas","Offres","Offerte","Акции","Ponude","优惠","\u002Fdeals","i-lucide-badge-percent",[],{"id":122,"title":123,"url":132,"target":62,"icon":133,"isActive":15,"productId":14,"categoryId":14,"shopCategoryId":14,"articleId":14,"pageId":14,"portfolioId":14,"children":134},"1-3",{"de":124,"en":125,"es":126,"fr":127,"it":128,"ru":129,"sr":130,"zh":131},"Release-Kalender","Release Calendar","Calendario de lanzamientos","Calendrier des sorties","Calendario delle uscite","Календарь релизов","Kalendar izdanja","发布日历","\u002Freleases","i-lucide-calendar-days",[],{"id":136,"title":137,"url":145,"target":62,"icon":146,"isActive":15,"type":105,"productId":14,"categoryId":14,"shopCategoryId":14,"articleId":14,"pageId":14,"portfolioId":14,"children":147},"3-2",{"de":138,"en":139,"es":140,"fr":141,"it":142,"sr":143,"zh":144},"amiibo-kompatible Spiele","amiibo-Compatible Games","Juegos compatibles con Amiibo","Jeux compatibles Amiibo","Giochi compatibili con amiibo","Amiibo kompatibilne igre","Amiibo 兼容游戏","\u002Freference\u002Fcompatibility","i-lucide-check-circle-2",[],{"id":149,"title":150,"url":159,"target":62,"icon":160,"isActive":15,"type":105,"productId":14,"categoryId":14,"shopCategoryId":14,"articleId":14,"pageId":14,"portfolioId":14,"children":161},"3-3",{"de":151,"en":152,"es":153,"fr":154,"it":155,"ru":156,"sr":157,"zh":158},"Wie amiibo funktionieren","How amiibo work","Cómo funcionan los Amiibo","Comment fonctionnent les Amiibo","Come funzionano gli Amiibo","Как работают Amiibo","Kako Amiibo rade","Amiibo 如何运作","\u002Fgames\u002Fhow-amiibo-unlocks-work","i-lucide-puzzle",[],{"id":163,"title":164,"url":173,"target":62,"icon":174,"isActive":15,"type":105,"productId":14,"categoryId":14,"shopCategoryId":14,"articleId":14,"pageId":14,"portfolioId":14,"children":175},"3-4",{"de":165,"en":166,"es":167,"fr":168,"it":169,"ru":170,"sr":171,"zh":172},"Beste Amiibo pro Spiel","Best amiibo per Game","Mejores Amiibo por juego","Meilleurs Amiibo par jeu","Migliori Amiibo per gioco","Лучшие Amiibo по играм","Najbolji Amiibo po igri","各游戏最佳 Amiibo","\u002Fgames\u002Fbest-amiibo-per-game","i-lucide-award",[],{"id":177,"title":178,"url":186,"target":62,"icon":187,"isActive":15,"type":105,"productId":14,"categoryId":14,"shopCategoryId":14,"articleId":14,"pageId":14,"portfolioId":14,"children":188},"3-5",{"de":179,"en":180,"es":181,"fr":182,"it":183,"ru":184,"sr":185},"Freischaltbares & Belohnungen","Unlockables & Rewards","Desbloqueables & Recompensas","Déblocables & Récompenses","Sbloccabili & Ricompense","Разблокировки & награды","Otključavanja & nagrade","\u002Fgames\u002Funlocks-and-benefits","i-lucide-gift",[],{"id":190,"title":191,"url":199,"target":62,"icon":200,"isActive":15,"productId":14,"categoryId":14,"shopCategoryId":14,"articleId":14,"pageId":14,"portfolioId":14,"children":201},"4",{"de":192,"en":193,"es":194,"it":195,"ru":196,"sr":197,"zh":198},"Sammeln","Collecting","Coleccionismo","Collezionismo","Коллекционирование","Kolekcionarstvo","收藏","\u002Fcollecting","i-lucide-gem",[202,213,227,240,254,268],{"id":203,"title":204,"url":211,"target":62,"icon":90,"isActive":15,"type":105,"productId":14,"categoryId":14,"shopCategoryId":14,"articleId":14,"pageId":14,"portfolioId":14,"children":212},"4-1",{"de":205,"en":206,"fr":207,"it":208,"ru":209,"zh":210},"Sammel-Hub","Collecting Hub","Hub de collecte","Hub di raccolta","Центр сбора","收藏中心","\u002Famiibo\u002Fcollecting",[],{"id":214,"title":215,"url":224,"target":62,"icon":225,"isActive":15,"type":105,"productId":14,"categoryId":14,"shopCategoryId":14,"articleId":14,"pageId":14,"portfolioId":14,"children":226},"4-2",{"de":216,"en":217,"es":218,"fr":219,"it":220,"ru":221,"sr":222,"zh":223},"Kaufberatung","Buying Guide","Guía de compra","Guide d'achat","Guida all'acquisto","Гид покупателя","Vodič za kupovinu","购买指南","\u002Famiibo-library\u002Famiibo-buying-smart","i-lucide-shopping-bag",[],{"id":228,"title":229,"url":237,"target":62,"icon":238,"isActive":15,"type":105,"productId":14,"categoryId":14,"shopCategoryId":14,"articleId":14,"pageId":14,"portfolioId":14,"children":239},"4-3",{"de":230,"en":230,"es":231,"fr":232,"it":233,"ru":234,"sr":235,"zh":236},"Fake vs Real","Falso vs. Real","Faux vs Vrai","Falso vs Vero","Фейк vs Оригинал","Lažno vs Pravo","真假对比","\u002Famiibo-library\u002Famiibo-buying-smart\u002Favoid-fake-listings","i-lucide-scan",[],{"id":241,"title":242,"url":251,"target":62,"icon":252,"isActive":15,"type":105,"productId":14,"categoryId":14,"shopCategoryId":14,"articleId":14,"pageId":14,"portfolioId":14,"children":253},"4-4",{"de":243,"en":244,"es":245,"fr":246,"it":247,"ru":248,"sr":249,"zh":250},"Aufbewahrung & Präsentation","Storage & Display","Almacenamiento & exhibición","Rangement & Présentation","Contenitori & Esposizione","Хранение & витрины","Odlaganje & izlaganje","存储 & 展示","\u002Famiibo-library\u002Famiibo-care-and-storage","i-lucide-box",[],{"id":255,"title":256,"url":265,"target":62,"icon":266,"isActive":15,"type":105,"productId":14,"categoryId":14,"shopCategoryId":14,"articleId":14,"pageId":14,"portfolioId":14,"children":267},"4-5",{"de":257,"en":258,"es":259,"fr":260,"it":261,"ru":262,"sr":263,"zh":264},"Preisübersicht","Price Guide","Guía de precios","Guide des prix","Guida ai prezzi","Гид по ценам","Cenovnik","价格指南","\u002Fcollections\u002Fprice-guide","i-lucide-bar-chart-3",[],{"id":269,"title":270,"url":279,"target":62,"icon":280,"isActive":15,"type":105,"productId":14,"categoryId":14,"shopCategoryId":14,"articleId":14,"pageId":14,"portfolioId":14,"children":281},"4-6",{"de":271,"en":272,"es":273,"fr":274,"it":275,"ru":276,"sr":277,"zh":278},"Raritäten","Rare Finds","Hallazgos únicos","Trouvailles rares","Pezzi rari","Редкие находки","Retki nalazi","稀有发现","\u002Fcollections\u002Frare-and-notable","i-lucide-trophy",[],{"id":283,"title":284,"url":292,"target":62,"icon":293,"isActive":15,"productId":14,"categoryId":14,"shopCategoryId":14,"articleId":14,"pageId":14,"portfolioId":14,"children":294},"5",{"de":285,"en":286,"es":287,"fr":286,"it":288,"ru":289,"sr":290,"zh":291},"Technik","Tech","Tecnología","Tecnologia","Технологии","Tehnologija","科技","\u002Ftech","i-lucide-cpu",[295,306,318,331,344],{"id":296,"title":297,"url":292,"target":62,"icon":90,"isActive":15,"productId":14,"categoryId":14,"shopCategoryId":14,"articleId":14,"pageId":14,"portfolioId":14,"children":305},"5-1",{"de":298,"en":298,"es":299,"fr":300,"it":301,"ru":302,"sr":303,"zh":304},"Tech Hub","Centro tecnológico","Hub Tech","Hub Tecnologico","Технохаб","Tehnološki centar","科技中心",[],{"id":307,"title":308,"url":315,"target":62,"icon":316,"isActive":15,"type":105,"productId":14,"categoryId":14,"shopCategoryId":14,"articleId":14,"pageId":14,"portfolioId":14,"children":317},"5-2",{"de":309,"en":310,"fr":311,"it":312,"ru":313,"zh":314},"Audio- & Mikrofonqualität","Audio & Mic Quality","Qualité audio & micro","Qualità audio & mic","Качество звука & микрофона","音频 & 麦克风质量","\u002Fgear\u002Faudio","i-lucide-mic",[],{"id":319,"title":320,"url":328,"target":62,"icon":329,"isActive":15,"type":105,"productId":14,"categoryId":14,"shopCategoryId":14,"articleId":14,"pageId":14,"portfolioId":14,"children":330},"5-3",{"de":321,"en":322,"es":323,"fr":324,"it":325,"sr":326,"zh":327},"Controller & Zubehör","Controllers & Accessories","Mandos & accesorios","Manettes & accessoires","Controller & Accessori","Kontroleri & dodatna oprema","控制器 & 配件","\u002Fgear\u002Fcontrols","i-lucide-joystick",[],{"id":332,"title":333,"url":341,"target":62,"icon":342,"isActive":15,"type":105,"productId":14,"categoryId":14,"shopCategoryId":14,"articleId":14,"pageId":14,"portfolioId":14,"children":343},"5-4",{"de":334,"en":335,"es":336,"fr":337,"it":338,"ru":339,"sr":340},"Displays & Aufnahme","Displays & Capture","Pantallas & Captura","Écrans & Capture","Display & Acquisizione","Дисплеи & Захват","Ekrani & snimanje","\u002Fgear\u002Fdisplays","i-lucide-monitor",[],{"id":345,"title":346,"url":355,"target":62,"icon":356,"isActive":15,"type":105,"productId":14,"categoryId":14,"shopCategoryId":14,"articleId":14,"pageId":14,"portfolioId":14,"children":357},"5-5",{"de":347,"en":348,"es":349,"fr":350,"it":351,"ru":352,"sr":353,"zh":354},"Netzwerkstabilität","Network Stability","Estabilidad de la red","Stabilité du réseau","Stabilità della rete","Стабильность сети","Stabilnost mreže","网络稳定性","\u002Fplaybooks\u002Fnetwork-stability","i-lucide-wifi",[],{"id":359,"title":360,"url":368,"target":62,"icon":369,"isActive":15,"productId":14,"categoryId":14,"shopCategoryId":14,"articleId":14,"pageId":14,"portfolioId":14,"children":370},"6",{"de":361,"en":361,"es":362,"fr":363,"it":364,"ru":365,"sr":366,"zh":367},"Shop","Tienda","Boutique","Negozio","Магазин","Prodavnica","商店","\u002Fshop","i-lucide-shopping-cart",[371,383,391,399,412],{"id":372,"title":373,"url":368,"target":62,"icon":381,"isActive":15,"productId":14,"categoryId":14,"shopCategoryId":14,"articleId":14,"pageId":14,"portfolioId":14,"children":382},"6-1",{"de":374,"en":375,"es":376,"fr":377,"ru":378,"sr":379,"zh":380},"Shop-Hub","Shop Hub","Centro de compras","Espace Boutique","Центр магазина","Centar za kupovinu","购物中心","i-lucide-store",[],{"id":384,"title":385,"url":388,"target":62,"icon":389,"isActive":15,"type":105,"productId":14,"categoryId":14,"shopCategoryId":14,"articleId":14,"pageId":14,"portfolioId":14,"children":390},"6-2",{"de":386,"en":386,"es":387,"fr":387,"it":387,"sr":387,"zh":387},"amiibo","Amiibo","\u002Famiibo-shop","i-lucide-scan-line",[],{"id":392,"title":393,"url":396,"target":62,"icon":397,"isActive":15,"type":105,"productId":14,"categoryId":14,"shopCategoryId":14,"articleId":14,"pageId":14,"portfolioId":14,"children":398},"6-3",{"de":394,"en":394,"es":394,"fr":394,"it":394,"ru":394,"sr":394,"zh":395},"LEGO","乐高","\u002Flego-shop","i-lucide-blocks",[],{"id":400,"title":401,"url":409,"target":62,"icon":410,"isActive":15,"productId":14,"categoryId":14,"shopCategoryId":14,"articleId":14,"pageId":14,"portfolioId":14,"children":411},"6-4",{"de":402,"en":403,"es":404,"fr":405,"it":406,"sr":407,"zh":408},"Figuren & Sammlerstücke","Figures & Collectibles","Figuras & Coleccionables","Figurines & Objets de collection","Figure & Collezionismo","Figure & kolekcionarstvo","手办 & 收藏品","\u002Fshop\u002Ffigures","i-lucide-package",[],{"id":413,"title":414,"url":421,"target":62,"icon":422,"isActive":15,"type":105,"productId":14,"categoryId":14,"shopCategoryId":14,"articleId":14,"pageId":14,"portfolioId":14,"children":423},"6-5",{"de":415,"en":415,"fr":416,"it":417,"ru":418,"sr":419,"zh":420},"Gaming Gear","Équipement gaming","Accessori gaming","Игровое снаряжение","Gejming oprema","游戏装备","\u002Fgaming-gear-shop","i-lucide-headphones",[],{"id":425,"title":426,"url":427,"target":62,"icon":389,"isActive":15,"type":105,"productId":14,"categoryId":14,"shopCategoryId":14,"articleId":14,"pageId":14,"portfolioId":14,"children":428},"2",{"de":386,"en":386,"es":386,"fr":386,"it":386,"ru":386,"sr":386,"zh":386},"\u002Famiibo",[429,440,447,461,475,489,503,517],{"id":430,"title":431,"url":427,"target":62,"icon":90,"isActive":15,"type":105,"productId":14,"categoryId":14,"shopCategoryId":14,"articleId":14,"pageId":14,"portfolioId":14,"children":439},"2-1",{"de":432,"en":433,"es":434,"fr":435,"it":435,"ru":436,"sr":437,"zh":438},"amiibo-Hub","amiibo Hub","Centro amiibo","Hub amiibo","Хаб amiibo","amiibo centar","amiibo 中心",[],{"id":441,"title":442,"url":444,"target":62,"icon":445,"isActive":15,"type":105,"productId":14,"categoryId":14,"shopCategoryId":14,"articleId":14,"pageId":14,"portfolioId":14,"children":446},"1772724791061",{"de":443,"en":443},"Franchise","\u002Famiibo\u002Ffranchise","i-lucide-star",[],{"id":448,"title":449,"url":458,"target":62,"icon":459,"isActive":15,"type":105,"productId":14,"categoryId":14,"shopCategoryId":14,"articleId":14,"pageId":14,"portfolioId":14,"children":460},"2-3",{"de":450,"en":451,"es":452,"fr":453,"it":454,"ru":455,"sr":456,"zh":457},"Region & Verpackung (EU\u002FUS\u002FJP)","Region & Packaging (EU\u002FUS\u002FJP)","Región y embalaje (UE\u002FEE. UU.\u002FJP)","Région & Emballage (UE\u002FUS\u002FJP)","Regione & Confezione (EU\u002FUS\u002FJP)","Регион & упаковка (EU\u002FUS\u002FJP)","Region & pakovanje (EU\u002FUS\u002FJP)","地区 & 包装 (EU\u002FUS\u002FJP)","\u002Famiibo-library\u002Famiibo-editions-and-regions","i-lucide-globe",[],{"id":462,"title":463,"url":472,"target":62,"icon":473,"isActive":15,"type":105,"productId":14,"categoryId":14,"shopCategoryId":14,"articleId":14,"pageId":14,"portfolioId":14,"children":474},"2-4",{"de":464,"en":465,"es":466,"fr":467,"it":468,"ru":469,"sr":470,"zh":471},"Schnellidentifikations-Checkliste","Fast Identification Checklist","Lista de identificación rápida","Liste d'identification rapide","Checklist identificazione rapida","Чек-лист быстрой идентификации","Kontrolna lista za brzu identifikaciju","快速识别清单","\u002Famiibo\u002Fidentify-checklist","i-lucide-list-checks",[],{"id":476,"title":477,"url":486,"target":62,"icon":487,"isActive":15,"productId":14,"categoryId":14,"shopCategoryId":14,"articleId":14,"pageId":14,"portfolioId":14,"children":488},"2-6",{"de":478,"en":479,"es":480,"fr":481,"it":482,"ru":483,"sr":484,"zh":485},"Tipps für Sealed-Sammlungen","Sealed Collection Tips","Consejos de colección sellada","Conseils collection scellée","Consigli collezione sigillata","Советы по коллекции Sealed","Saveti za zapečaćene kolekcije","密封收藏贴士","\u002Famiibo\u002Fsealed-tips","i-lucide-shield-check",[],{"id":490,"title":491,"url":500,"target":62,"icon":501,"isActive":15,"type":105,"productId":14,"categoryId":14,"shopCategoryId":14,"articleId":14,"pageId":14,"portfolioId":14,"children":502},"2-5",{"de":492,"en":493,"es":494,"fr":495,"it":496,"ru":497,"sr":498,"zh":499},"Zustand & Bewertung","Condition & Grading","Estado & Graduación","État & Évaluation","Condizioni & Valutazione","Состояние & Оценка","Stanje & ocenjivanje","品相与分级","\u002Famiibo-library\u002Famiibo-condition-and-grading","i-lucide-badge-check",[],{"id":504,"title":505,"url":514,"target":62,"icon":515,"isActive":15,"type":105,"productId":14,"categoryId":14,"shopCategoryId":14,"articleId":14,"pageId":14,"portfolioId":14,"children":516},"2-2",{"de":506,"en":507,"es":508,"fr":509,"it":510,"ru":511,"sr":512,"zh":513},"Ausgaben & Nachdrucke","Editions & Reprints","Ediciones & Reimpresiones","Éditions & Réimpressions","Edizioni & ristampe","Издания & переиздания","Izdanja & reizdanja","版本 & 重印","\u002Famiibo\u002Fcollecting\u002F","i-lucide-layers",[],{"id":518,"title":519,"url":528,"target":62,"icon":529,"isActive":15,"type":105,"productId":14,"categoryId":14,"shopCategoryId":14,"articleId":14,"pageId":14,"portfolioId":14,"children":530},"2-7",{"de":520,"en":521,"es":522,"fr":523,"it":524,"ru":525,"sr":526,"zh":527},"Nachdruck-Timeline","Reprint Timeline","Cronología de reimpresión","Chronologie des réimpressions","Cronologia ristampe","Хронология переизданий","Hronologija reprinta","重印时间轴","\u002Freference\u002Ftimelines\u002Famiibo-reprint-timeline","i-lucide-timer",[],{"statusCode":4,"data":532,"message":3134},{"id":533,"title":534,"slug":535,"content":536,"contentJson":537,"excerpt":1627,"featuredImage":1628,"featuredImageAlt":1629,"featuredImageCaption":14,"featuredImageTitle":14,"featuredImageCopyright":14,"featuredImageAuthor":14,"featuredImageSourceUrl":14,"featuredImageLicense":14,"featuredImageIsAiGenerated":603,"status":1630,"publishedAt":1631,"createdAt":1632,"updatedAt":1633,"seoLocalePaths":1634,"categories":1643,"author":1662,"translations":1665},"453","渲染更少，推理更多：AI正在打破传统游戏硬件升级周期吗？","render-less-infer-more-is-ai-breaking-the-traditional-gaming-hardware-upgrade-cycle","\u003Cp>几十年来，发烧级PC游戏遵循着一个熟悉的交易：游戏要求更多算力，GPU世代提供更多算力，玩家在差距足够大时升级。这个循环从未完美运作，但方向是清晰的。到了2026年，这个循环变得越来越难以解读，因为两件事正在同时发生：AI基础设施正在激烈争夺芯片和内存资源，而AI技术正直接进入游戏生成帧的方式中。\u003C\u002Fp>\n\u003Caside class=\"editorjs-callout editorjs-callout--info my-6 rounded-xl border p-5 border-blue-300 bg-blue-50 dark:border-blue-900 dark:bg-blue-950\u002F20\" role=\"note\">\u003Cstrong class=\"block mb-2 text-gray-900 dark:text-gray-100\">直接回答\u003C\u002Fstrong>\u003Cdiv class=\"text-gray-700 dark:text-gray-200\">\u003Cstrong>旧的游戏升级周期正承受压力，但它并非简单地被更昂贵的GPU所取代。\u003C\u002Fstrong>AI正在改变等式的两边：它加剧了对半导体和内存资源的竞争，同时也使神经渲染技术成为可能，这些技术可以用推理来替代存储数据或传统渲染工作。重要的可能性在于，游戏可能通过变得更具计算选择性来应对更紧张的硬件经济，而不仅仅是要求更大的GPU。\u003C\u002Fdiv>\u003C\u002Faside>\n\u003Caside class=\"editorjs-callout editorjs-callout--note my-6 rounded-xl border p-5 border-gray-300 bg-gray-50 dark:border-gray-700 dark:bg-gray-900\u002F40\" role=\"note\">\u003Cstrong class=\"block mb-2 text-gray-900 dark:text-gray-100\">如何阅读本文\u003C\u002Fstrong>\u003Cdiv class=\"text-gray-700 dark:text-gray-200\">这是一个\u003Cstrong>有证据支持的假设\u003C\u002Fstrong>。有记录的市场数据和已发布或宣布的技术被标记为事实或行业信号。它们之间的联系是技术推断。后面描述的未来架构是Figure Rocks的假设，并非关于未宣布的NVIDIA、AMD、Intel、Microsoft或游戏机路线图的声明。\u003C\u002Fdiv>\u003C\u002Faside>\n\u003Cnav class=\"editorjs-toc\" data-editorjs-toc=\"true\" aria-label=\"目录\">\u003Cstrong class=\"editorjs-toc__title\">目录\u003C\u002Fstrong>\u003Col class=\"editorjs-toc__list editorjs-toc__list--depth-0\">\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-5\" class=\"editorjs-toc__link\">规模不匹配已经显而易见\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-9\" class=\"editorjs-toc__link\">内存是压力变得具体的地方\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-13\" class=\"editorjs-toc__link\">恰在同时，帧本身正在改变\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-19\" class=\"editorjs-toc__link\">只有正确表述，5090 的问题才有意义\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-22\" class=\"editorjs-toc__link\">内存正开始成为一个计算问题，而不仅仅是容量问题\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-28\" class=\"editorjs-toc__link\">技术应对可能如何运作\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-31\" class=\"editorjs-toc__link\">游戏有一个通用工作负载通常不具备的优势：它已经知道部分未来\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-35\" class=\"editorjs-toc__link\">为什么类 Spark 架构变得有趣\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-39\" class=\"editorjs-toc__link\">要实现这一转变，必须具备哪些条件？\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-41\" class=\"editorjs-toc__link\">买家应该对类似 Spark 的转变有何期待？\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-43\" class=\"editorjs-toc__link\">什么会证伪这一假设？\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-46\" class=\"editorjs-toc__link\">当今游戏玩家的实际结论\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-49\" class=\"editorjs-toc__link\">主要来源\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-69\" class=\"editorjs-toc__link\">规模不匹配已经可见\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-73\" class=\"editorjs-toc__link\">内存是压力变得具体的地方\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-77\" class=\"editorjs-toc__link\">与此同时，帧本身也在改变\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-83\" class=\"editorjs-toc__link\">只有正确表述，5090的问题才有意义\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-86\" class=\"editorjs-toc__link\">内存正开始成为计算问题，而不仅仅是容量问题\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-92\" class=\"editorjs-toc__link\">技术响应可能如何运作\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-95\" class=\"editorjs-toc__link\">游戏拥有通用工作负载通常不具备的一个优势：它已经知道部分未来\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-99\" class=\"editorjs-toc__link\">为什么类Spark架构变得有趣\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-103\" class=\"editorjs-toc__link\">这一转变要成功，必须满足哪些条件？\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-105\" class=\"editorjs-toc__link\">买家应从类Spark转变中期待什么？\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-107\" class=\"editorjs-toc__link\">什么会证伪这一假设？\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-110\" class=\"editorjs-toc__link\">当今玩家的实用结论\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-113\" class=\"editorjs-toc__link\">主要来源\u003C\u002Fa>\u003C\u002Fli>\u003C\u002Fol>\u003C\u002Fnav>\n\u003Ch2 id=\"section-5\">规模不匹配已经显而易见\u003C\u002Fh2>\n\u003Cp>Newzoo预测2026年全球游戏市场将达到2139亿美元，同比增长6.1%。半导体行业协会报告2025年全球半导体销售额为7917亿美元，增长25.6%，并预测2026年约为1万亿美元。游戏业务和半导体业务仍然相互依赖，但它们的激励正变得不那么对称：使能高端游戏的那个先进制造生态系统，现在服务于拥有更大基础设施预算的AI和数据中心买家。这些并非直接可比的产品类别；它们确立了游戏周围上游计算经济的规模。\u003C\u002Fp>\n\u003Cdiv class=\"overflow-x-auto\">\u003Ctable class=\"w-full border-collapse\">\u003Cthead>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">参考点\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">报告或预测值\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">为何在此重要\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">全球游戏市场，2026年预测\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">2139亿美元，同比增长6.1%\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">庞大的消费娱乐市场，但处于先进硬件供应的下游\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">PC游戏，2026年预测\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">459亿美元，同比增长5.3%\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">最直接受PC组件经济影响的游戏细分市场\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">全球半导体销售额，2025年\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">7917亿美元，同比增长25.6%\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">显示了服务于游戏、AI、移动、汽车和云的上游产业的规模\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">全球半导体产业，2026年预测\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">约1万亿美元\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">提高了将高端芯片仅专用于消费游戏的機會成本\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\n\u003Caside class=\"editorjs-callout editorjs-callout--warning my-6 rounded-xl border p-5 border-amber-300 bg-amber-50 dark:border-amber-900 dark:bg-amber-950\u002F20\" role=\"note\">\u003Cstrong class=\"block mb-2 text-gray-900 dark:text-gray-100\">这些市场数字能证明什么——以及不能证明什么\u003C\u002Fstrong>\u003Cdiv class=\"text-gray-700 dark:text-gray-200\">这些数字确立了\u003Cstrong>经济背景，而非直接因果关系\u003C\u002Fstrong>。它们并不证明某个特定的晶圆、GPU或内存封装被从玩家手中拿走并重新分配给AI。它们也不意味着游戏和半导体是世界上最大的两个产业。更狭窄、更站得住脚的观点是，一个2000亿美元以上的游戏市场依赖于一个年规模接近万亿美元的上游半导体经济，而AI基础设施现在在非常不同的预算和利润条件下竞争先进计算和内存。\u003C\u002Fdiv>\u003C\u002Faside>\n\u003Ch2 id=\"section-9\">内存是压力变得具体的地方\u003C\u002Fh2>\n\u003Cp>TrendForce多次报告称，随着AI需求增长，DRAM供应商正在将产能转向服务器和HBM产品。2026年3月，它预测第二季度传统DRAM合约价格环比上涨58-63%，并表示有限的GDDR分配正在制约图形内存供应。到第三季度，TrendForce表示，尽管AI和服务器需求依然强劲，消费客户正达到承受能力极限。\u003C\u002Fp>\n\u003Csection class=\"editorjs-process my-6\">\u003Ch3 class=\"mb-3 text-lg font-semibold\">资源分配压力\u003C\u002Fh3>\u003Cdiv class=\"grid grid-cols-1 md:grid-cols-2 xl:grid-cols-3 gap-4\">\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">1\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">1. AI基础设施扩张\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">云提供商和AI建设者获取服务器内存、HBM、企业级SSD和先进计算。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">2\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">2. 供应商优化产品组合\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">产能流向需求更强、利润率更高或长期合同的产品。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">3\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">3. 消费级分配收紧\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">PC、图形和消费级内存买家面临更高价格或受限的供应。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">4\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">4. 游戏厂商吸收或转嫁成本\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">板卡设计、VRAM容量和系统价格成为更艰难的权衡。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">5\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">5. 旧升级周期减弱\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">更快的GPU仍然有用，但蛮力扩展在相同的消费价格点上变得经济上更难重复。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003C\u002Fdiv>\u003C\u002Fsection>\n\u003Caside class=\"editorjs-callout editorjs-callout--warning my-6 rounded-xl border p-5 border-amber-300 bg-amber-50 dark:border-amber-900 dark:bg-amber-950\u002F20\" role=\"note\">\u003Cstrong class=\"block mb-2 text-gray-900 dark:text-gray-100\">2026年9月的信号，而非旧的短缺故事\u003C\u002Fstrong>\u003Cdiv class=\"text-gray-700 dark:text-gray-200\">TrendForce在2026年9月30日报道称，供应商仍优先将先进工艺产能用于高性能服务器产品，预计第四季度传统DRAM合约价格将再环比上涨10-15%。其DRAM公告还表示，HBM继续挤占通用DRAM产能。因此，这里描述的压力在2026年末仍然可见；它不仅仅基于年初的暂时性飙升。\u003C\u002Fdiv>\u003C\u002Faside>\n\u003Ch2 id=\"section-13\">恰在同时，帧本身正在改变\u003C\u002Fh2>\n\u003Cp>微软现在将机器学习描述为正在成为实时图形的基础。DirectX正在添加着色器级和模型级的ML路径，使神经纹理压缩、降噪、超分辨率和其他学习型工作负载能够位于图形栈内部，而不是作为孤立功能被附加。NVIDIA的DLSS 5研究通过引入生成式神经渲染阶段进一步推进了这一理念，该阶段用学习到的外观先验补充传统渲染。神经渲染并不意味着传统渲染消失。新兴的模式是混合式的：直接渲染一些信息，当这样做更便宜或更有用时，推断或重建其他信息。\u003C\u002Fp>\n\u003Caside class=\"editorjs-referral my-6\">\u003Ca href=\"https:\u002F\u002Ffigure.rocks\u002Fzh\u002Fblog\u002Fdlss-5-is-not-just-upscaling-what-3d-guided-neural-rendering-actually-changes\" class=\"flex flex-col sm:flex-row gap-4 rounded-xl border border-gray-200 dark:border-gray-700 p-4 transition hover:border-primary-500\">\u003Cdiv class=\"min-w-0 flex-1\">\u003Cstrong class=\"block text-lg text-gray-900 dark:text-gray-100\">DLSS 5不仅仅是超分辨率：生成式神经渲染实际改变了什么\u003C\u002Fstrong>\u003Cp class=\"mt-2 text-sm text-gray-600 dark:text-gray-300\">为什么重要的变化不是又一个质量模式，而是视觉信息生产方式的新阶段。\u003C\u002Fp>\u003Cspan class=\"mt-3 inline-flex text-sm font-medium text-primary-600 dark:text-primary-400\">阅读DLSS 5分析 →\u003C\u002Fspan>\u003C\u002Fdiv>\u003C\u002Fa>\u003C\u002Faside>\n\u003Csection class=\"editorjs-comparison my-6\">\u003Ch3 class=\"mb-3 text-lg font-semibold\">传统扩展 vs 神经时代扩展\u003C\u002Fh3>\u003Cdiv class=\"overflow-x-auto\">\u003Ctable class=\"w-full border-collapse\">\u003Cthead>\u003Ctr>\u003Cth class=\"border border-gray-300 bg-gray-50 px-4 py-3 text-left dark:border-gray-700 dark:bg-gray-900\">\u003C\u002Fth>\u003Cth class=\"border border-gray-300 bg-gray-50 px-4 py-3 text-left font-semibold dark:border-gray-700 dark:bg-gray-900\">传统路径\u003C\u002Fth>\u003Cth class=\"border border-gray-300 bg-gray-50 px-4 py-3 text-left font-semibold dark:border-gray-700 dark:bg-gray-900\">神经时代路径\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">更高视觉保真度\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">存储更多细节并渲染更多工作\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">选择性存储\u002F渲染，重建或推断某些细节\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">内存响应\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">增加容量和带宽\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">在需要时增加容量，但也压缩和减少移动\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">性能响应\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">更多着色器\u002FRT吞吐量\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">混合着色器、RT和矩阵\u002FTensor风格计算\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">软件角色\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">引擎主要调度传统图形工作\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">引擎越来越多地调度渲染、数据移动和ML工作负载\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\u003C\u002Fsection>\n\u003Caside class=\"editorjs-callout editorjs-callout--success my-6 rounded-xl border p-5 border-emerald-300 bg-emerald-50 dark:border-emerald-900 dark:bg-emerald-950\u002F20\" role=\"note\">\u003Cstrong class=\"block mb-2 text-gray-900 dark:text-gray-100\">Figure Rocks的核心假设\u003C\u002Fstrong>\u003Cdiv class=\"text-gray-700 dark:text-gray-200\">\u003Cstrong>AI可能正在削弱蛮力游戏硬件的经济性，同时提供使蛮力扩展变得不那么必要的技术。\u003C\u002Fstrong>这并不意味着未来的游戏需要更少的硬件。它意味着更多的结果可能来自决定存储什么、移动什么、渲染什么、重建什么或推断什么。Figure Rocks的简写：\u003Cstrong>少存储 → 少移动 → 多预测 → 少渲染 → 多推断。\u003C\u002Fstrong>这是一个关于性能提升可能越来越多来自何处的假设，并非声称每个未来游戏都会使用单一管线。\u003C\u002Fdiv>\u003C\u002Faside>\n\u003Cdiv class=\"ce-delimiter cdx-block my-8\">\u003C\u002Fdiv>\n\u003Ch2 id=\"section-19\">只有正确表述，5090 的问题才有意义\u003C\u002Fh2>\n\u003Cp>RTX 5090 之所以是一个有用的例子，恰恰因为证据是混杂的。截至 2026 年 10 月 18 日，NVIDIA 仍将 GeForce RTX 5090 列为在售产品。与此同时，零售商的行为和媒体报道记录了异常糟糕的供货情况，以及未经证实的停产说法。这是市场信号，而非官方的停产公告。因此，站得住脚的问题不是 NVIDIA 是否秘密砍掉了某款显卡，而是当类似架构可以服务于经济逻辑截然不同的专业可视化和 AI 市场时，将顶级 Blackwell 级芯片分配给发烧友游戏的机会成本是否在上升。\u003C\u002Fp>\n\u003Caside class=\"editorjs-callout editorjs-callout--note my-6 rounded-xl border p-5 border-gray-300 bg-gray-50 dark:border-gray-700 dark:bg-gray-900\u002F40\" role=\"note\">\u003Cstrong class=\"block mb-2 text-gray-900 dark:text-gray-100\">事实 \u002F 信号 \u002F 假设\u003C\u002Fstrong>\u003Cdiv class=\"text-gray-700 dark:text-gray-200\">\u003Cstrong>事实：\u003C\u002Fstrong>NVIDIA 仍公开列出 RTX 5090，且 Blackwell 架构横跨游戏和专业产品。\u003Cbr>\u003Cstrong>行业信号：\u003C\u002Fstrong>存在零售商下架、极度缺货以及可能停产的报道，但 NVIDIA 尚未确认 RTX 5090 的 EOL 或停产。\u003Cbr>\u003Cstrong>技术推断：\u003C\u002Fstrong>高端芯片有多个具有经济价值的目的地。\u003Cbr>\u003Cstrong>FIGURE ROCKS 假设：\u003C\u002Fstrong>AI 和专业需求可能使传统的发烧友预期——即最大的实用芯片仍能作为游戏旗舰广泛获得——更难维持。\u003C\u002Fdiv>\u003C\u002Faside>\n\u003Ch2 id=\"section-22\">内存正开始成为一个计算问题，而不仅仅是容量问题\u003C\u002Fh2>\n\u003Cp>至少可以通过四种方式应对内存瓶颈，而不只是提高原始 DRAM 速度：压缩表示、用计算重建细节、预测接下来需要哪些数据，以及让正确的数据驻留在靠近执行单元的位置。这些方法都不能让物理内存芯片神奇地变快。它们改变的是系统如何从有限的容量和带宽预算中获得更多有效工作。\u003C\u002Fp>\n\u003Cdiv class=\"overflow-x-auto\">\u003Ctable class=\"w-full border-collapse\">\u003Cthead>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">技术\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">改变了什么\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">什么不会自动改变\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">压缩\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">用更少的物理字节表示相同的逻辑数据\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">DRAM 时钟和总线宽度\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">神经重建\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">部分细节由紧凑表示生成\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">成本消失；推理有其自身的计算\u002F延迟成本\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">预测性预取\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">数据在请求工作负载停顿之前到达\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">总物理带宽\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">驻留 \u002F 统一访问\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">减少不必要的迁移或复制\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">争用和延迟约束\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\n\u003Caside class=\"editorjs-callout editorjs-callout--info my-6 rounded-xl border p-5 border-blue-300 bg-blue-50 dark:border-blue-900 dark:bg-blue-950\u002F20\" role=\"note\">\u003Cstrong class=\"block mb-2 text-gray-900 dark:text-gray-100\">研究锚点：有效容量不等于原始速度\u003C\u002Fstrong>\u003Cdiv class=\"text-gray-700 dark:text-gray-200\">NVIDIA Research 的 Buddy Compression 报告称，HPC 工作负载的平均压缩比为 \u003Cstrong>2.2×，深度学习工作负载为 1.5×\u003C\u002Fstrong>，该研究中性能下降 1–2%。这一结果并非游戏基准测试。它在此处的相关性在于架构层面：压缩内存可以在不改变物理 DRAM 时钟的情况下增加有效容量并减少部分流量。\u003C\u002Fdiv>\u003C\u002Faside>\n\u003Caside class=\"editorjs-callout editorjs-callout--success my-6 rounded-xl border p-5 border-emerald-300 bg-emerald-50 dark:border-emerald-900 dark:bg-emerald-950\u002F20\" role=\"note\">\u003Cstrong class=\"block mb-2 text-gray-900 dark:text-gray-100\">游戏特定证据：神经纹理压缩\u003C\u002Fstrong>\u003Cdiv class=\"text-gray-700 dark:text-gray-200\">NVIDIA Research 的材质纹理 NTC 工作证明，与 BC High 对比，可实现 \u003Cstrong>4× 更高的线性纹理分辨率——16× 更多纹素——同时使用少 30% 的内存\u003C\u002Fstrong>。这并不意味着每款游戏都能获得 16× 的内存收益。它展示了此处关键的机制：部分视觉细节可以更紧凑地表示，并通过计算重建，而不是以相同保真度按常规存储。\u003C\u002Fdiv>\u003C\u002Faside>\n\u003Caside class=\"editorjs-callout editorjs-callout--warning my-6 rounded-xl border p-5 border-amber-300 bg-amber-50 dark:border-amber-900 dark:bg-amber-950\u002F20\" role=\"note\">\u003Cstrong class=\"block mb-2 text-gray-900 dark:text-gray-100\">“460×”这个数字并不意味着什么\u003C\u002Fstrong>\u003Cdiv class=\"text-gray-700 dark:text-gray-200\">有时与 NVIDIA 系统关联的 460× 数字，指的是跨 DGX 代际的 GPU 可访问内存容量的比较，\u003Cstrong>而非 460× 的内存压缩比\u003C\u002Fstrong>。本文刻意不将其用作压缩的证据。\u003C\u002Fdiv>\u003C\u002Faside>\n\u003Ch2 id=\"section-28\">技术应对可能如何运作\u003C\u002Fh2>\n\u003Csection class=\"editorjs-process my-6\">\u003Ch3 class=\"mb-3 text-lg font-semibold\">一种可能的未来帧管线\u003C\u002Fh3>\u003Cdiv class=\"grid grid-cols-1 md:grid-cols-2 xl:grid-cols-3 gap-4\">\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">1\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">压缩\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">在质量允许的地方，让纹理或学习到的表示更小。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">2\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">预测\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">利用确定性游戏状态和学习模型来估计接下来需要哪些数据。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">3\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">预取\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">在 GPU 因等待数据而停顿之前，移动有用的数据。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">4\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">渲染\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">将常规图形工作花在对可见结果影响最大的地方。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">5\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">推理\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">在推理比直接存储或模拟所有内容更便宜的地方，使用神经重建或生成。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">6\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">呈现\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">在保持延迟、伪影和帧节奏处于可接受范围内的同时交付帧。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003C\u002Fdiv>\u003C\u002Fsection>\n\u003Cp>这条管线并非某个已发布的产品。每个组件都已以某种形式存在，但整合后的链条是假设。神经纹理压缩已表明，学习到的表示可以用计算换取内存。GPU 预取研究表明，学习到的预测可以提高内存页命中率。DirectX 正在将 ML 正式纳入图形编程模型。DLSS 5 表明生成阶段正在进入实时渲染。\u003C\u002Fp>\n\u003Ch2 id=\"section-31\">游戏有一个通用工作负载通常不具备的优势：它已经知道部分未来\u003C\u002Fh2>\n\u003Cp>游戏引擎知道相机变换、玩家速度、当前世界单元、相邻流送单元、动画状态、活动实体、LOD 过渡、着色器依赖，并且通常知道可能的行进路径。这种确定性上下文并不能完美揭示未来，但它为预测器提供了比仅靠通用内存轨迹更丰富的起点。\u003C\u002Fp>\n\u003Csection class=\"editorjs-process my-6\">\u003Ch3 class=\"mb-3 text-lg font-semibold\">一个预测性游戏内存示例\u003C\u002Fh3>\u003Cdiv class=\"grid grid-cols-1 md:grid-cols-2 xl:grid-cols-3 gap-4\">\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">1\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">1. 观察确定性状态\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">相机方向、移动向量、当前房间、可见入口和已知资产依赖已存在于引擎状态中。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">2\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">2. 估计近期工作集\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">学习到的预测器可以对可能的下一个房间、纹理、几何体、着色器或神经表示进行排序。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">3\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">3. 在可用带宽期间预取\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">在内存系统有余量时移动可能的资源，而不是等待硬性需求未命中。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">4\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">4. 保持确定性保证\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">关键资产仍使用编写好的规则；预测是补充而非替代正确性。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">5\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">5. 衡量未命中成本\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">如果预测浪费带宽或驱逐有用数据，置信度或预取深度必须降低。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003C\u002Fdiv>\u003C\u002Fsection>\n\u003Caside class=\"editorjs-callout editorjs-callout--note my-6 rounded-xl border p-5 border-gray-300 bg-gray-50 dark:border-gray-700 dark:bg-gray-900\u002F40\" role=\"note\">\u003Cstrong class=\"block mb-2 text-gray-900 dark:text-gray-100\">这个想法有研究先例，但尚无游戏领域的证明\u003C\u002Fstrong>\u003Cdiv class=\"text-gray-700 dark:text-gray-200\">一项已发表的 CPU–GPU 统一虚拟内存研究报告称，使用学习到的预取后，\u003Cstrong>性能提升 10.89%\u003C\u002Fstrong>，页命中率从 \u003Cstrong>76.10% 升至 89.02%\u003C\u002Fstrong>，且 \u003Cstrong>CPU–GPU 互连流量降低 11.05%\u003C\u002Fstrong>。一篇 2026 年 AAAI 关于异步 KV 缓存预取的论文报告称，在其 LLM 工作负载上实现了 \u003Cstrong>2.15× 的注意力内核效率和高达 1.97× 的端到端吞吐量\u003C\u002Fstrong>。这两项结果都不应直接照搬到游戏上；它们都表明，预测和预取可以在真实 GPU 工作负载中隐藏内存成本。\u003C\u002Fdiv>\u003C\u002Faside>\n\u003Ch2 id=\"section-35\">为什么类 Spark 架构变得有趣\u003C\u002Fh2>\n\u003Cp>DGX Spark 不是游戏 PC，将其视为游戏 PC 会掩盖有用的教训。其 GB10 设计结合了 Arm CPU 核心、Blackwell GPU、Tensor 和 RT 硬件，以及高达 128 GB 的一致性统一系统内存。这种架构提出了与经典台式机不同的问题：不是将系统清晰地分为 CPU 内存和独立 VRAM，而是当多个计算引擎共享一个大内存池时会发生什么？\u003C\u002Fp>\n\u003Cdiv class=\"overflow-x-auto\">\u003Ctable class=\"w-full border-collapse\">\u003Cthead>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">属性\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">DGX Spark 参考\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">如果该原则被采用，对游戏的影响\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">内存\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">高达 128 GB 一致性统一 LPDDR5X\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">大型共享池可以减少重复，并让本地 AI 和图形访问公共数据\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">内存带宽\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">273 GB\u002Fs\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">容量不够；高端图形仍需激进的带宽和延迟管理\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">计算组合\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">CPU + Blackwell GPU + Tensor + RT\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">未来的游戏工作负载可以拆分到更多专用引擎上\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">外形\u002F功耗\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">紧凑型 GB10 系统\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">集成可以提高效率，但可能降低模块化可升级性\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\n\u003Caside class=\"editorjs-callout editorjs-callout--warning my-6 rounded-xl border p-5 border-amber-300 bg-amber-50 dark:border-amber-900 dark:bg-amber-950\u002F20\" role=\"note\">\u003Cstrong class=\"block mb-2 text-gray-900 dark:text-gray-100\">容量不等于带宽\u003C\u002Fstrong>\u003Cdiv class=\"text-gray-700 dark:text-gray-200\">大型统一池\u003Cstrong>不会\u003C\u002Fstrong>自动取代高端独立 GPU 的带宽。DGX Spark 列出 273 GB\u002Fs 的内存带宽，而 NVIDIA 的 RTX PRO 6000 Blackwell 工作站版列出 1,792 GB\u002Fs。这些是非常不同的产品，比较并非游戏基准测试；它只是让架构约束变得可见。类似 Spark 的游戏设计需要解决争用、延迟和带宽问题——而不仅仅是提供更多共享千兆字节。\u003C\u002Fdiv>\u003C\u002Faside>\n\u003Ch2 id=\"section-39\">要实现这一转变，必须具备哪些条件？\u003C\u002Fh2>\n\u003Csection class=\"editorjs-process my-6\">\u003Ch3 class=\"mb-3 text-lg font-semibold\">神经游戏可行性测试\u003C\u002Fh3>\u003Cdiv class=\"grid grid-cols-1 md:grid-cols-2 xl:grid-cols-3 gap-4\">\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">1\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">1. 推理节省必须大于其成本\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">神经工作必须消除足够的存储、带宽或传统渲染成本，以证明其自身计算和延迟的合理性。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">2\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">2. 引擎必须暴露有用的状态\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">当系统知道摄像机方向、资产依赖关系、移动和可能的可见性时，预测效果更好。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">3\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">3. 内存移动必须变得有目的性\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">没有良好的驻留和预取策略，压缩仍可能导致停顿。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">4\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">4. 质量必须保持稳定\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">生成的细节不能引入不可接受的伪影、时间不稳定或不正确的游戏信息。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">5\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">5. 延迟必须有界\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">更漂亮的帧并不自动意味着更灵敏的游戏。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">6\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">6. 技术栈必须经受住供应商差异\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">有用的技术需要不会分裂成一次性演示的 API、驱动和引擎路径。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003C\u002Fdiv>\u003C\u002Fsection>\n\u003Ch2 id=\"section-41\">买家应该对类似 Spark 的转变有何期待？\u003C\u002Fh2>\n\u003Csection class=\"editorjs-comparison my-6\">\u003Ch3 class=\"mb-3 text-lg font-semibold\">潜在收益 vs 实际成本\u003C\u002Fh3>\u003Cdiv class=\"overflow-x-auto\">\u003Ctable class=\"w-full border-collapse\">\u003Cthead>\u003Ctr>\u003Cth class=\"border border-gray-300 bg-gray-50 px-4 py-3 text-left dark:border-gray-700 dark:bg-gray-900\">\u003C\u002Fth>\u003Cth class=\"border border-gray-300 bg-gray-50 px-4 py-3 text-left font-semibold dark:border-gray-700 dark:bg-gray-900\">潜在收益\u003C\u002Fth>\u003Cth class=\"border border-gray-300 bg-gray-50 px-4 py-3 text-left font-semibold dark:border-gray-700 dark:bg-gray-900\">成本 \u002F 风险\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">大型一致性池\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">CPU、图形和本地 AI 之间更灵活的共享\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">共享客户端可能争用同一带宽\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">减少重复\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">独立的系统内存和显存池之间更少的副本\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">一致性和调度增加了复杂性\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">更多本地 AI 余量\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">更大的模型、代理或神经资产可以在本地共存\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">AI 可能与图形争夺计算和内存\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">系统级优化\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">平台可以一起优化数据放置和加速器\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">更紧密的集成可能降低可升级性并增加锁定\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">新的工作负载灵活性\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">如果未来的游戏转向神经工作负载，机器可能更耐用\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">只有当引擎和运行时实际使用这些能力时\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\u003C\u002Fsection>\n\u003Ch2 id=\"section-43\">什么会证伪这一假设？\u003C\u002Fh2>\n\u003Cp>这一论点刻意可证伪。下表定义了市场、软件、硬件和游戏采用信号，这些信号会随着时间的推移使论点更强或更弱。\u003C\u002Fp>\n\u003Csection class=\"editorjs-comparison my-6\">\u003Ch3 class=\"mb-3 text-lg font-semibold\">加强或削弱论点的证据\u003C\u002Fh3>\u003Cdiv class=\"overflow-x-auto\">\u003Ctable class=\"w-full border-collapse\">\u003Cthead>\u003Ctr>\u003Cth class=\"border border-gray-300 bg-gray-50 px-4 py-3 text-left dark:border-gray-700 dark:bg-gray-900\">\u003C\u002Fth>\u003Cth class=\"border border-gray-300 bg-gray-50 px-4 py-3 text-left font-semibold dark:border-gray-700 dark:bg-gray-900\">会加强它\u003C\u002Fth>\u003Cth class=\"border border-gray-300 bg-gray-50 px-4 py-3 text-left font-semibold dark:border-gray-700 dark:bg-gray-900\">会削弱它\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">内存经济\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">服务器\u002FHBM 优先级继续挤压消费级分配\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">消费级容量扩大，价格压力正常化\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">图形栈\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">ML 原语成为常规引擎\u002FDirectX 功能\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">神经图形仍然是供应商特定的可选后处理\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">硬件设计\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">更多集成异构计算和更大的一致性内存池\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">独立 GPU + 独立内存 在成本、灵活性和延迟方面仍然明显优越\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">游戏采用\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">工作室广泛发布神经压缩、生成和预测数据路径\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">技术仍限于演示或少数旗舰游戏\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\u003C\u002Fsection>\n\u003Ch2 id=\"section-46\">当今游戏玩家的实际结论\u003C\u002Fh2>\n\u003Cp>不要因为推测性的架构可能出现而购买硬件。根据你实际拥有的游戏、分辨率、延迟目标和工作负载来购买。但不要假设面向未来仍然可以简化为一个数字，例如光栅性能或显存容量。软件栈的方向表明，内存行为、神经计算和引擎支持可能成为价值方程中越来越重要的部分。\u003C\u002Fp>\n\u003Caside class=\"editorjs-referral my-6\">\u003Ca href=\"https:\u002F\u002Ffigure.rocks\u002Fzh\u002Fblog\u002Fvram-usage-is-not-vram-requirement-why-a-full-memory-meter-does-not-tell-the-whole-story\" class=\"flex flex-col sm:flex-row gap-4 rounded-xl border border-gray-200 dark:border-gray-700 p-4 transition hover:border-primary-500\">\u003Cdiv class=\"min-w-0 flex-1\">\u003Cstrong class=\"block text-lg text-gray-900 dark:text-gray-100\">显存使用量不是显存需求量：为什么完整的内存计量器不能说明全部情况\u003C\u002Fstrong>\u003Cp class=\"mt-2 text-sm text-gray-600 dark:text-gray-300\">实用解释工作集、驻留以及为什么分配的显存数字与硬性需求不是一回事。\u003C\u002Fp>\u003Cspan class=\"mt-3 inline-flex text-sm font-medium text-primary-600 dark:text-primary-400\">阅读显存指南 →\u003C\u002Fspan>\u003C\u002Fdiv>\u003C\u002Fa>\u003C\u002Faside>\n\u003Ch2 id=\"section-49\">主要来源\u003C\u002Fh2>\n\u003Ca href=\"https:\u002F\u002Fnewzoo.com\u002Farticles\u002F2026-global-games-market-key-numbers\" target=\"_blank\" rel=\"noopener noreferrer\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">Newzoo — 2026 全球游戏市场关键数字\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">用于规模比较的 2026 年游戏市场预测。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fwww.semiconductors.org\u002Fglobal-annual-semiconductor-sales-increase-25-6-to-791-7-billion-in-2025\u002F\" target=\"_blank\" rel=\"noopener noreferrer\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">半导体行业协会 — 2025 年全球销售额\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">2025 年半导体销售额以及该协会约 1 万亿美元的 2026 年预测。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fwww.trendforce.com\u002Fpresscenter\u002Fnews\u002F20260930-13258.html\" target=\"_blank\" rel=\"noopener noreferrer\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">TrendForce — AI 服务器需求维持 2026 年第四季度内存价格上涨\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">2026 年末的证据，表明服务器优先的产能分配持续存在以及消费级压力。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fresearch.nvidia.com\u002Fpublication\u002F2019-04_buddy-compression-enabling-larger-memory-deep-learning-and-hpc-workloads-gpus\" target=\"_blank\" rel=\"noopener noreferrer\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">NVIDIA 研究 — Buddy 压缩\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">HPC 平均压缩比 2.2 倍和 DL 平均压缩比 1.5 倍的研究来源。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fwww.sciencedirect.com\u002Fscience\u002Farticle\u002Fabs\u002Fpii\u002FS0743731522002490\" target=\"_blank\" rel=\"noopener noreferrer\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">基于深度学习的 CPU–GPU UVM 预取\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">学习页面预取结果的研究来源。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fojs.aaai.org\u002Findex.php\u002FAAAI\u002Farticle\u002Fview\u002F39224\" target=\"_blank\" rel=\"noopener noreferrer\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">AAAI 2026 — 异步 KV 缓存预取\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">研究来源，展示通过主动 GPU 缓存预取实现延迟隐藏。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fwww.trendforce.com\u002Fpresscenter\u002Fnews\u002F20260331-12995.html\" target=\"_blank\" rel=\"noopener noreferrer\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">TrendForce — AI 服务器需求与 2026 年第二季度内存价格\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">服务器\u002FHBM 优先化、GDDR 限制与消费端压力的当前证据。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fdevblogs.microsoft.com\u002Fdirectx\u002Fevolving-directx-for-the-ml-era-on-windows\u002F\" target=\"_blank\" rel=\"noopener noreferrer\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">Microsoft DirectX — 为 ML 时代演进 DirectX\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">GDC 2026 官方概述，介绍 DirectX 中着色器级和模型级 ML。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fresearch.nvidia.com\u002Flabs\u002Fadlr\u002FDLSS5\u002F\" target=\"_blank\" rel=\"noopener noreferrer\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">NVIDIA Research — DLSS 5：生成式神经渲染\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">主要研究页面，将 DLSS 5 描述为实时生成式渲染阶段。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fwww.nvidia.com\u002Fen-us\u002Fproducts\u002Fworkstations\u002Fdgx-spark\u002F\" target=\"_blank\" rel=\"noopener noreferrer\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">NVIDIA — DGX Spark 规格\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">官方架构、统一内存和带宽规格。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fwww.nvidia.com\u002Fen-us\u002Fgeforce\u002Fgraphics-cards\u002F50-series\u002Frtx-5090\u002F\" target=\"_blank\" rel=\"noopener noreferrer\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">NVIDIA — GeForce RTX 5090\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">官方当前 RTX 5090 产品页面，用于区分已确认的产品状态与停产传闻。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fwww.tomshardware.com\u002Fpc-components\u002Fgpus\u002Fnvidia-rtx-5090-disappears-from-micro-center-amid-rumors-of-a-production-halt-retailer-flags-flagship-as-not-carried-as-stock-mysteriously-vanishes\" target=\"_blank\" rel=\"noopener noreferrer\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">Tom’s Hardware — RTX 5090 零售消失与停产传闻\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">当前市场报告，仅作为行业信号使用；报告明确指出 NVIDIA 尚未确认停产。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fresearch.nvidia.com\u002Flabs\u002Frtr\u002Fneural_texture_compression\u002F\" target=\"_blank\" rel=\"noopener noreferrer\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">NVIDIA Research — 材质纹理的随机访问神经压缩\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">主要来源，说明在内存比 BC High 少 30% 的情况下实现 4 倍线性分辨率 \u002F 16 倍纹素比较。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fwww.nvidia.com\u002Fde-de\u002Fproducts\u002Fworkstations\u002Fprofessional-desktop-gpus\u002Frtx-pro-6000\u002F\" target=\"_blank\" rel=\"noopener noreferrer\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">NVIDIA — RTX PRO 6000 Blackwell 工作站版\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">官方来源，提供 96 GB GDDR7 和 1,792 GB\u002Fs 内存带宽，用作带宽参考点。\u003C\u002Fp>\u003C\u002Fa>\n\u003Csection class=\"editorjs-glossary my-6 rounded-xl border border-gray-200 dark:border-gray-700 p-5\">\u003Ch3 class=\"mb-3 text-lg font-semibold\">关键概念\u003C\u002Fh3>\u003Cdl>\u003Cdiv id=\"neural-rendering\" class=\"border-t border-gray-200 dark:border-gray-700 py-3 first:border-t-0\">\u003Cdt class=\"font-semibold text-gray-900 dark:text-gray-100\">神经渲染\u003C\u002Fdt>\u003Cdd class=\"mt-1 text-gray-600 dark:text-gray-300\">在生成、重建或增强视觉输出的过程中使用训练模型作为一部分的渲染方法。\u003C\u002Fdd>\u003C\u002Fdiv>\u003Cdiv id=\"heterogeneous-compute\" class=\"border-t border-gray-200 dark:border-gray-700 py-3 first:border-t-0\">\u003Cdt class=\"font-semibold text-gray-900 dark:text-gray-100\">异构计算\u003C\u002Fdt>\u003Cdd class=\"mt-1 text-gray-600 dark:text-gray-300\">将不同工作负载分配给不同计算引擎（如 CPU、GPU、RT 和矩阵\u002FTensor 硬件）的系统。\u003C\u002Fdd>\u003C\u002Fdiv>\u003Cdiv id=\"effective-memory\" class=\"border-t border-gray-200 dark:border-gray-700 py-3 first:border-t-0\">\u003Cdt class=\"font-semibold text-gray-900 dark:text-gray-100\">有效内存\u003C\u002Fdt>\u003Cdd class=\"mt-1 text-gray-600 dark:text-gray-300\">在考虑压缩、驻留和移动效率后，对内存系统能够保留并供给计算的有用数据量的实际描述。它不是物理容量的正式替代。\u003C\u002Fdd>\u003C\u002Fdiv>\u003Cdiv id=\"predictive-memory\" class=\"border-t border-gray-200 dark:border-gray-700 py-3 first:border-t-0\">\u003Cdt class=\"font-semibold text-gray-900 dark:text-gray-100\">预测性内存\u003C\u002Fdt>\u003Cdd class=\"mt-1 text-gray-600 dark:text-gray-300\">Figure Rocks 的简写，指结合已知游戏状态和学习预测，在停顿发生前准备可能需要的数 据。它是一种假设，不是行业标准。\u003C\u002Fdd>\u003C\u002Fdiv>\u003C\u002Fdl>\u003C\u002Fsection>\n\u003Cp>几十年来，发烧级 PC 游戏遵循一个熟悉的交易：游戏要求更多计算，GPU 世代提供更多计算，玩家在差距足够大时升级。这个循环从未完美运作，但方向是可以理解的。到了 2026 年，这个循环变得越来越难以解读，因为两件事同时发生：AI 基础设施正在积极争夺芯片和内存，而 AI 技术正直接进入游戏创建帧的方式。\u003C\u002Fp>\n\u003Caside class=\"editorjs-callout editorjs-callout--info my-6 rounded-xl border p-5 border-blue-300 bg-blue-50 dark:border-blue-900 dark:bg-blue-950\u002F20\" role=\"note\">\u003Cstrong class=\"block mb-2 text-gray-900 dark:text-gray-100\">直接回答\u003C\u002Fstrong>\u003Cdiv class=\"text-gray-700 dark:text-gray-200\">\u003Cstrong>旧的游戏升级周期正面临压力，但它并不是简单地被更昂贵的 GPU 所取代。\u003C\u002Fstrong> AI 正在改变等式的两边：它加剧了对半导体和内存资源的竞争，同时也使神经渲染技术成为可能，这些技术可以用推理来换取存储数据或传统渲染工作。重要的可能性是，游戏可能通过变得更具计算选择性来应对更紧张的硬件经济，而不仅仅是要求更大的 GPU。\u003C\u002Fdiv>\u003C\u002Faside>\n\u003Caside class=\"editorjs-callout editorjs-callout--note my-6 rounded-xl border p-5 border-gray-300 bg-gray-50 dark:border-gray-700 dark:bg-gray-900\u002F40\" role=\"note\">\u003Cstrong class=\"block mb-2 text-gray-900 dark:text-gray-100\">如何阅读本文\u003C\u002Fstrong>\u003Cdiv class=\"text-gray-700 dark:text-gray-200\">这是一个\u003Cstrong>有证据支持的假设\u003C\u002Fstrong>。已记录的市场数据和已发布或宣布的技术被标记为事实或行业信号。它们之间的联系是技术推断。后面描述的未来架构是 Figure Rocks 的假设，不是关于未宣布的 NVIDIA、AMD、Intel、Microsoft 或游戏机路线图的声明。\u003C\u002Fdiv>\u003C\u002Faside>\n\u003Cnav class=\"editorjs-toc\" data-editorjs-toc=\"true\" aria-label=\"目录\">\u003Cstrong class=\"editorjs-toc__title\">目录\u003C\u002Fstrong>\u003Col class=\"editorjs-toc__list editorjs-toc__list--depth-0\">\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-5\" class=\"editorjs-toc__link\">规模不匹配已经显而易见\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-9\" class=\"editorjs-toc__link\">内存是压力变得具体的地方\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-13\" class=\"editorjs-toc__link\">恰在同时，帧本身正在改变\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-19\" class=\"editorjs-toc__link\">只有正确表述，5090 的问题才有意义\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-22\" class=\"editorjs-toc__link\">内存正开始成为一个计算问题，而不仅仅是容量问题\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-28\" class=\"editorjs-toc__link\">技术应对可能如何运作\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-31\" class=\"editorjs-toc__link\">游戏有一个通用工作负载通常不具备的优势：它已经知道部分未来\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-35\" class=\"editorjs-toc__link\">为什么类 Spark 架构变得有趣\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-39\" class=\"editorjs-toc__link\">要实现这一转变，必须具备哪些条件？\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-41\" class=\"editorjs-toc__link\">买家应该对类似 Spark 的转变有何期待？\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-43\" class=\"editorjs-toc__link\">什么会证伪这一假设？\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-46\" class=\"editorjs-toc__link\">当今游戏玩家的实际结论\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-49\" class=\"editorjs-toc__link\">主要来源\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-69\" class=\"editorjs-toc__link\">规模不匹配已经可见\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-73\" class=\"editorjs-toc__link\">内存是压力变得具体的地方\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-77\" class=\"editorjs-toc__link\">与此同时，帧本身也在改变\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-83\" class=\"editorjs-toc__link\">只有正确表述，5090的问题才有意义\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-86\" class=\"editorjs-toc__link\">内存正开始成为计算问题，而不仅仅是容量问题\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-92\" class=\"editorjs-toc__link\">技术响应可能如何运作\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-95\" class=\"editorjs-toc__link\">游戏拥有通用工作负载通常不具备的一个优势：它已经知道部分未来\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-99\" class=\"editorjs-toc__link\">为什么类Spark架构变得有趣\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-103\" class=\"editorjs-toc__link\">这一转变要成功，必须满足哪些条件？\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-105\" class=\"editorjs-toc__link\">买家应从类Spark转变中期待什么？\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-107\" class=\"editorjs-toc__link\">什么会证伪这一假设？\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-110\" class=\"editorjs-toc__link\">当今玩家的实用结论\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-113\" class=\"editorjs-toc__link\">主要来源\u003C\u002Fa>\u003C\u002Fli>\u003C\u002Fol>\u003C\u002Fnav>\n\u003Ch2 id=\"section-69\">规模不匹配已经可见\u003C\u002Fh2>\n\u003Cp>Newzoo 预测 2026 年全球游戏市场将达到 2139 亿美元，同比增长 6.1%。半导体行业协会报告 2025 年全球半导体销售额为 7917 亿美元，增长 25.6%，并预测 2026 年约为 1 万亿美元。游戏业务和半导体业务仍然相互依赖，但它们的激励正变得越来越不对称：同一个使能高端游戏的先进制造生态系统，现在服务于拥有更大基础设施预算的 AI 和数据中心买家。这些不是直接可比的产品类别；它们确立了游戏周围上游计算经济的规模。\u003C\u002Fp>\n\u003Cdiv class=\"overflow-x-auto\">\u003Ctable class=\"w-full border-collapse\">\u003Cthead>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">参考点\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">报告或预测值\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">为何在此重要\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">全球游戏市场，2026 年预测\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">2139 亿美元，同比增长 6.1%\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">大型消费娱乐市场，但处于先进硬件供应的下游\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">PC 游戏，2026 年预测\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">459 亿美元，同比增长 5.3%\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">最直接受 PC 组件经济影响的游戏细分市场\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">全球半导体销售额，2025 年\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">7917 亿美元，同比增长 25.6%\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">显示服务于游戏、AI、移动、汽车和云的上游产业的规模\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">全球半导体产业，2026 年预测\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">约 1 万亿美元\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">提高了将高端芯片仅专用于消费游戏的机 会成本\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\n\u003Caside class=\"editorjs-callout editorjs-callout--warning my-6 rounded-xl border p-5 border-amber-300 bg-amber-50 dark:border-amber-900 dark:bg-amber-950\u002F20\" role=\"note\">\u003Cstrong class=\"block mb-2 text-gray-900 dark:text-gray-100\">这些市场数字能证明什么——以及不能证明什么\u003C\u002Fstrong>\u003Cdiv class=\"text-gray-700 dark:text-gray-200\">这些数字确立了\u003Cstrong>经济背景，而非直接因果关系\u003C\u002Fstrong>。它们不能证明某个特定的晶圆、GPU 或内存封装从玩家手中被拿走并重新分配给 AI。它们也不意味着游戏和半导体是世界上最大的两个产业。更狭窄、更有说服力的观点是，一个 2000 亿美元以上的游戏市场依赖于年规模接近万亿美元的上游半导体经济，其中 AI 基础设施现在在非常不同的预算和利润条件下竞争先进计算和内存。\u003C\u002Fdiv>\u003C\u002Faside>\n\u003Ch2 id=\"section-73\">内存是压力变得具体的地方\u003C\u002Fh2>\n\u003Cp>TrendForce多次报告称，随着AI需求增长，DRAM供应商正将产能转向服务器和HBM产品。2026年3月，它预测第二季度传统DRAM合约价格环比上涨58–63%，并表示有限的GDDR分配正在制约图形内存供应。到第三季度，TrendForce表示，尽管AI和服务器需求依然强劲，消费类客户正达到可负担性极限。\u003C\u002Fp>\n\u003Csection class=\"editorjs-process my-6\">\u003Ch3 class=\"mb-3 text-lg font-semibold\">资源分配压力\u003C\u002Fh3>\u003Cdiv class=\"grid grid-cols-1 md:grid-cols-2 xl:grid-cols-3 gap-4\">\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">1\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">1. AI基础设施扩张\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">云提供商和AI建设者获取服务器内存、HBM、企业级SSD和先进计算资源。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">2\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">2. 供应商优化产品组合\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">产能转向需求更强、利润率更高或拥有长期合同的产品。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">3\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">3. 消费级分配收紧\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">PC、图形和消费级内存买家面临更高价格或受限的供应。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">4\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">4. 游戏厂商吸收或转嫁成本\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">板卡设计、VRAM容量和系统价格成为更艰难的权衡。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">5\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">5. 旧升级周期减弱\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">更快的GPU仍然有用，但在相同的消费价格点上，暴力堆料式扩展在经济上越来越难以重复。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003C\u002Fdiv>\u003C\u002Fsection>\n\u003Caside class=\"editorjs-callout editorjs-callout--warning my-6 rounded-xl border p-5 border-amber-300 bg-amber-50 dark:border-amber-900 dark:bg-amber-950\u002F20\" role=\"note\">\u003Cstrong class=\"block mb-2 text-gray-900 dark:text-gray-100\">2026年9月的信号，而非旧短缺故事\u003C\u002Fstrong>\u003Cdiv class=\"text-gray-700 dark:text-gray-200\">TrendForce于2026年9月30日报道称，供应商仍优先将先进制程产能用于高性能服务器产品，预计第四季度传统DRAM合约价格将再环比上涨10–15%。其DRAM公告还表示，HBM继续挤占通用DRAM产能。因此，这里描述的压力在2026年末依然可见；它并非仅基于年初的暂时性飙升。\u003C\u002Fdiv>\u003C\u002Faside>\n\u003Ch2 id=\"section-77\">与此同时，帧本身也在改变\u003C\u002Fh2>\n\u003Cp>微软现在将机器学习描述为正在成为实时图形的基础。DirectX正在添加着色器级和模型级的ML路径，使神经纹理压缩、降噪、超分辨率和其他学习型工作负载能够融入图形栈，而不是作为孤立功能附加。NVIDIA的DLSS 5研究通过引入生成式神经渲染阶段进一步推进了这一理念，该阶段用学习到的外观先验补充传统渲染。神经渲染并不意味着传统渲染消失。新兴的模式是混合式的：直接渲染一些信息，在更便宜或更有用时推断或重建其他信息。\u003C\u002Fp>\n\u003Caside class=\"editorjs-referral my-6\">\u003Ca href=\"https:\u002F\u002Ffigure.rocks\u002Fzh\u002Fblog\u002Fdlss-5-is-not-just-upscaling-what-3d-guided-neural-rendering-actually-changes\" class=\"flex flex-col sm:flex-row gap-4 rounded-xl border border-gray-200 dark:border-gray-700 p-4 transition hover:border-primary-500\">\u003Cdiv class=\"min-w-0 flex-1\">\u003Cstrong class=\"block text-lg text-gray-900 dark:text-gray-100\">DLSS 5不只是超分辨率：生成式神经渲染真正改变了什么\u003C\u002Fstrong>\u003Cp class=\"mt-2 text-sm text-gray-600 dark:text-gray-300\">为什么重要的变化不是又一个质量模式，而是视觉信息生产方式的新阶段。\u003C\u002Fp>\u003Cspan class=\"mt-3 inline-flex text-sm font-medium text-primary-600 dark:text-primary-400\">阅读DLSS 5分析 →\u003C\u002Fspan>\u003C\u002Fdiv>\u003C\u002Fa>\u003C\u002Faside>\n\u003Csection class=\"editorjs-comparison my-6\">\u003Ch3 class=\"mb-3 text-lg font-semibold\">传统缩放 vs 神经时代缩放\u003C\u002Fh3>\u003Cdiv class=\"overflow-x-auto\">\u003Ctable class=\"w-full border-collapse\">\u003Cthead>\u003Ctr>\u003Cth class=\"border border-gray-300 bg-gray-50 px-4 py-3 text-left dark:border-gray-700 dark:bg-gray-900\">\u003C\u002Fth>\u003Cth class=\"border border-gray-300 bg-gray-50 px-4 py-3 text-left font-semibold dark:border-gray-700 dark:bg-gray-900\">传统路径\u003C\u002Fth>\u003Cth class=\"border border-gray-300 bg-gray-50 px-4 py-3 text-left font-semibold dark:border-gray-700 dark:bg-gray-900\">神经时代路径\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">更高视觉保真度\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">存储更多细节并渲染更多工作\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">选择性存储\u002F渲染，重建或推断部分细节\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">内存响应\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">增加容量和带宽\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">在需要时增加容量，但也压缩并减少移动\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">性能响应\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">更多着色器\u002FRT吞吐量\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">混合着色器、RT和矩阵\u002FTensor风格计算\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">软件角色\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">引擎主要调度传统图形工作\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">引擎越来越多地调度渲染、数据移动和ML工作负载\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\u003C\u002Fsection>\n\u003Caside class=\"editorjs-callout editorjs-callout--success my-6 rounded-xl border p-5 border-emerald-300 bg-emerald-50 dark:border-emerald-900 dark:bg-emerald-950\u002F20\" role=\"note\">\u003Cstrong class=\"block mb-2 text-gray-900 dark:text-gray-100\">Figure Rocks的核心假设\u003C\u002Fstrong>\u003Cdiv class=\"text-gray-700 dark:text-gray-200\">\u003Cstrong>AI可能正在削弱暴力式游戏硬件的经济性，同时提供使暴力式缩放变得不那么必要的技术。\u003C\u002Fstrong>这并不意味着未来的游戏需要更少的硬件。它意味着更多结果可能来自决定存储什么、移动什么、渲染什么、重建什么或推断什么。Figure Rocks的简写：\u003Cstrong>少存储 → 少移动 → 多预测 → 少渲染 → 多推断。\u003C\u002Fstrong>这是一个关于性能提升可能越来越多来自何处的假设，而不是声称未来每款游戏都会使用单一管线。\u003C\u002Fdiv>\u003C\u002Faside>\n\u003Cdiv class=\"ce-delimiter cdx-block my-8\">\u003C\u002Fdiv>\n\u003Ch2 id=\"section-83\">只有正确表述，5090的问题才有意义\u003C\u002Fh2>\n\u003Cp>RTX 5090是一个有用的例子，恰恰因为证据是混合的。截至2026年10月18日，NVIDIA仍将GeForce RTX 5090列为当前产品。与此同时，零售商行为和媒体报道记录了异常糟糕的可用性以及未经证实的停产说法。这是一个市场信号，而非官方的停产公告。因此，站得住脚的问题不是NVIDIA是否秘密杀死了一款显卡，而是当类似架构可以服务于经济性截然不同的专业可视化和AI市场时，将顶级Blackwell级硅片分配给发烧级游戏的机会成本是否上升。\u003C\u002Fp>\n\u003Caside class=\"editorjs-callout editorjs-callout--note my-6 rounded-xl border p-5 border-gray-300 bg-gray-50 dark:border-gray-700 dark:bg-gray-900\u002F40\" role=\"note\">\u003Cstrong class=\"block mb-2 text-gray-900 dark:text-gray-100\">事实 \u002F 信号 \u002F 假设\u003C\u002Fstrong>\u003Cdiv class=\"text-gray-700 dark:text-gray-200\">\u003Cstrong>事实：\u003C\u002Fstrong>NVIDIA仍公开列出RTX 5090，且Blackwell架构横跨游戏和专业产品。\u003Cbr>\u003Cstrong>行业信号：\u003C\u002Fstrong>零售商下架、极度稀缺和可能停产的报道存在，但NVIDIA尚未确认RTX 5090的EOL或停产。\u003Cbr>\u003Cstrong>技术推断：\u003C\u002Fstrong>高端芯片有多个具有经济价值的目的地。\u003Cbr>\u003Cstrong>FIGURE ROCKS假设：\u003C\u002Fstrong>AI和专业需求可能使传统的发烧友期望——即最大的实用芯片仍能作为游戏旗舰广泛获得——更难维持。\u003C\u002Fdiv>\u003C\u002Faside>\n\u003Ch2 id=\"section-86\">内存正开始成为计算问题，而不仅仅是容量问题\u003C\u002Fh2>\n\u003Cp>至少有两种方法可以在不单纯提高原始DRAM速度的情况下攻击内存瓶颈：压缩表示、用计算重建细节、预测接下来需要什么数据，以及将正确的数据保持在执行单元附近。这些都不会让物理内存芯片神奇地变快。它们改变了系统从有限容量和带宽预算中获得多少有用工作。\u003C\u002Fp>\n\u003Cdiv class=\"overflow-x-auto\">\u003Ctable class=\"w-full border-collapse\">\u003Cthead>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">技术\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">改变了什么\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">什么不会自动改变\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">压缩\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">更少的物理字节表示相同的逻辑数据\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">DRAM时钟和总线宽度\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">神经重建\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">部分细节从紧凑表示生成\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">成本消失；推断有其自身的计算\u002F延迟成本\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">预测性预取\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">数据在请求工作负载停顿之前到达\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">总物理带宽\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">驻留\u002F统一访问\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">减少不必要的迁移或复制\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">争用和延迟约束\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\n\u003Caside class=\"editorjs-callout editorjs-callout--info my-6 rounded-xl border p-5 border-blue-300 bg-blue-50 dark:border-blue-900 dark:bg-blue-950\u002F20\" role=\"note\">\u003Cstrong class=\"block mb-2 text-gray-900 dark:text-gray-100\">研究锚点：有效容量不是原始速度\u003C\u002Fstrong>\u003Cdiv class=\"text-gray-700 dark:text-gray-200\">NVIDIA Research的Buddy Compression报告称，HPC工作负载平均压缩比为\u003Cstrong>2.2倍，深度学习工作负载为1.5倍\u003C\u002Fstrong>，该研究中减速为1–2%。该结果不是游戏基准测试。它在此处的相关性是架构性的：压缩内存可以增加有效容量并减少部分流量，而无需改变物理DRAM时钟。\u003C\u002Fdiv>\u003C\u002Faside>\n\u003Caside class=\"editorjs-callout editorjs-callout--success my-6 rounded-xl border p-5 border-emerald-300 bg-emerald-50 dark:border-emerald-900 dark:bg-emerald-950\u002F20\" role=\"note\">\u003Cstrong class=\"block mb-2 text-gray-900 dark:text-gray-100\">游戏特定证据：神经纹理压缩\u003C\u002Fstrong>\u003Cdiv class=\"text-gray-700 dark:text-gray-200\">NVIDIA Research的材质纹理NTC工作证明，与BC High比较，\u003Cstrong>线性纹理分辨率提高4倍——纹素多16倍——同时内存使用减少30%\u003C\u002Fstrong>。这并不意味着每款游戏都能获得16倍的内存增益。它展示了这里重要的机制：一些视觉细节可以更紧凑地表示，并通过计算重建，而不是以相同保真度传统存储。\u003C\u002Fdiv>\u003C\u002Faside>\n\u003Caside class=\"editorjs-callout editorjs-callout--warning my-6 rounded-xl border p-5 border-amber-300 bg-amber-50 dark:border-amber-900 dark:bg-amber-950\u002F20\" role=\"note\">\u003Cstrong class=\"block mb-2 text-gray-900 dark:text-gray-100\">“460×”这个数字并不代表什么\u003C\u002Fstrong>\u003Cdiv class=\"text-gray-700 dark:text-gray-200\">有时与NVIDIA系统关联的460×数字，指的是跨DGX代际的GPU可访问内存容量比较，\u003Cstrong>而非460倍内存压缩比\u003C\u002Fstrong>。本文刻意不将其作为压缩的证据。\u003C\u002Fdiv>\u003C\u002Faside>\n\u003Ch2 id=\"section-92\">技术响应可能如何运作\u003C\u002Fh2>\n\u003Csection class=\"editorjs-process my-6\">\u003Ch3 class=\"mb-3 text-lg font-semibold\">一种可能的未来帧管线\u003C\u002Fh3>\u003Cdiv class=\"grid grid-cols-1 md:grid-cols-2 xl:grid-cols-3 gap-4\">\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">1\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">压缩\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">在质量允许的情况下，保持纹理或学习到的表示更小。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">2\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">预测\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">使用确定性游戏状态和学习模型来估计接下来需要哪些数据。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">3\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">预取\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">在GPU因等待数据而停滞之前，提前移动有用的数据。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">4\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">渲染\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">将传统图形工作集中在最能改变可见结果的地方。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">5\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">推理\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">在推理比直接存储或模拟所有内容更便宜的地方，使用神经重建或生成。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">6\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">呈现\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">在保持延迟、伪影和帧节奏在可接受范围内的同时交付帧。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003C\u002Fdiv>\u003C\u002Fsection>\n\u003Cp>这条管线并非某个已发布的产品。每个组件都已以某种形式存在，但整合后的链条仍是假设。神经纹理压缩已表明，学习到的表示可以用计算换取内存。GPU预取研究表明，学习到的预测可以提高内存页命中率。DirectX正在将机器学习形式化到图形编程模型中。DLSS 5显示生成阶段正在进入实时渲染。\u003C\u002Fp>\n\u003Ch2 id=\"section-95\">游戏拥有通用工作负载通常不具备的一个优势：它已经知道部分未来\u003C\u002Fh2>\n\u003Cp>游戏引擎知道相机变换、玩家速度、当前世界单元、相邻流送单元、动画状态、活动实体、LOD过渡、着色器依赖关系，以及通常可能的遍历路径。这种确定性上下文并不能完美揭示未来，但它为预测器提供了比仅靠通用内存跟踪更丰富的起点。\u003C\u002Fp>\n\u003Csection class=\"editorjs-process my-6\">\u003Ch3 class=\"mb-3 text-lg font-semibold\">一个预测性游戏内存示例\u003C\u002Fh3>\u003Cdiv class=\"grid grid-cols-1 md:grid-cols-2 xl:grid-cols-3 gap-4\">\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">1\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">1. 观察确定性状态\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">相机方向、移动向量、当前房间、可见传送门和已知资产依赖关系已存在于引擎状态中。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">2\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">2. 估计近期工作集\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">学习到的预测器可以对可能的下一房间、纹理、几何体、着色器或神经表示进行排序。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">3\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">3. 在可用带宽期间预取\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">在内存系统有余量时移动可能的资源，而不是等待硬性需求缺失。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">4\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">4. 保持确定性保证\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">关键资产仍使用编写规则；预测补充而非替代正确性。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">5\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">5. 衡量缺失成本\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">如果预测浪费带宽或驱逐有用数据，置信度或预取深度必须降低。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003C\u002Fdiv>\u003C\u002Fsection>\n\u003Caside class=\"editorjs-callout editorjs-callout--note my-6 rounded-xl border p-5 border-gray-300 bg-gray-50 dark:border-gray-700 dark:bg-gray-900\u002F40\" role=\"note\">\u003Cstrong class=\"block mb-2 text-gray-900 dark:text-gray-100\">这一想法有研究先例，但尚无游戏领域的证明\u003C\u002Fstrong>\u003Cdiv class=\"text-gray-700 dark:text-gray-200\">一项已发表的CPU–GPU统一虚拟内存研究报告称，使用学习到的预取实现了\u003Cstrong>10.89%的性能提升\u003C\u002Fstrong>，页命中率从\u003Cstrong>76.10%升至89.02%\u003C\u002Fstrong>，且\u003Cstrong>CPU–GPU互连流量降低11.05%\u003C\u002Fstrong>。2026年AAAI关于异步KV缓存预取的论文在其LLM工作负载上报告了\u003Cstrong>2.15倍注意力内核效率和高达1.97倍的端到端吞吐量\u003C\u002Fstrong>。这两项结果都不应直接照搬到游戏上；它们都表明预测和预取可以在真实GPU工作负载中隐藏内存成本。\u003C\u002Fdiv>\u003C\u002Faside>\n\u003Ch2 id=\"section-99\">为什么类Spark架构变得有趣\u003C\u002Fh2>\n\u003Cp>DGX Spark不是游戏PC，将其视为游戏PC会掩盖有用的教训。其GB10设计结合了Arm CPU核心、Blackwell GPU、Tensor和RT硬件，以及高达128 GB的一致性统一系统内存。这种架构提出了与经典台式机不同的问题：不是将系统清晰地分为CPU内存和独立VRAM，而是当多个计算引擎共享一个大内存池时会发生什么？\u003C\u002Fp>\n\u003Cdiv class=\"overflow-x-auto\">\u003Ctable class=\"w-full border-collapse\">\u003Cthead>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">属性\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">DGX Spark参考\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">若该原则被适配，对游戏的影响\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">内存\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">高达128 GB一致性统一LPDDR5X\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">大型共享池可减少重复，并让本地AI和图形访问公共数据\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">内存带宽\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">273 GB\u002Fs\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">容量不够；高端图形仍需激进的带宽和延迟管理\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">计算组合\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">CPU + Blackwell GPU + Tensor + RT\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">未来的游戏工作负载可以拆分到更多专用引擎上\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">形态\u002F功耗\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">紧凑型GB10系统\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">集成可提高效率，但可能降低模块化可升级性\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\n\u003Caside class=\"editorjs-callout editorjs-callout--warning my-6 rounded-xl border p-5 border-amber-300 bg-amber-50 dark:border-amber-900 dark:bg-amber-950\u002F20\" role=\"note\">\u003Cstrong class=\"block mb-2 text-gray-900 dark:text-gray-100\">容量不等于带宽\u003C\u002Fstrong>\u003Cdiv class=\"text-gray-700 dark:text-gray-200\">大型统一池\u003Cstrong>不会\u003C\u002Fstrong>自动替代高端独立GPU的带宽。DGX Spark列出273 GB\u002Fs的内存带宽，而NVIDIA的RTX PRO 6000 Blackwell工作站版列出1,792 GB\u002Fs。这些是非常不同的产品，比较并非游戏基准测试；它只是让架构约束变得可见。类Spark的游戏设计需要解决争用、延迟和带宽问题——而不仅仅是提供更多共享千兆字节。\u003C\u002Fdiv>\u003C\u002Faside>\n\u003Ch2 id=\"section-103\">这一转变要成功，必须满足哪些条件？\u003C\u002Fh2>\n\u003Csection class=\"editorjs-process my-6\">\u003Ch3 class=\"mb-3 text-lg font-semibold\">神经游戏可行性测试\u003C\u002Fh3>\u003Cdiv class=\"grid grid-cols-1 md:grid-cols-2 xl:grid-cols-3 gap-4\">\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">1\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">1. 推理节省必须大于其成本\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">神经工作必须消除足够的存储、带宽或传统渲染成本，以证明其自身计算和延迟的合理性。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">2\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">2. 引擎必须暴露有用的状态\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">当系统知道相机方向、资产依赖关系、移动和可能的可见性时，预测效果更好。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">3\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">3. 内存移动必须变得有意识\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">没有良好的驻留和预取策略，压缩仍可能导致停滞。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">4\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">4. 质量必须保持稳定\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">生成的细节不能引入不可接受的伪影、时间不稳定或不正确的游戏信息。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">5\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">5. 延迟必须有界\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">更漂亮的帧并不自动意味着更响应灵敏的游戏。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">6\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">6. 技术栈必须经受供应商差异\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">有用的技术需要API、驱动和引擎路径，不能碎片化为一次性演示。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003C\u002Fdiv>\u003C\u002Fsection>\n\u003Ch2 id=\"section-105\">买家应从类Spark转变中期待什么？\u003C\u002Fh2>\n\u003Csection class=\"editorjs-comparison my-6\">\u003Ch3 class=\"mb-3 text-lg font-semibold\">潜在收益 vs 实际成本\u003C\u002Fh3>\u003Cdiv class=\"overflow-x-auto\">\u003Ctable class=\"w-full border-collapse\">\u003Cthead>\u003Ctr>\u003Cth class=\"border border-gray-300 bg-gray-50 px-4 py-3 text-left dark:border-gray-700 dark:bg-gray-900\">\u003C\u002Fth>\u003Cth class=\"border border-gray-300 bg-gray-50 px-4 py-3 text-left font-semibold dark:border-gray-700 dark:bg-gray-900\">潜在收益\u003C\u002Fth>\u003Cth class=\"border border-gray-300 bg-gray-50 px-4 py-3 text-left font-semibold dark:border-gray-700 dark:bg-gray-900\">成本\u002F风险\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">大型一致性池\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">CPU、图形和本地AI之间更灵活的共享\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">共享客户端可能争用同一带宽\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">减少重复\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">独立系统RAM和VRAM池之间的副本更少\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">一致性和调度增加复杂性\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">更多本地AI余量\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">更大的模型、代理或神经资产可以在本地共存\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">AI可能与图形争夺计算和内存\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">系统级优化\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">平台可以一起优化数据放置和加速器\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">更紧密的集成可能降低可升级性并增加锁定\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">新的工作负载灵活性\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">如果未来游戏转向神经工作负载，机器可能更耐用\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">只有当引擎和运行时实际使用这些能力时\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\u003C\u002Fsection>\n\u003Ch2 id=\"section-107\">什么会证伪这一假设？\u003C\u002Fh2>\n\u003Cp>这一论点刻意可证伪。下表定义了市场、软件、硬件和游戏采用信号，这些信号会随时间使论点更强或更弱。\u003C\u002Fp>\n\u003Csection class=\"editorjs-comparison my-6\">\u003Ch3 class=\"mb-3 text-lg font-semibold\">强化或削弱论点的证据\u003C\u002Fh3>\u003Cdiv class=\"overflow-x-auto\">\u003Ctable class=\"w-full border-collapse\">\u003Cthead>\u003Ctr>\u003Cth class=\"border border-gray-300 bg-gray-50 px-4 py-3 text-left dark:border-gray-700 dark:bg-gray-900\">\u003C\u002Fth>\u003Cth class=\"border border-gray-300 bg-gray-50 px-4 py-3 text-left font-semibold dark:border-gray-700 dark:bg-gray-900\">会强化论点\u003C\u002Fth>\u003Cth class=\"border border-gray-300 bg-gray-50 px-4 py-3 text-left font-semibold dark:border-gray-700 dark:bg-gray-900\">会削弱论点\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">内存经济学\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">服务器\u002FHBM 优先级持续挤压消费级分配\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">消费者产能扩大，价格压力趋于正常\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">图形技术栈\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">机器学习原语成为常规引擎\u002FDirectX 特性\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">神经图形仍为厂商特定的可选后处理\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">硬件设计\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">更集成的异构计算和更大的连贯内存池\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">独立 GPU + 独立内存仍在成本、灵活性和延迟上明显占优\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">游戏采用\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">工作室广泛部署神经压缩、生成和预测数据路径\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">技术仍局限于演示或少数旗舰游戏\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\u003C\u002Fsection>\n\u003Ch2 id=\"section-110\">当今玩家的实用结论\u003C\u002Fh2>\n\u003Cp>不要因为某种推测性架构可能到来而购买硬件。应根据你实际拥有的游戏、分辨率、延迟目标和工作负载来购买。但不要假设面向未来的考量仍能简化为光栅性能或显存容量这样的单一数字。软件栈的发展方向表明，内存行为、神经计算和引擎支持可能日益成为价值方程中的重要部分。\u003C\u002Fp>\n\u003Caside class=\"editorjs-referral my-6\">\u003Ca href=\"https:\u002F\u002Ffigure.rocks\u002Fzh\u002Fblog\u002Fvram-usage-is-not-vram-requirement-why-a-full-memory-meter-does-not-tell-the-whole-story\" class=\"flex flex-col sm:flex-row gap-4 rounded-xl border border-gray-200 dark:border-gray-700 p-4 transition hover:border-primary-500\">\u003Cdiv class=\"min-w-0 flex-1\">\u003Cstrong class=\"block text-lg text-gray-900 dark:text-gray-100\">显存使用量不等于显存需求：为什么显存占用条满并不说明全部问题\u003C\u002Fstrong>\u003Cp class=\"mt-2 text-sm text-gray-600 dark:text-gray-300\">实用解释工作集、驻留，以及为什么已分配的显存数值与硬性需求不是一回事。\u003C\u002Fp>\u003Cspan class=\"mt-3 inline-flex text-sm font-medium text-primary-600 dark:text-primary-400\">阅读显存指南 →\u003C\u002Fspan>\u003C\u002Fdiv>\u003C\u002Fa>\u003C\u002Faside>\n\u003Ch2 id=\"section-113\">主要来源\u003C\u002Fh2>\n\u003Ca href=\"https:\u002F\u002Fnewzoo.com\u002Farticles\u002F2026-global-games-market-key-numbers\" target=\"_blank\" rel=\"noopener noreferrer\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">Newzoo — 2026 全球游戏市场关键数据\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">用于规模比较的 2026 年游戏市场预测。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fwww.semiconductors.org\u002Fglobal-annual-semiconductor-sales-increase-25-6-to-791-7-billion-in-2025\u002F\" target=\"_blank\" rel=\"noopener noreferrer\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">半导体行业协会 — 2025 年全球销售额\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">2025 年半导体销售额及该协会约 1 万亿美元的 2026 年预测。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fwww.trendforce.com\u002Fpresscenter\u002Fnews\u002F20260930-13258.html\" target=\"_blank\" rel=\"noopener noreferrer\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">TrendForce — AI 服务器需求支撑 2026 年第四季度内存价格上涨\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">2026 年末关于服务器优先产能分配持续及消费端压力的证据。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fresearch.nvidia.com\u002Fpublication\u002F2019-04_buddy-compression-enabling-larger-memory-deep-learning-and-hpc-workloads-gpus\" target=\"_blank\" rel=\"noopener noreferrer\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">NVIDIA Research — Buddy Compression\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">HPC 平均压缩比 2.2 倍、DL 平均压缩比 1.5 倍的研究来源。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fwww.sciencedirect.com\u002Fscience\u002Farticle\u002Fabs\u002Fpii\u002FS0743731522002490\" target=\"_blank\" rel=\"noopener noreferrer\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">基于深度学习的 CPU–GPU UVM 预取\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">学习型页面预取结果的研究来源。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fojs.aaai.org\u002Findex.php\u002FAAAI\u002Farticle\u002Fview\u002F39224\" target=\"_blank\" rel=\"noopener noreferrer\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">AAAI 2026 — 异步 KV 缓存预取\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">通过主动 GPU 缓存预取实现延迟隐藏的研究来源。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fwww.trendforce.com\u002Fpresscenter\u002Fnews\u002F20260331-12995.html\" target=\"_blank\" rel=\"noopener noreferrer\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">TrendForce — AI 服务器需求与 2026 年第二季度内存价格\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">关于服务器\u002FHBM 优先、GDDR 限制和消费端压力的当前证据。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fdevblogs.microsoft.com\u002Fdirectx\u002Fevolving-directx-for-the-ml-era-on-windows\u002F\" target=\"_blank\" rel=\"noopener noreferrer\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">Microsoft DirectX — 为 ML 时代演进 DirectX\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">GDC 2026 关于 DirectX 中着色器级和模型级 ML 的官方概述。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fresearch.nvidia.com\u002Flabs\u002Fadlr\u002FDLSS5\u002F\" target=\"_blank\" rel=\"noopener noreferrer\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">NVIDIA Research — DLSS 5：生成式神经渲染\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">将 DLSS 5 描述为实时生成式渲染阶段的主要研究页面。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fwww.nvidia.com\u002Fen-us\u002Fproducts\u002Fworkstations\u002Fdgx-spark\u002F\" target=\"_blank\" rel=\"noopener noreferrer\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">NVIDIA — DGX Spark 规格\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">官方架构、统一内存和带宽规格。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fwww.nvidia.com\u002Fen-us\u002Fgeforce\u002Fgraphics-cards\u002F50-series\u002Frtx-5090\u002F\" target=\"_blank\" rel=\"noopener noreferrer\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">NVIDIA — GeForce RTX 5090\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">官方当前 RTX 5090 产品页面，用于区分已确认的产品状态与停产传闻。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fwww.tomshardware.com\u002Fpc-components\u002Fgpus\u002Fnvidia-rtx-5090-disappears-from-micro-center-amid-rumors-of-a-production-halt-retailer-flags-flagship-as-not-carried-as-stock-mysteriously-vanishes\" target=\"_blank\" rel=\"noopener noreferrer\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">Tom’s Hardware — RTX 5090 零售消失与停产传闻\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">仅作为行业信号的当前市场报道；该报道明确指出 NVIDIA 尚未确认停产。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fresearch.nvidia.com\u002Flabs\u002Frtr\u002Fneural_texture_compression\u002F\" target=\"_blank\" rel=\"noopener noreferrer\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">NVIDIA Research — 材质纹理的随机访问神经压缩\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">关于相比 BC High 内存减少 30% 时实现 4 倍线性分辨率 \u002F 16 倍纹素比较的主要来源。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fwww.nvidia.com\u002Fde-de\u002Fproducts\u002Fworkstations\u002Fprofessional-desktop-gpus\u002Frtx-pro-6000\u002F\" target=\"_blank\" rel=\"noopener noreferrer\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">NVIDIA — RTX PRO 6000 Blackwell 工作站版\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">96 GB GDDR7 和 1,792 GB\u002Fs 内存带宽的官方来源，用作带宽参考点。\u003C\u002Fp>\u003C\u002Fa>\n\u003Csection class=\"editorjs-glossary my-6 rounded-xl border border-gray-200 dark:border-gray-700 p-5\">\u003Ch3 class=\"mb-3 text-lg font-semibold\">关键概念\u003C\u002Fh3>\u003Cdl>\u003Cdiv id=\"neural-rendering\" class=\"border-t border-gray-200 dark:border-gray-700 py-3 first:border-t-0\">\u003Cdt class=\"font-semibold text-gray-900 dark:text-gray-100\">神经渲染\u003C\u002Fdt>\u003Cdd class=\"mt-1 text-gray-600 dark:text-gray-300\">使用训练好的模型作为生成、重建或增强视觉输出的一部分的渲染方法。\u003C\u002Fdd>\u003C\u002Fdiv>\u003Cdiv id=\"heterogeneous-compute\" class=\"border-t border-gray-200 dark:border-gray-700 py-3 first:border-t-0\">\u003Cdt class=\"font-semibold text-gray-900 dark:text-gray-100\">异构计算\u003C\u002Fdt>\u003Cdd class=\"mt-1 text-gray-600 dark:text-gray-300\">将不同工作负载分配给不同计算引擎（如 CPU、GPU、RT 和矩阵\u002FTensor 硬件）的系统。\u003C\u002Fdd>\u003C\u002Fdiv>\u003Cdiv id=\"effective-memory\" class=\"border-t border-gray-200 dark:border-gray-700 py-3 first:border-t-0\">\u003Cdt class=\"font-semibold text-gray-900 dark:text-gray-100\">有效内存\u003C\u002Fdt>\u003Cdd class=\"mt-1 text-gray-600 dark:text-gray-300\">在考虑压缩、驻留和移动效率后，对内存系统能够保留并供给计算的有用数据量的实际描述。它不是对物理容量的正式替代。\u003C\u002Fdd>\u003C\u002Fdiv>\u003Cdiv id=\"predictive-memory\" class=\"border-t border-gray-200 dark:border-gray-700 py-3 first:border-t-0\">\u003Cdt class=\"font-semibold text-gray-900 dark:text-gray-100\">预测性内存\u003C\u002Fdt>\u003Cdd class=\"mt-1 text-gray-600 dark:text-gray-300\">Figure Rocks 的简写，指结合已知游戏状态和学习预测，在停顿发生前准备可能需要的数​​据。这是一种假设，并非行业标准。\u003C\u002Fdd>\u003C\u002Fdiv>\u003C\u002Fdl>\u003C\u002Fsection>",{"time":538,"blocks":539,"version":1626},1792337215764,[540,546,554,561,568,574,579,606,613,618,623,646,652,657,662,671,709,716,721,726,731,737,742,747,772,778,784,790,795,818,823,828,833,854,860,865,870,895,901,906,930,935,978,983,988,1024,1029,1034,1042,1047,1057,1066,1075,1084,1093,1102,1111,1120,1129,1138,1147,1156,1165,1174,1197,1201,1205,1209,1212,1216,1220,1241,1245,1248,1252,1264,1269,1273,1277,1281,1299,1303,1306,1310,1314,1318,1322,1326,1340,1345,1349,1354,1358,1373,1377,1381,1385,1399,1404,1408,1412,1428,1432,1436,1452,1456,1480,1483,1487,1515,1519,1523,1528,1531,1537,1543,1550,1557,1563,1569,1575,1581,1587,1592,1597,1603,1609,1615],{"id":541,"data":542,"type":544,"tunes":545},"p-0001",{"text":543},"几十年来，发烧级PC游戏遵循着一个熟悉的交易：游戏要求更多算力，GPU世代提供更多算力，玩家在差距足够大时升级。这个循环从未完美运作，但方向是清晰的。到了2026年，这个循环变得越来越难以解读，因为两件事正在同时发生：AI基础设施正在激烈争夺芯片和内存资源，而AI技术正直接进入游戏生成帧的方式中。","paragraph",{},{"id":547,"data":548,"type":552,"tunes":553},"c-0002",{"body":549,"title":550,"variant":551},"\u003Cstrong>旧的游戏升级周期正承受压力，但它并非简单地被更昂贵的GPU所取代。\u003C\u002Fstrong>AI正在改变等式的两边：它加剧了对半导体和内存资源的竞争，同时也使神经渲染技术成为可能，这些技术可以用推理来替代存储数据或传统渲染工作。重要的可能性在于，游戏可能通过变得更具计算选择性来应对更紧张的硬件经济，而不仅仅是要求更大的GPU。","直接回答","info","callout",{},{"id":555,"data":556,"type":552,"tunes":560},"c-0003",{"body":557,"title":558,"variant":559},"这是一个\u003Cstrong>有证据支持的假设\u003C\u002Fstrong>。有记录的市场数据和已发布或宣布的技术被标记为事实或行业信号。它们之间的联系是技术推断。后面描述的未来架构是Figure Rocks的假设，并非关于未宣布的NVIDIA、AMD、Intel、Microsoft或游戏机路线图的声明。","如何阅读本文","note",{},{"id":562,"data":563,"type":566,"tunes":567},"toc-0004",{"title":564,"maxLevel":565,"minLevel":47},"目录",3,"tableOfContents",{},{"id":569,"data":570,"type":572,"tunes":573},"h-0005",{"text":571,"level":47},"规模不匹配已经显而易见","header",{},{"id":575,"data":576,"type":544,"tunes":578},"p-0006",{"text":577},"Newzoo预测2026年全球游戏市场将达到2139亿美元，同比增长6.1%。半导体行业协会报告2025年全球半导体销售额为7917亿美元，增长25.6%，并预测2026年约为1万亿美元。游戏业务和半导体业务仍然相互依赖，但它们的激励正变得不那么对称：使能高端游戏的那个先进制造生态系统，现在服务于拥有更大基础设施预算的AI和数据中心买家。这些并非直接可比的产品类别；它们确立了游戏周围上游计算经济的规模。",{},{"id":580,"data":581,"type":604,"tunes":605},"t-0007",{"content":582,"stretched":603,"withHeadings":15},[583,587,591,595,599],[584,585,586],"参考点","报告或预测值","为何在此重要",[588,589,590],"全球游戏市场，2026年预测","2139亿美元，同比增长6.1%","庞大的消费娱乐市场，但处于先进硬件供应的下游",[592,593,594],"PC游戏，2026年预测","459亿美元，同比增长5.3%","最直接受PC组件经济影响的游戏细分市场",[596,597,598],"全球半导体销售额，2025年","7917亿美元，同比增长25.6%","显示了服务于游戏、AI、移动、汽车和云的上游产业的规模",[600,601,602],"全球半导体产业，2026年预测","约1万亿美元","提高了将高端芯片仅专用于消费游戏的機會成本",false,"table",{},{"id":607,"data":608,"type":552,"tunes":612},"c-0008",{"body":609,"title":610,"variant":611},"这些数字确立了\u003Cstrong>经济背景，而非直接因果关系\u003C\u002Fstrong>。它们并不证明某个特定的晶圆、GPU或内存封装被从玩家手中拿走并重新分配给AI。它们也不意味着游戏和半导体是世界上最大的两个产业。更狭窄、更站得住脚的观点是，一个2000亿美元以上的游戏市场依赖于一个年规模接近万亿美元的上游半导体经济，而AI基础设施现在在非常不同的预算和利润条件下竞争先进计算和内存。","这些市场数字能证明什么——以及不能证明什么","warning",{},{"id":614,"data":615,"type":572,"tunes":617},"h-0009",{"text":616,"level":47},"内存是压力变得具体的地方",{},{"id":619,"data":620,"type":544,"tunes":622},"p-0010",{"text":621},"TrendForce多次报告称，随着AI需求增长，DRAM供应商正在将产能转向服务器和HBM产品。2026年3月，它预测第二季度传统DRAM合约价格环比上涨58-63%，并表示有限的GDDR分配正在制约图形内存供应。到第三季度，TrendForce表示，尽管AI和服务器需求依然强劲，消费客户正达到承受能力极限。",{},{"id":624,"data":625,"type":644,"tunes":645},"flow-0011",{"steps":626,"title":642,"orientation":643},[627,630,633,636,639],{"label":628,"description":629},"1. AI基础设施扩张","云提供商和AI建设者获取服务器内存、HBM、企业级SSD和先进计算。",{"label":631,"description":632},"2. 供应商优化产品组合","产能流向需求更强、利润率更高或长期合同的产品。",{"label":634,"description":635},"3. 消费级分配收紧","PC、图形和消费级内存买家面临更高价格或受限的供应。",{"label":637,"description":638},"4. 游戏厂商吸收或转嫁成本","板卡设计、VRAM容量和系统价格成为更艰难的权衡。",{"label":640,"description":641},"5. 旧升级周期减弱","更快的GPU仍然有用，但蛮力扩展在相同的消费价格点上变得经济上更难重复。","资源分配压力","auto","processFlow",{},{"id":647,"data":648,"type":552,"tunes":651},"v3-sep-signal",{"body":649,"title":650,"variant":611},"TrendForce在2026年9月30日报道称，供应商仍优先将先进工艺产能用于高性能服务器产品，预计第四季度传统DRAM合约价格将再环比上涨10-15%。其DRAM公告还表示，HBM继续挤占通用DRAM产能。因此，这里描述的压力在2026年末仍然可见；它不仅仅基于年初的暂时性飙升。","2026年9月的信号，而非旧的短缺故事",{},{"id":653,"data":654,"type":572,"tunes":656},"h-0012",{"text":655,"level":47},"恰在同时，帧本身正在改变",{},{"id":658,"data":659,"type":544,"tunes":661},"p-0013",{"text":660},"微软现在将机器学习描述为正在成为实时图形的基础。DirectX正在添加着色器级和模型级的ML路径，使神经纹理压缩、降噪、超分辨率和其他学习型工作负载能够位于图形栈内部，而不是作为孤立功能被附加。NVIDIA的DLSS 5研究通过引入生成式神经渲染阶段进一步推进了这一理念，该阶段用学习到的外观先验补充传统渲染。神经渲染并不意味着传统渲染消失。新兴的模式是混合式的：直接渲染一些信息，当这样做更便宜或更有用时，推断或重建其他信息。",{},{"id":663,"data":664,"type":669,"tunes":670},"ref-0014",{"url":665,"title":666,"excerpt":667,"ctaLabel":668},"https:\u002F\u002Ffigure.rocks\u002Fzh\u002Fblog\u002Fdlss-5-is-not-just-upscaling-what-3d-guided-neural-rendering-actually-changes","DLSS 5不仅仅是超分辨率：生成式神经渲染实际改变了什么","为什么重要的变化不是又一个质量模式，而是视觉信息生产方式的新阶段。","阅读DLSS 5分析","referralArticle",{},{"id":672,"data":673,"type":707,"tunes":708},"cmp-0015",{"rows":674,"title":699,"layout":604,"columns":700},[675,681,687,693],{"id":676,"label":677,"values":678},"fidelity","更高视觉保真度",{"new":679,"old":680},"选择性存储\u002F渲染，重建或推断某些细节","存储更多细节并渲染更多工作",{"id":682,"label":683,"values":684},"memory","内存响应",{"new":685,"old":686},"在需要时增加容量，但也压缩和减少移动","增加容量和带宽",{"id":688,"label":689,"values":690},"performance","性能响应",{"new":691,"old":692},"混合着色器、RT和矩阵\u002FTensor风格计算","更多着色器\u002FRT吞吐量",{"id":694,"label":695,"values":696},"software","软件角色",{"new":697,"old":698},"引擎越来越多地调度渲染、数据移动和ML工作负载","引擎主要调度传统图形工作","传统扩展 vs 神经时代扩展",[701,704],{"id":702,"label":703},"old","传统路径",{"id":705,"label":706},"new","神经时代路径","comparison",{},{"id":710,"data":711,"type":552,"tunes":715},"c-0016",{"body":712,"title":713,"variant":714},"\u003Cstrong>AI可能正在削弱蛮力游戏硬件的经济性，同时提供使蛮力扩展变得不那么必要的技术。\u003C\u002Fstrong>这并不意味着未来的游戏需要更少的硬件。它意味着更多的结果可能来自决定存储什么、移动什么、渲染什么、重建什么或推断什么。Figure Rocks的简写：\u003Cstrong>少存储 → 少移动 → 多预测 → 少渲染 → 多推断。\u003C\u002Fstrong>这是一个关于性能提升可能越来越多来自何处的假设，并非声称每个未来游戏都会使用单一管线。","Figure Rocks的核心假设","success",{},{"id":717,"data":718,"type":719,"tunes":720},"d-0017",{},"delimiter",{},{"id":722,"data":723,"type":572,"tunes":725},"h-0018",{"text":724,"level":47},"只有正确表述，5090 的问题才有意义",{},{"id":727,"data":728,"type":544,"tunes":730},"p-0019",{"text":729},"RTX 5090 之所以是一个有用的例子，恰恰因为证据是混杂的。截至 2026 年 10 月 18 日，NVIDIA 仍将 GeForce RTX 5090 列为在售产品。与此同时，零售商的行为和媒体报道记录了异常糟糕的供货情况，以及未经证实的停产说法。这是市场信号，而非官方的停产公告。因此，站得住脚的问题不是 NVIDIA 是否秘密砍掉了某款显卡，而是当类似架构可以服务于经济逻辑截然不同的专业可视化和 AI 市场时，将顶级 Blackwell 级芯片分配给发烧友游戏的机会成本是否在上升。",{},{"id":732,"data":733,"type":552,"tunes":736},"c-0020",{"body":734,"title":735,"variant":559},"\u003Cstrong>事实：\u003C\u002Fstrong>NVIDIA 仍公开列出 RTX 5090，且 Blackwell 架构横跨游戏和专业产品。\u003Cbr>\u003Cstrong>行业信号：\u003C\u002Fstrong>存在零售商下架、极度缺货以及可能停产的报道，但 NVIDIA 尚未确认 RTX 5090 的 EOL 或停产。\u003Cbr>\u003Cstrong>技术推断：\u003C\u002Fstrong>高端芯片有多个具有经济价值的目的地。\u003Cbr>\u003Cstrong>FIGURE ROCKS 假设：\u003C\u002Fstrong>AI 和专业需求可能使传统的发烧友预期——即最大的实用芯片仍能作为游戏旗舰广泛获得——更难维持。","事实 \u002F 信号 \u002F 假设",{},{"id":738,"data":739,"type":572,"tunes":741},"v3-memory-compute",{"text":740,"level":47},"内存正开始成为一个计算问题，而不仅仅是容量问题",{},{"id":743,"data":744,"type":544,"tunes":746},"v3-memory-compute-p1",{"text":745},"至少可以通过四种方式应对内存瓶颈，而不只是提高原始 DRAM 速度：压缩表示、用计算重建细节、预测接下来需要哪些数据，以及让正确的数据驻留在靠近执行单元的位置。这些方法都不能让物理内存芯片神奇地变快。它们改变的是系统如何从有限的容量和带宽预算中获得更多有效工作。",{},{"id":748,"data":749,"type":604,"tunes":771},"v3-four-memory-levers",{"content":750,"stretched":603,"withHeadings":15},[751,755,759,763,767],[752,753,754],"技术","改变了什么","什么不会自动改变",[756,757,758],"压缩","用更少的物理字节表示相同的逻辑数据","DRAM 时钟和总线宽度",[760,761,762],"神经重建","部分细节由紧凑表示生成","成本消失；推理有其自身的计算\u002F延迟成本",[764,765,766],"预测性预取","数据在请求工作负载停顿之前到达","总物理带宽",[768,769,770],"驻留 \u002F 统一访问","减少不必要的迁移或复制","争用和延迟约束",{},{"id":773,"data":774,"type":552,"tunes":777},"v3-buddy-data",{"body":775,"title":776,"variant":551},"NVIDIA Research 的 Buddy Compression 报告称，HPC 工作负载的平均压缩比为 \u003Cstrong>2.2×，深度学习工作负载为 1.5×\u003C\u002Fstrong>，该研究中性能下降 1–2%。这一结果并非游戏基准测试。它在此处的相关性在于架构层面：压缩内存可以在不改变物理 DRAM 时钟的情况下增加有效容量并减少部分流量。","研究锚点：有效容量不等于原始速度",{},{"id":779,"data":780,"type":552,"tunes":783},"v5-ntc-evidence",{"body":781,"title":782,"variant":714},"NVIDIA Research 的材质纹理 NTC 工作证明，与 BC High 对比，可实现 \u003Cstrong>4× 更高的线性纹理分辨率——16× 更多纹素——同时使用少 30% 的内存\u003C\u002Fstrong>。这并不意味着每款游戏都能获得 16× 的内存收益。它展示了此处关键的机制：部分视觉细节可以更紧凑地表示，并通过计算重建，而不是以相同保真度按常规存储。","游戏特定证据：神经纹理压缩",{},{"id":785,"data":786,"type":552,"tunes":789},"v3-460-correction",{"body":787,"title":788,"variant":611},"有时与 NVIDIA 系统关联的 460× 数字，指的是跨 DGX 代际的 GPU 可访问内存容量的比较，\u003Cstrong>而非 460× 的内存压缩比\u003C\u002Fstrong>。本文刻意不将其用作压缩的证据。","“460×”这个数字并不意味着什么",{},{"id":791,"data":792,"type":572,"tunes":794},"h-0021",{"text":793,"level":47},"技术应对可能如何运作",{},{"id":796,"data":797,"type":644,"tunes":817},"flow-0022",{"steps":798,"title":816,"orientation":643},[799,801,804,807,810,813],{"label":756,"description":800},"在质量允许的地方，让纹理或学习到的表示更小。",{"label":802,"description":803},"预测","利用确定性游戏状态和学习模型来估计接下来需要哪些数据。",{"label":805,"description":806},"预取","在 GPU 因等待数据而停顿之前，移动有用的数据。",{"label":808,"description":809},"渲染","将常规图形工作花在对可见结果影响最大的地方。",{"label":811,"description":812},"推理","在推理比直接存储或模拟所有内容更便宜的地方，使用神经重建或生成。",{"label":814,"description":815},"呈现","在保持延迟、伪影和帧节奏处于可接受范围内的同时交付帧。","一种可能的未来帧管线",{},{"id":819,"data":820,"type":544,"tunes":822},"p-0023",{"text":821},"这条管线并非某个已发布的产品。每个组件都已以某种形式存在，但整合后的链条是假设。神经纹理压缩已表明，学习到的表示可以用计算换取内存。GPU 预取研究表明，学习到的预测可以提高内存页命中率。DirectX 正在将 ML 正式纳入图形编程模型。DLSS 5 表明生成阶段正在进入实时渲染。",{},{"id":824,"data":825,"type":572,"tunes":827},"v3-predictive-loop",{"text":826,"level":47},"游戏有一个通用工作负载通常不具备的优势：它已经知道部分未来",{},{"id":829,"data":830,"type":544,"tunes":832},"v3-predictive-loop-p1",{"text":831},"游戏引擎知道相机变换、玩家速度、当前世界单元、相邻流送单元、动画状态、活动实体、LOD 过渡、着色器依赖，并且通常知道可能的行进路径。这种确定性上下文并不能完美揭示未来，但它为预测器提供了比仅靠通用内存轨迹更丰富的起点。",{},{"id":834,"data":835,"type":644,"tunes":853},"v3-predictive-door",{"steps":836,"title":852,"orientation":643},[837,840,843,846,849],{"label":838,"description":839},"1. 观察确定性状态","相机方向、移动向量、当前房间、可见入口和已知资产依赖已存在于引擎状态中。",{"label":841,"description":842},"2. 估计近期工作集","学习到的预测器可以对可能的下一个房间、纹理、几何体、着色器或神经表示进行排序。",{"label":844,"description":845},"3. 在可用带宽期间预取","在内存系统有余量时移动可能的资源，而不是等待硬性需求未命中。",{"label":847,"description":848},"4. 保持确定性保证","关键资产仍使用编写好的规则；预测是补充而非替代正确性。",{"label":850,"description":851},"5. 衡量未命中成本","如果预测浪费带宽或驱逐有用数据，置信度或预取深度必须降低。","一个预测性游戏内存示例",{},{"id":855,"data":856,"type":552,"tunes":859},"v3-learned-evidence",{"body":857,"title":858,"variant":559},"一项已发表的 CPU–GPU 统一虚拟内存研究报告称，使用学习到的预取后，\u003Cstrong>性能提升 10.89%\u003C\u002Fstrong>，页命中率从 \u003Cstrong>76.10% 升至 89.02%\u003C\u002Fstrong>，且 \u003Cstrong>CPU–GPU 互连流量降低 11.05%\u003C\u002Fstrong>。一篇 2026 年 AAAI 关于异步 KV 缓存预取的论文报告称，在其 LLM 工作负载上实现了 \u003Cstrong>2.15× 的注意力内核效率和高达 1.97× 的端到端吞吐量\u003C\u002Fstrong>。这两项结果都不应直接照搬到游戏上；它们都表明，预测和预取可以在真实 GPU 工作负载中隐藏内存成本。","这个想法有研究先例，但尚无游戏领域的证明",{},{"id":861,"data":862,"type":572,"tunes":864},"h-0024",{"text":863,"level":47},"为什么类 Spark 架构变得有趣",{},{"id":866,"data":867,"type":544,"tunes":869},"p-0025",{"text":868},"DGX Spark 不是游戏 PC，将其视为游戏 PC 会掩盖有用的教训。其 GB10 设计结合了 Arm CPU 核心、Blackwell GPU、Tensor 和 RT 硬件，以及高达 128 GB 的一致性统一系统内存。这种架构提出了与经典台式机不同的问题：不是将系统清晰地分为 CPU 内存和独立 VRAM，而是当多个计算引擎共享一个大内存池时会发生什么？",{},{"id":871,"data":872,"type":604,"tunes":894},"t-0026",{"content":873,"stretched":603,"withHeadings":15},[874,878,882,886,890],[875,876,877],"属性","DGX Spark 参考","如果该原则被采用，对游戏的影响",[879,880,881],"内存","高达 128 GB 一致性统一 LPDDR5X","大型共享池可以减少重复，并让本地 AI 和图形访问公共数据",[883,884,885],"内存带宽","273 GB\u002Fs","容量不够；高端图形仍需激进的带宽和延迟管理",[887,888,889],"计算组合","CPU + Blackwell GPU + Tensor + RT","未来的游戏工作负载可以拆分到更多专用引擎上",[891,892,893],"外形\u002F功耗","紧凑型 GB10 系统","集成可以提高效率，但可能降低模块化可升级性",{},{"id":896,"data":897,"type":552,"tunes":900},"c-0027",{"body":898,"title":899,"variant":611},"大型统一池\u003Cstrong>不会\u003C\u002Fstrong>自动取代高端独立 GPU 的带宽。DGX Spark 列出 273 GB\u002Fs 的内存带宽，而 NVIDIA 的 RTX PRO 6000 Blackwell 工作站版列出 1,792 GB\u002Fs。这些是非常不同的产品，比较并非游戏基准测试；它只是让架构约束变得可见。类似 Spark 的游戏设计需要解决争用、延迟和带宽问题——而不仅仅是提供更多共享千兆字节。","容量不等于带宽",{},{"id":902,"data":903,"type":572,"tunes":905},"h-0028",{"text":904,"level":47},"要实现这一转变，必须具备哪些条件？",{},{"id":907,"data":908,"type":644,"tunes":929},"flow-0029",{"steps":909,"title":928,"orientation":643},[910,913,916,919,922,925],{"label":911,"description":912},"1. 推理节省必须大于其成本","神经工作必须消除足够的存储、带宽或传统渲染成本，以证明其自身计算和延迟的合理性。",{"label":914,"description":915},"2. 引擎必须暴露有用的状态","当系统知道摄像机方向、资产依赖关系、移动和可能的可见性时，预测效果更好。",{"label":917,"description":918},"3. 内存移动必须变得有目的性","没有良好的驻留和预取策略，压缩仍可能导致停顿。",{"label":920,"description":921},"4. 质量必须保持稳定","生成的细节不能引入不可接受的伪影、时间不稳定或不正确的游戏信息。",{"label":923,"description":924},"5. 延迟必须有界","更漂亮的帧并不自动意味着更灵敏的游戏。",{"label":926,"description":927},"6. 技术栈必须经受住供应商差异","有用的技术需要不会分裂成一次性演示的 API、驱动和引擎路径。","神经游戏可行性测试",{},{"id":931,"data":932,"type":572,"tunes":934},"v3-reader-spark-expect",{"text":933,"level":47},"买家应该对类似 Spark 的转变有何期待？",{},{"id":936,"data":937,"type":707,"tunes":977},"v3-buyer-transition",{"rows":938,"title":969,"layout":604,"columns":970},[939,945,951,957,963],{"id":940,"label":941,"values":942},"pool","大型一致性池",{"cost":943,"gain":944},"共享客户端可能争用同一带宽","CPU、图形和本地 AI 之间更灵活的共享",{"id":946,"label":947,"values":948},"movement","减少重复",{"cost":949,"gain":950},"一致性和调度增加了复杂性","独立的系统内存和显存池之间更少的副本",{"id":952,"label":953,"values":954},"ai","更多本地 AI 余量",{"cost":955,"gain":956},"AI 可能与图形争夺计算和内存","更大的模型、代理或神经资产可以在本地共存",{"id":958,"label":959,"values":960},"efficiency","系统级优化",{"cost":961,"gain":962},"更紧密的集成可能降低可升级性并增加锁定","平台可以一起优化数据放置和加速器",{"id":964,"label":965,"values":966},"longevity","新的工作负载灵活性",{"cost":967,"gain":968},"只有当引擎和运行时实际使用这些能力时","如果未来的游戏转向神经工作负载，机器可能更耐用","潜在收益 vs 实际成本",[971,974],{"id":972,"label":973},"gain","潜在收益",{"id":975,"label":976},"cost","成本 \u002F 风险",{},{"id":979,"data":980,"type":572,"tunes":982},"h-0030",{"text":981,"level":47},"什么会证伪这一假设？",{},{"id":984,"data":985,"type":544,"tunes":987},"p-0031",{"text":986},"这一论点刻意可证伪。下表定义了市场、软件、硬件和游戏采用信号，这些信号会随着时间的推移使论点更强或更弱。",{},{"id":989,"data":990,"type":707,"tunes":1023},"cmp-0032",{"rows":991,"title":1015,"layout":604,"columns":1016},[992,997,1003,1009],{"id":682,"label":993,"values":994},"内存经济",{"weakens":995,"strengthens":996},"消费级容量扩大，价格压力正常化","服务器\u002FHBM 优先级继续挤压消费级分配",{"id":998,"label":999,"values":1000},"graphics","图形栈",{"weakens":1001,"strengthens":1002},"神经图形仍然是供应商特定的可选后处理","ML 原语成为常规引擎\u002FDirectX 功能",{"id":1004,"label":1005,"values":1006},"hardware","硬件设计",{"weakens":1007,"strengthens":1008},"独立 GPU + 独立内存 在成本、灵活性和延迟方面仍然明显优越","更多集成异构计算和更大的一致性内存池",{"id":1010,"label":1011,"values":1012},"games","游戏采用",{"weakens":1013,"strengthens":1014},"技术仍限于演示或少数旗舰游戏","工作室广泛发布神经压缩、生成和预测数据路径","加强或削弱论点的证据",[1017,1020],{"id":1018,"label":1019},"strengthens","会加强它",{"id":1021,"label":1022},"weakens","会削弱它",{},{"id":1025,"data":1026,"type":572,"tunes":1028},"h-0033",{"text":1027,"level":47},"当今游戏玩家的实际结论",{},{"id":1030,"data":1031,"type":544,"tunes":1033},"p-0034",{"text":1032},"不要因为推测性的架构可能出现而购买硬件。根据你实际拥有的游戏、分辨率、延迟目标和工作负载来购买。但不要假设面向未来仍然可以简化为一个数字，例如光栅性能或显存容量。软件栈的方向表明，内存行为、神经计算和引擎支持可能成为价值方程中越来越重要的部分。",{},{"id":1035,"data":1036,"type":669,"tunes":1041},"ref-0035",{"url":1037,"title":1038,"excerpt":1039,"ctaLabel":1040},"https:\u002F\u002Ffigure.rocks\u002Fzh\u002Fblog\u002Fvram-usage-is-not-vram-requirement-why-a-full-memory-meter-does-not-tell-the-whole-story","显存使用量不是显存需求量：为什么完整的内存计量器不能说明全部情况","实用解释工作集、驻留以及为什么分配的显存数字与硬性需求不是一回事。","阅读显存指南",{},{"id":1043,"data":1044,"type":572,"tunes":1046},"h-0036",{"text":1045,"level":47},"主要来源",{},{"id":1048,"data":1049,"type":1055,"tunes":1056},"v3-src-newzoo",{"link":1050,"meta":1051},"https:\u002F\u002Fnewzoo.com\u002Farticles\u002F2026-global-games-market-key-numbers",{"image":1052,"title":1053,"description":1054},{"url":13},"Newzoo — 2026 全球游戏市场关键数字","用于规模比较的 2026 年游戏市场预测。","linkTool",{},{"id":1058,"data":1059,"type":1055,"tunes":1065},"src-0038",{"link":1060,"meta":1061},"https:\u002F\u002Fwww.semiconductors.org\u002Fglobal-annual-semiconductor-sales-increase-25-6-to-791-7-billion-in-2025\u002F",{"image":1062,"title":1063,"description":1064},{"url":13},"半导体行业协会 — 2025 年全球销售额","2025 年半导体销售额以及该协会约 1 万亿美元的 2026 年预测。",{},{"id":1067,"data":1068,"type":1055,"tunes":1074},"v3-src-trend-sep",{"link":1069,"meta":1070},"https:\u002F\u002Fwww.trendforce.com\u002Fpresscenter\u002Fnews\u002F20260930-13258.html",{"image":1071,"title":1072,"description":1073},{"url":13},"TrendForce — AI 服务器需求维持 2026 年第四季度内存价格上涨","2026 年末的证据，表明服务器优先的产能分配持续存在以及消费级压力。",{},{"id":1076,"data":1077,"type":1055,"tunes":1083},"v3-src-buddy",{"link":1078,"meta":1079},"https:\u002F\u002Fresearch.nvidia.com\u002Fpublication\u002F2019-04_buddy-compression-enabling-larger-memory-deep-learning-and-hpc-workloads-gpus",{"image":1080,"title":1081,"description":1082},{"url":13},"NVIDIA 研究 — Buddy 压缩","HPC 平均压缩比 2.2 倍和 DL 平均压缩比 1.5 倍的研究来源。",{},{"id":1085,"data":1086,"type":1055,"tunes":1092},"v3-src-uvm",{"link":1087,"meta":1088},"https:\u002F\u002Fwww.sciencedirect.com\u002Fscience\u002Farticle\u002Fabs\u002Fpii\u002FS0743731522002490",{"image":1089,"title":1090,"description":1091},{"url":13},"基于深度学习的 CPU–GPU UVM 预取","学习页面预取结果的研究来源。",{},{"id":1094,"data":1095,"type":1055,"tunes":1101},"v3-src-aaai",{"link":1096,"meta":1097},"https:\u002F\u002Fojs.aaai.org\u002Findex.php\u002FAAAI\u002Farticle\u002Fview\u002F39224",{"image":1098,"title":1099,"description":1100},{"url":13},"AAAI 2026 — 异步 KV 缓存预取","研究来源，展示通过主动 GPU 缓存预取实现延迟隐藏。",{},{"id":1103,"data":1104,"type":1055,"tunes":1110},"src-0039",{"link":1105,"meta":1106},"https:\u002F\u002Fwww.trendforce.com\u002Fpresscenter\u002Fnews\u002F20260331-12995.html",{"image":1107,"title":1108,"description":1109},{"url":13},"TrendForce — AI 服务器需求与 2026 年第二季度内存价格","服务器\u002FHBM 优先化、GDDR 限制与消费端压力的当前证据。",{},{"id":1112,"data":1113,"type":1055,"tunes":1119},"src-0040",{"link":1114,"meta":1115},"https:\u002F\u002Fdevblogs.microsoft.com\u002Fdirectx\u002Fevolving-directx-for-the-ml-era-on-windows\u002F",{"image":1116,"title":1117,"description":1118},{"url":13},"Microsoft DirectX — 为 ML 时代演进 DirectX","GDC 2026 官方概述，介绍 DirectX 中着色器级和模型级 ML。",{},{"id":1121,"data":1122,"type":1055,"tunes":1128},"src-0041",{"link":1123,"meta":1124},"https:\u002F\u002Fresearch.nvidia.com\u002Flabs\u002Fadlr\u002FDLSS5\u002F",{"image":1125,"title":1126,"description":1127},{"url":13},"NVIDIA Research — DLSS 5：生成式神经渲染","主要研究页面，将 DLSS 5 描述为实时生成式渲染阶段。",{},{"id":1130,"data":1131,"type":1055,"tunes":1137},"src-0042",{"link":1132,"meta":1133},"https:\u002F\u002Fwww.nvidia.com\u002Fen-us\u002Fproducts\u002Fworkstations\u002Fdgx-spark\u002F",{"image":1134,"title":1135,"description":1136},{"url":13},"NVIDIA — DGX Spark 规格","官方架构、统一内存和带宽规格。",{},{"id":1139,"data":1140,"type":1055,"tunes":1146},"v5-src-5090-official",{"link":1141,"meta":1142},"https:\u002F\u002Fwww.nvidia.com\u002Fen-us\u002Fgeforce\u002Fgraphics-cards\u002F50-series\u002Frtx-5090\u002F",{"image":1143,"title":1144,"description":1145},{"url":13},"NVIDIA — GeForce RTX 5090","官方当前 RTX 5090 产品页面，用于区分已确认的产品状态与停产传闻。",{},{"id":1148,"data":1149,"type":1055,"tunes":1155},"v5-src-5090-signal",{"link":1150,"meta":1151},"https:\u002F\u002Fwww.tomshardware.com\u002Fpc-components\u002Fgpus\u002Fnvidia-rtx-5090-disappears-from-micro-center-amid-rumors-of-a-production-halt-retailer-flags-flagship-as-not-carried-as-stock-mysteriously-vanishes",{"image":1152,"title":1153,"description":1154},{"url":13},"Tom’s Hardware — RTX 5090 零售消失与停产传闻","当前市场报告，仅作为行业信号使用；报告明确指出 NVIDIA 尚未确认停产。",{},{"id":1157,"data":1158,"type":1055,"tunes":1164},"v5-src-ntc",{"link":1159,"meta":1160},"https:\u002F\u002Fresearch.nvidia.com\u002Flabs\u002Frtr\u002Fneural_texture_compression\u002F",{"image":1161,"title":1162,"description":1163},{"url":13},"NVIDIA Research — 材质纹理的随机访问神经压缩","主要来源，说明在内存比 BC High 少 30% 的情况下实现 4 倍线性分辨率 \u002F 16 倍纹素比较。",{},{"id":1166,"data":1167,"type":1055,"tunes":1173},"v5-src-rtxpro6000",{"link":1168,"meta":1169},"https:\u002F\u002Fwww.nvidia.com\u002Fde-de\u002Fproducts\u002Fworkstations\u002Fprofessional-desktop-gpus\u002Frtx-pro-6000\u002F",{"image":1170,"title":1171,"description":1172},{"url":13},"NVIDIA — RTX PRO 6000 Blackwell 工作站版","官方来源，提供 96 GB GDDR7 和 1,792 GB\u002Fs 内存带宽，用作带宽参考点。",{},{"id":1175,"data":1176,"type":1195,"tunes":1196},"gloss-0044",{"title":1177,"entries":1178},"关键概念",[1179,1183,1187,1191],{"term":1180,"anchor":1181,"definition":1182},"神经渲染","neural-rendering","在生成、重建或增强视觉输出的过程中使用训练模型作为一部分的渲染方法。",{"term":1184,"anchor":1185,"definition":1186},"异构计算","heterogeneous-compute","将不同工作负载分配给不同计算引擎（如 CPU、GPU、RT 和矩阵\u002FTensor 硬件）的系统。",{"term":1188,"anchor":1189,"definition":1190},"有效内存","effective-memory","在考虑压缩、驻留和移动效率后，对内存系统能够保留并供给计算的有用数据量的实际描述。它不是物理容量的正式替代。",{"term":1192,"anchor":1193,"definition":1194},"预测性内存","predictive-memory","Figure Rocks 的简写，指结合已知游戏状态和学习预测，在停顿发生前准备可能需要的数 据。它是一种假设，不是行业标准。","glossary",{},{"id":541,"data":1198,"type":544,"tunes":1200},{"text":1199},"几十年来，发烧级 PC 游戏遵循一个熟悉的交易：游戏要求更多计算，GPU 世代提供更多计算，玩家在差距足够大时升级。这个循环从未完美运作，但方向是可以理解的。到了 2026 年，这个循环变得越来越难以解读，因为两件事同时发生：AI 基础设施正在积极争夺芯片和内存，而 AI 技术正直接进入游戏创建帧的方式。",{},{"id":547,"data":1202,"type":552,"tunes":1204},{"body":1203,"title":550,"variant":551},"\u003Cstrong>旧的游戏升级周期正面临压力，但它并不是简单地被更昂贵的 GPU 所取代。\u003C\u002Fstrong> AI 正在改变等式的两边：它加剧了对半导体和内存资源的竞争，同时也使神经渲染技术成为可能，这些技术可以用推理来换取存储数据或传统渲染工作。重要的可能性是，游戏可能通过变得更具计算选择性来应对更紧张的硬件经济，而不仅仅是要求更大的 GPU。",{},{"id":555,"data":1206,"type":552,"tunes":1208},{"body":1207,"title":558,"variant":559},"这是一个\u003Cstrong>有证据支持的假设\u003C\u002Fstrong>。已记录的市场数据和已发布或宣布的技术被标记为事实或行业信号。它们之间的联系是技术推断。后面描述的未来架构是 Figure Rocks 的假设，不是关于未宣布的 NVIDIA、AMD、Intel、Microsoft 或游戏机路线图的声明。",{},{"id":562,"data":1210,"type":566,"tunes":1211},{"title":564,"maxLevel":565,"minLevel":47},{},{"id":569,"data":1213,"type":572,"tunes":1215},{"text":1214,"level":47},"规模不匹配已经可见",{},{"id":575,"data":1217,"type":544,"tunes":1219},{"text":1218},"Newzoo 预测 2026 年全球游戏市场将达到 2139 亿美元，同比增长 6.1%。半导体行业协会报告 2025 年全球半导体销售额为 7917 亿美元，增长 25.6%，并预测 2026 年约为 1 万亿美元。游戏业务和半导体业务仍然相互依赖，但它们的激励正变得越来越不对称：同一个使能高端游戏的先进制造生态系统，现在服务于拥有更大基础设施预算的 AI 和数据中心买家。这些不是直接可比的产品类别；它们确立了游戏周围上游计算经济的规模。",{},{"id":580,"data":1221,"type":604,"tunes":1240},{"content":1222,"stretched":603,"withHeadings":15},[1223,1224,1228,1232,1236],[584,585,586],[1225,1226,1227],"全球游戏市场，2026 年预测","2139 亿美元，同比增长 6.1%","大型消费娱乐市场，但处于先进硬件供应的下游",[1229,1230,1231],"PC 游戏，2026 年预测","459 亿美元，同比增长 5.3%","最直接受 PC 组件经济影响的游戏细分市场",[1233,1234,1235],"全球半导体销售额，2025 年","7917 亿美元，同比增长 25.6%","显示服务于游戏、AI、移动、汽车和云的上游产业的规模",[1237,1238,1239],"全球半导体产业，2026 年预测","约 1 万亿美元","提高了将高端芯片仅专用于消费游戏的机 会成本",{},{"id":607,"data":1242,"type":552,"tunes":1244},{"body":1243,"title":610,"variant":611},"这些数字确立了\u003Cstrong>经济背景，而非直接因果关系\u003C\u002Fstrong>。它们不能证明某个特定的晶圆、GPU 或内存封装从玩家手中被拿走并重新分配给 AI。它们也不意味着游戏和半导体是世界上最大的两个产业。更狭窄、更有说服力的观点是，一个 2000 亿美元以上的游戏市场依赖于年规模接近万亿美元的上游半导体经济，其中 AI 基础设施现在在非常不同的预算和利润条件下竞争先进计算和内存。",{},{"id":614,"data":1246,"type":572,"tunes":1247},{"text":616,"level":47},{},{"id":619,"data":1249,"type":544,"tunes":1251},{"text":1250},"TrendForce多次报告称，随着AI需求增长，DRAM供应商正将产能转向服务器和HBM产品。2026年3月，它预测第二季度传统DRAM合约价格环比上涨58–63%，并表示有限的GDDR分配正在制约图形内存供应。到第三季度，TrendForce表示，尽管AI和服务器需求依然强劲，消费类客户正达到可负担性极限。",{},{"id":624,"data":1253,"type":644,"tunes":1263},{"steps":1254,"title":642,"orientation":643},[1255,1257,1259,1260,1261],{"label":628,"description":1256},"云提供商和AI建设者获取服务器内存、HBM、企业级SSD和先进计算资源。",{"label":631,"description":1258},"产能转向需求更强、利润率更高或拥有长期合同的产品。",{"label":634,"description":635},{"label":637,"description":638},{"label":640,"description":1262},"更快的GPU仍然有用，但在相同的消费价格点上，暴力堆料式扩展在经济上越来越难以重复。",{},{"id":647,"data":1265,"type":552,"tunes":1268},{"body":1266,"title":1267,"variant":611},"TrendForce于2026年9月30日报道称，供应商仍优先将先进制程产能用于高性能服务器产品，预计第四季度传统DRAM合约价格将再环比上涨10–15%。其DRAM公告还表示，HBM继续挤占通用DRAM产能。因此，这里描述的压力在2026年末依然可见；它并非仅基于年初的暂时性飙升。","2026年9月的信号，而非旧短缺故事",{},{"id":653,"data":1270,"type":572,"tunes":1272},{"text":1271,"level":47},"与此同时，帧本身也在改变",{},{"id":658,"data":1274,"type":544,"tunes":1276},{"text":1275},"微软现在将机器学习描述为正在成为实时图形的基础。DirectX正在添加着色器级和模型级的ML路径，使神经纹理压缩、降噪、超分辨率和其他学习型工作负载能够融入图形栈，而不是作为孤立功能附加。NVIDIA的DLSS 5研究通过引入生成式神经渲染阶段进一步推进了这一理念，该阶段用学习到的外观先验补充传统渲染。神经渲染并不意味着传统渲染消失。新兴的模式是混合式的：直接渲染一些信息，在更便宜或更有用时推断或重建其他信息。",{},{"id":663,"data":1278,"type":669,"tunes":1280},{"url":665,"title":1279,"excerpt":667,"ctaLabel":668},"DLSS 5不只是超分辨率：生成式神经渲染真正改变了什么",{},{"id":672,"data":1282,"type":707,"tunes":1298},{"rows":1283,"title":1294,"layout":604,"columns":1295},[1284,1287,1290,1292],{"id":676,"label":677,"values":1285},{"new":1286,"old":680},"选择性存储\u002F渲染，重建或推断部分细节",{"id":682,"label":683,"values":1288},{"new":1289,"old":686},"在需要时增加容量，但也压缩并减少移动",{"id":688,"label":689,"values":1291},{"new":691,"old":692},{"id":694,"label":695,"values":1293},{"new":697,"old":698},"传统缩放 vs 神经时代缩放",[1296,1297],{"id":702,"label":703},{"id":705,"label":706},{},{"id":710,"data":1300,"type":552,"tunes":1302},{"body":1301,"title":713,"variant":714},"\u003Cstrong>AI可能正在削弱暴力式游戏硬件的经济性，同时提供使暴力式缩放变得不那么必要的技术。\u003C\u002Fstrong>这并不意味着未来的游戏需要更少的硬件。它意味着更多结果可能来自决定存储什么、移动什么、渲染什么、重建什么或推断什么。Figure Rocks的简写：\u003Cstrong>少存储 → 少移动 → 多预测 → 少渲染 → 多推断。\u003C\u002Fstrong>这是一个关于性能提升可能越来越多来自何处的假设，而不是声称未来每款游戏都会使用单一管线。",{},{"id":717,"data":1304,"type":719,"tunes":1305},{},{},{"id":722,"data":1307,"type":572,"tunes":1309},{"text":1308,"level":47},"只有正确表述，5090的问题才有意义",{},{"id":727,"data":1311,"type":544,"tunes":1313},{"text":1312},"RTX 5090是一个有用的例子，恰恰因为证据是混合的。截至2026年10月18日，NVIDIA仍将GeForce RTX 5090列为当前产品。与此同时，零售商行为和媒体报道记录了异常糟糕的可用性以及未经证实的停产说法。这是一个市场信号，而非官方的停产公告。因此，站得住脚的问题不是NVIDIA是否秘密杀死了一款显卡，而是当类似架构可以服务于经济性截然不同的专业可视化和AI市场时，将顶级Blackwell级硅片分配给发烧级游戏的机会成本是否上升。",{},{"id":732,"data":1315,"type":552,"tunes":1317},{"body":1316,"title":735,"variant":559},"\u003Cstrong>事实：\u003C\u002Fstrong>NVIDIA仍公开列出RTX 5090，且Blackwell架构横跨游戏和专业产品。\u003Cbr>\u003Cstrong>行业信号：\u003C\u002Fstrong>零售商下架、极度稀缺和可能停产的报道存在，但NVIDIA尚未确认RTX 5090的EOL或停产。\u003Cbr>\u003Cstrong>技术推断：\u003C\u002Fstrong>高端芯片有多个具有经济价值的目的地。\u003Cbr>\u003Cstrong>FIGURE ROCKS假设：\u003C\u002Fstrong>AI和专业需求可能使传统的发烧友期望——即最大的实用芯片仍能作为游戏旗舰广泛获得——更难维持。",{},{"id":738,"data":1319,"type":572,"tunes":1321},{"text":1320,"level":47},"内存正开始成为计算问题，而不仅仅是容量问题",{},{"id":743,"data":1323,"type":544,"tunes":1325},{"text":1324},"至少有两种方法可以在不单纯提高原始DRAM速度的情况下攻击内存瓶颈：压缩表示、用计算重建细节、预测接下来需要什么数据，以及将正确的数据保持在执行单元附近。这些都不会让物理内存芯片神奇地变快。它们改变了系统从有限容量和带宽预算中获得多少有用工作。",{},{"id":748,"data":1327,"type":604,"tunes":1339},{"content":1328,"stretched":603,"withHeadings":15},[1329,1330,1333,1336,1337],[752,753,754],[756,1331,1332],"更少的物理字节表示相同的逻辑数据","DRAM时钟和总线宽度",[760,1334,1335],"部分细节从紧凑表示生成","成本消失；推断有其自身的计算\u002F延迟成本",[764,765,766],[1338,769,770],"驻留\u002F统一访问",{},{"id":773,"data":1341,"type":552,"tunes":1344},{"body":1342,"title":1343,"variant":551},"NVIDIA Research的Buddy Compression报告称，HPC工作负载平均压缩比为\u003Cstrong>2.2倍，深度学习工作负载为1.5倍\u003C\u002Fstrong>，该研究中减速为1–2%。该结果不是游戏基准测试。它在此处的相关性是架构性的：压缩内存可以增加有效容量并减少部分流量，而无需改变物理DRAM时钟。","研究锚点：有效容量不是原始速度",{},{"id":779,"data":1346,"type":552,"tunes":1348},{"body":1347,"title":782,"variant":714},"NVIDIA Research的材质纹理NTC工作证明，与BC High比较，\u003Cstrong>线性纹理分辨率提高4倍——纹素多16倍——同时内存使用减少30%\u003C\u002Fstrong>。这并不意味着每款游戏都能获得16倍的内存增益。它展示了这里重要的机制：一些视觉细节可以更紧凑地表示，并通过计算重建，而不是以相同保真度传统存储。",{},{"id":785,"data":1350,"type":552,"tunes":1353},{"body":1351,"title":1352,"variant":611},"有时与NVIDIA系统关联的460×数字，指的是跨DGX代际的GPU可访问内存容量比较，\u003Cstrong>而非460倍内存压缩比\u003C\u002Fstrong>。本文刻意不将其作为压缩的证据。","“460×”这个数字并不代表什么",{},{"id":791,"data":1355,"type":572,"tunes":1357},{"text":1356,"level":47},"技术响应可能如何运作",{},{"id":796,"data":1359,"type":644,"tunes":1372},{"steps":1360,"title":816,"orientation":643},[1361,1363,1365,1367,1369,1370],{"label":756,"description":1362},"在质量允许的情况下，保持纹理或学习到的表示更小。",{"label":802,"description":1364},"使用确定性游戏状态和学习模型来估计接下来需要哪些数据。",{"label":805,"description":1366},"在GPU因等待数据而停滞之前，提前移动有用的数据。",{"label":808,"description":1368},"将传统图形工作集中在最能改变可见结果的地方。",{"label":811,"description":812},{"label":814,"description":1371},"在保持延迟、伪影和帧节奏在可接受范围内的同时交付帧。",{},{"id":819,"data":1374,"type":544,"tunes":1376},{"text":1375},"这条管线并非某个已发布的产品。每个组件都已以某种形式存在，但整合后的链条仍是假设。神经纹理压缩已表明，学习到的表示可以用计算换取内存。GPU预取研究表明，学习到的预测可以提高内存页命中率。DirectX正在将机器学习形式化到图形编程模型中。DLSS 5显示生成阶段正在进入实时渲染。",{},{"id":824,"data":1378,"type":572,"tunes":1380},{"text":1379,"level":47},"游戏拥有通用工作负载通常不具备的一个优势：它已经知道部分未来",{},{"id":829,"data":1382,"type":544,"tunes":1384},{"text":1383},"游戏引擎知道相机变换、玩家速度、当前世界单元、相邻流送单元、动画状态、活动实体、LOD过渡、着色器依赖关系，以及通常可能的遍历路径。这种确定性上下文并不能完美揭示未来，但它为预测器提供了比仅靠通用内存跟踪更丰富的起点。",{},{"id":834,"data":1386,"type":644,"tunes":1398},{"steps":1387,"title":852,"orientation":643},[1388,1390,1392,1394,1396],{"label":838,"description":1389},"相机方向、移动向量、当前房间、可见传送门和已知资产依赖关系已存在于引擎状态中。",{"label":841,"description":1391},"学习到的预测器可以对可能的下一房间、纹理、几何体、着色器或神经表示进行排序。",{"label":844,"description":1393},"在内存系统有余量时移动可能的资源，而不是等待硬性需求缺失。",{"label":847,"description":1395},"关键资产仍使用编写规则；预测补充而非替代正确性。",{"label":1397,"description":851},"5. 衡量缺失成本",{},{"id":855,"data":1400,"type":552,"tunes":1403},{"body":1401,"title":1402,"variant":559},"一项已发表的CPU–GPU统一虚拟内存研究报告称，使用学习到的预取实现了\u003Cstrong>10.89%的性能提升\u003C\u002Fstrong>，页命中率从\u003Cstrong>76.10%升至89.02%\u003C\u002Fstrong>，且\u003Cstrong>CPU–GPU互连流量降低11.05%\u003C\u002Fstrong>。2026年AAAI关于异步KV缓存预取的论文在其LLM工作负载上报告了\u003Cstrong>2.15倍注意力内核效率和高达1.97倍的端到端吞吐量\u003C\u002Fstrong>。这两项结果都不应直接照搬到游戏上；它们都表明预测和预取可以在真实GPU工作负载中隐藏内存成本。","这一想法有研究先例，但尚无游戏领域的证明",{},{"id":861,"data":1405,"type":572,"tunes":1407},{"text":1406,"level":47},"为什么类Spark架构变得有趣",{},{"id":866,"data":1409,"type":544,"tunes":1411},{"text":1410},"DGX Spark不是游戏PC，将其视为游戏PC会掩盖有用的教训。其GB10设计结合了Arm CPU核心、Blackwell GPU、Tensor和RT硬件，以及高达128 GB的一致性统一系统内存。这种架构提出了与经典台式机不同的问题：不是将系统清晰地分为CPU内存和独立VRAM，而是当多个计算引擎共享一个大内存池时会发生什么？",{},{"id":871,"data":1413,"type":604,"tunes":1427},{"content":1414,"stretched":603,"withHeadings":15},[1415,1418,1421,1422,1423],[875,1416,1417],"DGX Spark参考","若该原则被适配，对游戏的影响",[879,1419,1420],"高达128 GB一致性统一LPDDR5X","大型共享池可减少重复，并让本地AI和图形访问公共数据",[883,884,885],[887,888,889],[1424,1425,1426],"形态\u002F功耗","紧凑型GB10系统","集成可提高效率，但可能降低模块化可升级性",{},{"id":896,"data":1429,"type":552,"tunes":1431},{"body":1430,"title":899,"variant":611},"大型统一池\u003Cstrong>不会\u003C\u002Fstrong>自动替代高端独立GPU的带宽。DGX Spark列出273 GB\u002Fs的内存带宽，而NVIDIA的RTX PRO 6000 Blackwell工作站版列出1,792 GB\u002Fs。这些是非常不同的产品，比较并非游戏基准测试；它只是让架构约束变得可见。类Spark的游戏设计需要解决争用、延迟和带宽问题——而不仅仅是提供更多共享千兆字节。",{},{"id":902,"data":1433,"type":572,"tunes":1435},{"text":1434,"level":47},"这一转变要成功，必须满足哪些条件？",{},{"id":907,"data":1437,"type":644,"tunes":1451},{"steps":1438,"title":928,"orientation":643},[1439,1440,1442,1445,1446,1448],{"label":911,"description":912},{"label":914,"description":1441},"当系统知道相机方向、资产依赖关系、移动和可能的可见性时，预测效果更好。",{"label":1443,"description":1444},"3. 内存移动必须变得有意识","没有良好的驻留和预取策略，压缩仍可能导致停滞。",{"label":920,"description":921},{"label":923,"description":1447},"更漂亮的帧并不自动意味着更响应灵敏的游戏。",{"label":1449,"description":1450},"6. 技术栈必须经受供应商差异","有用的技术需要API、驱动和引擎路径，不能碎片化为一次性演示。",{},{"id":931,"data":1453,"type":572,"tunes":1455},{"text":1454,"level":47},"买家应从类Spark转变中期待什么？",{},{"id":936,"data":1457,"type":707,"tunes":1479},{"rows":1458,"title":969,"layout":604,"columns":1475},[1459,1462,1466,1470,1472],{"id":940,"label":941,"values":1460},{"cost":943,"gain":1461},"CPU、图形和本地AI之间更灵活的共享",{"id":946,"label":947,"values":1463},{"cost":1464,"gain":1465},"一致性和调度增加复杂性","独立系统RAM和VRAM池之间的副本更少",{"id":952,"label":1467,"values":1468},"更多本地AI余量",{"cost":1469,"gain":956},"AI可能与图形争夺计算和内存",{"id":958,"label":959,"values":1471},{"cost":961,"gain":962},{"id":964,"label":965,"values":1473},{"cost":967,"gain":1474},"如果未来游戏转向神经工作负载，机器可能更耐用",[1476,1477],{"id":972,"label":973},{"id":975,"label":1478},"成本\u002F风险",{},{"id":979,"data":1481,"type":572,"tunes":1482},{"text":981,"level":47},{},{"id":984,"data":1484,"type":544,"tunes":1486},{"text":1485},"这一论点刻意可证伪。下表定义了市场、软件、硬件和游戏采用信号，这些信号会随时间使论点更强或更弱。",{},{"id":989,"data":1488,"type":707,"tunes":1514},{"rows":1489,"title":1508,"layout":604,"columns":1509},[1490,1495,1500,1504],{"id":682,"label":1491,"values":1492},"内存经济学",{"weakens":1493,"strengthens":1494},"消费者产能扩大，价格压力趋于正常","服务器\u002FHBM 优先级持续挤压消费级分配",{"id":998,"label":1496,"values":1497},"图形技术栈",{"weakens":1498,"strengthens":1499},"神经图形仍为厂商特定的可选后处理","机器学习原语成为常规引擎\u002FDirectX 特性",{"id":1004,"label":1005,"values":1501},{"weakens":1502,"strengthens":1503},"独立 GPU + 独立内存仍在成本、灵活性和延迟上明显占优","更集成的异构计算和更大的连贯内存池",{"id":1010,"label":1011,"values":1505},{"weakens":1506,"strengthens":1507},"技术仍局限于演示或少数旗舰游戏","工作室广泛部署神经压缩、生成和预测数据路径","强化或削弱论点的证据",[1510,1512],{"id":1018,"label":1511},"会强化论点",{"id":1021,"label":1513},"会削弱论点",{},{"id":1025,"data":1516,"type":572,"tunes":1518},{"text":1517,"level":47},"当今玩家的实用结论",{},{"id":1030,"data":1520,"type":544,"tunes":1522},{"text":1521},"不要因为某种推测性架构可能到来而购买硬件。应根据你实际拥有的游戏、分辨率、延迟目标和工作负载来购买。但不要假设面向未来的考量仍能简化为光栅性能或显存容量这样的单一数字。软件栈的发展方向表明，内存行为、神经计算和引擎支持可能日益成为价值方程中的重要部分。",{},{"id":1035,"data":1524,"type":669,"tunes":1527},{"url":1037,"title":1525,"excerpt":1526,"ctaLabel":1040},"显存使用量不等于显存需求：为什么显存占用条满并不说明全部问题","实用解释工作集、驻留，以及为什么已分配的显存数值与硬性需求不是一回事。",{},{"id":1043,"data":1529,"type":572,"tunes":1530},{"text":1045,"level":47},{},{"id":1048,"data":1532,"type":1055,"tunes":1536},{"link":1050,"meta":1533},{"image":1534,"title":1535,"description":1054},{"url":13},"Newzoo — 2026 全球游戏市场关键数据",{},{"id":1058,"data":1538,"type":1055,"tunes":1542},{"link":1060,"meta":1539},{"image":1540,"title":1063,"description":1541},{"url":13},"2025 年半导体销售额及该协会约 1 万亿美元的 2026 年预测。",{},{"id":1067,"data":1544,"type":1055,"tunes":1549},{"link":1069,"meta":1545},{"image":1546,"title":1547,"description":1548},{"url":13},"TrendForce — AI 服务器需求支撑 2026 年第四季度内存价格上涨","2026 年末关于服务器优先产能分配持续及消费端压力的证据。",{},{"id":1076,"data":1551,"type":1055,"tunes":1556},{"link":1078,"meta":1552},{"image":1553,"title":1554,"description":1555},{"url":13},"NVIDIA Research — Buddy Compression","HPC 平均压缩比 2.2 倍、DL 平均压缩比 1.5 倍的研究来源。",{},{"id":1085,"data":1558,"type":1055,"tunes":1562},{"link":1087,"meta":1559},{"image":1560,"title":1090,"description":1561},{"url":13},"学习型页面预取结果的研究来源。",{},{"id":1094,"data":1564,"type":1055,"tunes":1568},{"link":1096,"meta":1565},{"image":1566,"title":1099,"description":1567},{"url":13},"通过主动 GPU 缓存预取实现延迟隐藏的研究来源。",{},{"id":1103,"data":1570,"type":1055,"tunes":1574},{"link":1105,"meta":1571},{"image":1572,"title":1108,"description":1573},{"url":13},"关于服务器\u002FHBM 优先、GDDR 限制和消费端压力的当前证据。",{},{"id":1112,"data":1576,"type":1055,"tunes":1580},{"link":1114,"meta":1577},{"image":1578,"title":1117,"description":1579},{"url":13},"GDC 2026 关于 DirectX 中着色器级和模型级 ML 的官方概述。",{},{"id":1121,"data":1582,"type":1055,"tunes":1586},{"link":1123,"meta":1583},{"image":1584,"title":1126,"description":1585},{"url":13},"将 DLSS 5 描述为实时生成式渲染阶段的主要研究页面。",{},{"id":1130,"data":1588,"type":1055,"tunes":1591},{"link":1132,"meta":1589},{"image":1590,"title":1135,"description":1136},{"url":13},{},{"id":1139,"data":1593,"type":1055,"tunes":1596},{"link":1141,"meta":1594},{"image":1595,"title":1144,"description":1145},{"url":13},{},{"id":1148,"data":1598,"type":1055,"tunes":1602},{"link":1150,"meta":1599},{"image":1600,"title":1153,"description":1601},{"url":13},"仅作为行业信号的当前市场报道；该报道明确指出 NVIDIA 尚未确认停产。",{},{"id":1157,"data":1604,"type":1055,"tunes":1608},{"link":1159,"meta":1605},{"image":1606,"title":1162,"description":1607},{"url":13},"关于相比 BC High 内存减少 30% 时实现 4 倍线性分辨率 \u002F 16 倍纹素比较的主要来源。",{},{"id":1166,"data":1610,"type":1055,"tunes":1614},{"link":1168,"meta":1611},{"image":1612,"title":1171,"description":1613},{"url":13},"96 GB GDDR7 和 1,792 GB\u002Fs 内存带宽的官方来源，用作带宽参考点。",{},{"id":1175,"data":1616,"type":1195,"tunes":1625},{"title":1177,"entries":1617},[1618,1620,1621,1623],{"term":1180,"anchor":1181,"definition":1619},"使用训练好的模型作为生成、重建或增强视觉输出的一部分的渲染方法。",{"term":1184,"anchor":1185,"definition":1186},{"term":1188,"anchor":1189,"definition":1622},"在考虑压缩、驻留和移动效率后，对内存系统能够保留并供给计算的有用数据量的实际描述。它不是对物理容量的正式替代。",{"term":1192,"anchor":1193,"definition":1624},"Figure Rocks 的简写，指结合已知游戏状态和学习预测，在停顿发生前准备可能需要的数​​据。这是一种假设，并非行业标准。",{},"2.31","人工智能正使高性能芯片和存储器在游戏之外更具价值，同时也为游戏提供了新的渲染、压缩和预测方式。本分析探讨这两种趋势是否会汇聚成一种新的硬件模式。","\u002Fuploads\u002F2026\u002F10\u002Frender-less-infer-more-is-ai-breaking-the-traditional-gaming-hardware-upgrade-cycle-1792323493963-l2x8qb.webp","render-less-infer-more-is-ai-breaking-the-traditional-gaming-hardware-upgrade-cycle-1792323493963-l2x8qb","PUBLISHED","2026-10-18T03:34:00.000Z","2026-10-18T11:34:48.220Z","2026-10-18T15:26:55.786Z",{"en":1635,"de":1636,"sr":1637,"es":1638,"fr":1639,"it":1640,"ru":1641,"zh":1642},"\u002Fblog\u002Frender-less-infer-more-is-ai-breaking-the-traditional-gaming-hardware-upgrade-cycle","\u002Fde\u002Fblog\u002Frender-less-infer-more-is-ai-breaking-the-traditional-gaming-hardware-upgrade-cycle","\u002Fsr\u002Fblog\u002Frender-less-infer-more-is-ai-breaking-the-traditional-gaming-hardware-upgrade-cycle","\u002Fes\u002Fblog\u002Frender-less-infer-more-is-ai-breaking-the-traditional-gaming-hardware-upgrade-cycle","\u002Ffr\u002Fblog\u002Frender-less-infer-more-is-ai-breaking-the-traditional-gaming-hardware-upgrade-cycle","\u002Fit\u002Fblog\u002Frender-less-infer-more-is-ai-breaking-the-traditional-gaming-hardware-upgrade-cycle","\u002Fru\u002Fblog\u002Frender-less-infer-more-is-ai-breaking-the-traditional-gaming-hardware-upgrade-cycle","\u002Fzh\u002Fblog\u002Frender-less-infer-more-is-ai-breaking-the-traditional-gaming-hardware-upgrade-cycle",[1644,1647,1651,1655,1659],{"id":1645,"name":291,"slug":1646},144,"tech",{"id":1648,"name":1649,"slug":1650},166,"游戏新闻","gaming-news",{"id":1652,"name":1653,"slug":1654},165,"新闻","news",{"id":1656,"name":1657,"slug":1658},170,"科技新闻","tech-news",{"id":1660,"name":1661,"slug":705},427,"新品",{"id":283,"login":1663,"email":13,"displayName":1664},"aleksandar","Aleksandar Stajic",[1666,2517],{"lang":8,"title":1667,"content":1668,"contentJson":1669,"excerpt":2516},"Render Less, Infer More: Is AI Breaking the Traditional Gaming Hardware Upgrade Cycle?","{\"time\":1792336677504,\"blocks\":[{\"id\":\"p-0001\",\"type\":\"paragraph\",\"data\":{\"text\":\"For decades, enthusiast PC gaming followed a familiar bargain: games asked for more compute, GPU generations delivered more compute, and players upgraded when the gap became large enough. That loop never worked perfectly, but the direction was understandable. In 2026, the loop is becoming harder to read because two changes are happening at the same time: AI infrastructure is competing aggressively for silicon and memory, while AI techniques are moving directly into the way games create frames.\"},\"tunes\":{}},{\"id\":\"c-0002\",\"type\":\"callout\",\"data\":{\"variant\":\"info\",\"title\":\"Direct answer\",\"body\":\"\u003Cstrong>The old gaming upgrade cycle is under pressure, but it is not simply being replaced by more expensive GPUs.\u003C\u002Fstrong> AI is changing both sides of the equation: it increases competition for semiconductor and memory resources, and it also enables neural rendering techniques that can trade stored data or conventional rendering work for inference. The important possibility is that gaming may respond to tighter hardware economics by becoming more computationally selective, not merely by demanding larger GPUs.\"},\"tunes\":{}},{\"id\":\"c-0003\",\"type\":\"callout\",\"data\":{\"variant\":\"note\",\"title\":\"How to read this article\",\"body\":\"This is an \u003Cstrong>evidence-backed hypothesis\u003C\u002Fstrong>. Documented market data and shipping or announced technologies are labeled as facts or industry signals. The connections between them are technical inference. The future architecture described later is a Figure Rocks hypothesis, not a claim about an unannounced NVIDIA, AMD, Intel, Microsoft or console roadmap.\"},\"tunes\":{}},{\"id\":\"toc-0004\",\"type\":\"tableOfContents\",\"data\":{\"title\":\"Contents\",\"minLevel\":2,\"maxLevel\":3},\"tunes\":{}},{\"id\":\"h-0005\",\"type\":\"header\",\"data\":{\"text\":\"The scale mismatch is already visible\",\"level\":2},\"tunes\":{}},{\"id\":\"p-0006\",\"type\":\"paragraph\",\"data\":{\"text\":\"Newzoo forecasts the global games market at $213.9 billion in 2026, up 6.1% year over year. The Semiconductor Industry Association reported $791.7 billion in global semiconductor sales for 2025, up 25.6%, and projected roughly $1 trillion for 2026. The games business and semiconductor business still depend on each other, but their incentives are becoming less symmetric: the same advanced manufacturing ecosystem that enables premium gaming now serves AI and data-center buyers with much larger infrastructure budgets. These are not directly comparable product categories; they establish the scale of the upstream compute economy around gaming.\"},\"tunes\":{}},{\"id\":\"t-0007\",\"type\":\"table\",\"data\":{\"withHeadings\":true,\"stretched\":false,\"content\":[[\"Reference point\",\"Reported or forecast value\",\"Why it matters here\"],[\"Global games market, 2026 forecast\",\"$213.9B, +6.1% YoY\",\"Large consumer entertainment market, but downstream of advanced hardware supply\"],[\"PC games, 2026 forecast\",\"$45.9B, +5.3% YoY\",\"The gaming segment most directly exposed to PC component economics\"],[\"Global semiconductor sales, 2025\",\"$791.7B, +25.6% YoY\",\"Shows the scale of the upstream industry serving gaming, AI, mobile, automotive and cloud\"],[\"Global semiconductor industry, 2026 projection\",\"~$1T\",\"Raises the opportunity cost of dedicating premium silicon only to consumer gaming\"]]},\"tunes\":{}},{\"id\":\"c-0008\",\"type\":\"callout\",\"data\":{\"variant\":\"warning\",\"title\":\"What these market numbers do — and do not — prove\",\"body\":\"The figures establish \u003Cstrong>economic context, not direct causation\u003C\u002Fstrong>. They do not prove that a specific wafer, GPU or memory package was taken from a gamer and reassigned to AI. They also do not mean gaming and semiconductors are the two largest industries in the world. The narrower, defensible point is that a $200B+ games market depends on an upstream semiconductor economy approaching trillion-dollar annual scale, where AI infrastructure now competes for advanced compute and memory under very different budget and margin conditions.\"},\"tunes\":{}},{\"id\":\"h-0009\",\"type\":\"header\",\"data\":{\"text\":\"Memory is where the pressure becomes concrete\",\"level\":2},\"tunes\":{}},{\"id\":\"p-0010\",\"type\":\"paragraph\",\"data\":{\"text\":\"TrendForce has repeatedly reported that DRAM suppliers are shifting capacity toward server and HBM products as AI demand grows. In March 2026 it forecast conventional DRAM contract prices rising 58–63% quarter over quarter in Q2 and said limited GDDR allocation was constraining graphics memory supply. By Q3, TrendForce said consumer customers were reaching an affordability limit even as AI and server demand remained strong.\"},\"tunes\":{}},{\"id\":\"flow-0011\",\"type\":\"processFlow\",\"data\":{\"title\":\"The resource-allocation pressure\",\"orientation\":\"auto\",\"steps\":[{\"label\":\"1. AI infrastructure expands\",\"description\":\"Cloud providers and AI builders secure server memory, HBM, enterprise SSDs and advanced compute.\"},{\"label\":\"2. Suppliers optimize product mix\",\"description\":\"Capacity moves toward products with stronger demand, margins or long-term contracts.\"},{\"label\":\"3. Consumer allocation tightens\",\"description\":\"PC, graphics and consumer memory buyers face higher prices or constrained availability.\"},{\"label\":\"4. Gaming vendors absorb or pass through cost\",\"description\":\"Board design, VRAM capacity and system price become harder trade-offs.\"},{\"label\":\"5. The old upgrade cycle weakens\",\"description\":\"A faster GPU is still useful, but brute-force scaling becomes economically harder to repeat at the same consumer price points.\"}]},\"tunes\":{}},{\"id\":\"v3-sep-signal\",\"type\":\"callout\",\"data\":{\"variant\":\"warning\",\"title\":\"A September 2026 signal, not an old shortage story\",\"body\":\"TrendForce reported on September 30, 2026 that suppliers still prioritize advanced-process capacity for high-performance server products, with conventional DRAM contract prices projected to rise another 10–15% QoQ in Q4. Its DRAM bulletin also says HBM continues to crowd out general DRAM capacity. The pressure described here therefore remains visible late in 2026; it is not based only on a temporary early-year spike.\"},\"tunes\":{}},{\"id\":\"h-0012\",\"type\":\"header\",\"data\":{\"text\":\"At exactly the same time, the frame itself is changing\",\"level\":2},\"tunes\":{}},{\"id\":\"p-0013\",\"type\":\"paragraph\",\"data\":{\"text\":\"Microsoft now describes machine learning as becoming foundational to real-time graphics. DirectX is adding shader-level and model-level ML paths so neural texture compression, denoising, super resolution and other learned workloads can sit inside the graphics stack instead of being bolted on as isolated features. NVIDIA's DLSS 5 research pushes the idea further by introducing a generative neural rendering stage that complements conventional rendering with learned appearance priors. Neural rendering does not mean conventional rendering disappears. The emerging model is hybrid: render some information directly, infer or reconstruct other information when doing so is cheaper or more useful.\"},\"tunes\":{}},{\"id\":\"ref-0014\",\"type\":\"referralArticle\",\"data\":{\"url\":\"https:\u002F\u002Ffigure.rocks\u002Fblog\u002Fdlss-5-is-not-just-upscaling-what-3d-guided-neural-rendering-actually-changes\",\"title\":\"DLSS 5 Is Not Just Upscaling: What Generative Neural Rendering Actually Changes\",\"excerpt\":\"Why the important change is not another quality mode, but a new stage in how visual information can be produced.\",\"ctaLabel\":\"Read the DLSS 5 analysis\"},\"tunes\":{}},{\"id\":\"cmp-0015\",\"type\":\"comparison\",\"data\":{\"title\":\"Traditional scaling vs neural-era scaling\",\"layout\":\"table\",\"columns\":[{\"id\":\"old\",\"label\":\"Traditional path\"},{\"id\":\"new\",\"label\":\"Neural-era path\"}],\"rows\":[{\"id\":\"fidelity\",\"label\":\"More visual fidelity\",\"values\":{\"old\":\"Store more detail and render more work\",\"new\":\"Store\u002Frender selectively, reconstruct or infer some detail\"}},{\"id\":\"memory\",\"label\":\"Memory response\",\"values\":{\"old\":\"Add capacity and bandwidth\",\"new\":\"Add capacity where needed, but also compress and reduce movement\"}},{\"id\":\"performance\",\"label\":\"Performance response\",\"values\":{\"old\":\"More shader\u002FRT throughput\",\"new\":\"Mix shader, RT and matrix\u002FTensor-style compute\"}},{\"id\":\"software\",\"label\":\"Software role\",\"values\":{\"old\":\"Engine schedules mostly conventional graphics work\",\"new\":\"Engine increasingly schedules rendering, data movement and ML workloads\"}}]},\"tunes\":{}},{\"id\":\"c-0016\",\"type\":\"callout\",\"data\":{\"variant\":\"success\",\"title\":\"The central Figure Rocks hypothesis\",\"body\":\"\u003Cstrong>AI may be weakening the economics of brute-force gaming hardware while simultaneously supplying techniques that make brute-force scaling less necessary.\u003C\u002Fstrong> That does not mean future games need less hardware. It means more of the result may come from deciding what to store, move, render, reconstruct or infer. Figure Rocks shorthand: \u003Cstrong>store less → move less → predict more → render less → infer more.\u003C\u002Fstrong> This is a hypothesis about where performance gains may increasingly come from, not a claim that every future game will use one pipeline.\"},\"tunes\":{}},{\"id\":\"d-0017\",\"type\":\"delimiter\",\"data\":{},\"tunes\":{}},{\"id\":\"h-0018\",\"type\":\"header\",\"data\":{\"text\":\"The 5090 question is useful only if we phrase it correctly\",\"level\":2},\"tunes\":{}},{\"id\":\"p-0019\",\"type\":\"paragraph\",\"data\":{\"text\":\"The RTX 5090 is a useful example precisely because the evidence is mixed. As of October 18, 2026, NVIDIA still lists the GeForce RTX 5090 as a current product. At the same time, retailer behavior and press reports have documented unusually poor availability and unconfirmed claims of a production halt. That is a market signal, not an official end-of-production announcement. The defensible question is therefore not whether NVIDIA secretly killed one card, but whether the opportunity cost of allocating top-tier Blackwell-class silicon to enthusiast gaming rises when similar architecture can serve professional visualization and AI markets with very different economics.\"},\"tunes\":{}},{\"id\":\"c-0020\",\"type\":\"callout\",\"data\":{\"variant\":\"note\",\"title\":\"FACT \u002F SIGNAL \u002F HYPOTHESIS\",\"body\":\"\u003Cstrong>FACT:\u003C\u002Fstrong> NVIDIA still publicly lists the RTX 5090, and Blackwell architecture spans gaming and professional products.\u003Cbr>\u003Cstrong>INDUSTRY SIGNAL:\u003C\u002Fstrong> retailer disappearance, extreme scarcity and reports of a possible production halt exist, but NVIDIA has not confirmed an RTX 5090 EOL or production stop.\u003Cbr>\u003Cstrong>TECHNICAL INFERENCE:\u003C\u002Fstrong> a high-end die has multiple economically valuable destinations.\u003Cbr>\u003Cstrong>FIGURE ROCKS HYPOTHESIS:\u003C\u002Fstrong> AI and professional demand can make the traditional enthusiast expectation — that the largest practical die remains broadly attainable as a gaming flagship — harder to sustain.\"},\"tunes\":{}},{\"id\":\"v3-memory-compute\",\"type\":\"header\",\"data\":{\"text\":\"Memory is starting to become a compute problem, not only a capacity problem\",\"level\":2},\"tunes\":{}},{\"id\":\"v3-memory-compute-p1\",\"type\":\"paragraph\",\"data\":{\"text\":\"There are at least four ways to attack a memory bottleneck without simply increasing raw DRAM speed: compress the representation, reconstruct detail with compute, predict what data will be needed next, and keep the right data resident close to the execution unit. None of these makes a physical memory chip magically faster. They change how much useful work the system obtains from a finite capacity and bandwidth budget.\"},\"tunes\":{}},{\"id\":\"v3-four-memory-levers\",\"type\":\"table\",\"data\":{\"withHeadings\":true,\"stretched\":false,\"content\":[[\"Technique\",\"What changes\",\"What does not change automatically\"],[\"Compression\",\"Fewer physical bytes represent the same logical data\",\"The DRAM clock and bus width\"],[\"Neural reconstruction\",\"Some detail is generated from a compact representation\",\"The cost disappears; inference has its own compute\u002Flatency cost\"],[\"Predictive prefetch\",\"Data arrives before the requesting workload stalls\",\"Total physical bandwidth\"],[\"Residency \u002F unified access\",\"Less unnecessary migration or duplication\",\"Contention and latency constraints\"]]},\"tunes\":{}},{\"id\":\"v3-buddy-data\",\"type\":\"callout\",\"data\":{\"variant\":\"info\",\"title\":\"Research anchor: effective capacity is not raw speed\",\"body\":\"NVIDIA Research's Buddy Compression reported an average \u003Cstrong>2.2× compression ratio for HPC workloads and 1.5× for deep-learning workloads\u003C\u002Fstrong>, with 1–2% slowdown in that study. That result is not a gaming benchmark. Its relevance here is architectural: compressed memory can increase effective capacity and reduce some traffic without changing the physical DRAM clock.\"},\"tunes\":{}},{\"id\":\"v5-ntc-evidence\",\"type\":\"callout\",\"data\":{\"variant\":\"success\",\"title\":\"Gaming-specific evidence: neural texture compression\",\"body\":\"NVIDIA Research's material-texture NTC work demonstrated \u003Cstrong>4× higher linear texture resolution — 16× more texels — while using 30% less memory\u003C\u002Fstrong> than its BC High comparison. That does not mean every game gets a 16× memory gain. It demonstrates the mechanism that matters here: some visual detail can be represented more compactly and reconstructed with compute instead of being stored conventionally at the same fidelity.\"},\"tunes\":{}},{\"id\":\"v3-460-correction\",\"type\":\"callout\",\"data\":{\"variant\":\"warning\",\"title\":\"What the “460×” number does not mean\",\"body\":\"A 460× figure sometimes associated with NVIDIA systems refers to a comparison in GPU-accessible memory capacity across DGX generations, \u003Cstrong>not a 460× memory-compression ratio\u003C\u002Fstrong>. This article deliberately does not use it as evidence for compression.\"},\"tunes\":{}},{\"id\":\"h-0021\",\"type\":\"header\",\"data\":{\"text\":\"How the technical response could work\",\"level\":2},\"tunes\":{}},{\"id\":\"flow-0022\",\"type\":\"processFlow\",\"data\":{\"title\":\"One possible future frame pipeline\",\"orientation\":\"auto\",\"steps\":[{\"label\":\"Compress\",\"description\":\"Keep textures or learned representations smaller where quality permits.\"},{\"label\":\"Predict\",\"description\":\"Use deterministic game state and learned models to estimate what data will be needed next.\"},{\"label\":\"Prefetch\",\"description\":\"Move useful data before the GPU stalls waiting for it.\"},{\"label\":\"Render\",\"description\":\"Spend conventional graphics work where it changes the visible result most.\"},{\"label\":\"Infer\",\"description\":\"Use neural reconstruction or generation where inference is cheaper than storing or simulating everything directly.\"},{\"label\":\"Present\",\"description\":\"Deliver the frame while keeping latency, artifacts and frame pacing inside acceptable limits.\"}]},\"tunes\":{}},{\"id\":\"p-0023\",\"type\":\"paragraph\",\"data\":{\"text\":\"This pipeline is not one announced product. Every component already exists in some form, but the integrated chain is the hypothesis. Neural texture compression has shown that learned representations can exchange compute for memory. GPU prefetch research has shown that learned prediction can improve memory-page hit rates. DirectX is formalizing ML inside the graphics programming model. DLSS 5 shows a generative stage entering real-time rendering.\"},\"tunes\":{}},{\"id\":\"v3-predictive-loop\",\"type\":\"header\",\"data\":{\"text\":\"A game has one advantage general workloads often do not: it already knows part of its future\",\"level\":2},\"tunes\":{}},{\"id\":\"v3-predictive-loop-p1\",\"type\":\"paragraph\",\"data\":{\"text\":\"A game engine knows the camera transform, player velocity, current world cell, neighboring streaming cells, animation state, active entities, LOD transitions, shader dependencies and often likely traversal paths. That deterministic context does not reveal the future perfectly, but it gives a predictor a richer starting point than a generic memory trace alone.\"},\"tunes\":{}},{\"id\":\"v3-predictive-door\",\"type\":\"processFlow\",\"data\":{\"title\":\"A predictive gaming-memory example\",\"orientation\":\"auto\",\"steps\":[{\"label\":\"1. Observe deterministic state\",\"description\":\"Camera direction, movement vector, current room, visible portals and known asset dependencies already exist in engine state.\"},{\"label\":\"2. Estimate the near-future working set\",\"description\":\"A learned predictor can rank likely next rooms, textures, geometry, shaders or neural representations.\"},{\"label\":\"3. Prefetch during available bandwidth\",\"description\":\"Move likely resources while the memory system has headroom instead of waiting for a hard demand miss.\"},{\"label\":\"4. Keep deterministic guarantees\",\"description\":\"Critical assets still use authored rules; prediction supplements rather than replaces correctness.\"},{\"label\":\"5. Measure the miss cost\",\"description\":\"If prediction wastes bandwidth or evicts useful data, confidence or prefetch depth must fall.\"}]},\"tunes\":{}},{\"id\":\"v3-learned-evidence\",\"type\":\"callout\",\"data\":{\"variant\":\"note\",\"title\":\"This idea has a research precedent, but not yet a gaming proof\",\"body\":\"A published CPU–GPU unified-virtual-memory study reported a \u003Cstrong>10.89% performance improvement\u003C\u002Fstrong>, page-hit rate rising from \u003Cstrong>76.10% to 89.02%\u003C\u002Fstrong>, and \u003Cstrong>11.05% lower CPU–GPU interconnect traffic\u003C\u002Fstrong> using learned prefetching. A 2026 AAAI paper on asynchronous KV-cache prefetching reported \u003Cstrong>2.15× attention-kernel efficiency and up to 1.97× end-to-end throughput\u003C\u002Fstrong> on its LLM workload. Neither result should be copied onto games; both demonstrate that prediction and prefetch can hide memory cost in real GPU workloads.\"},\"tunes\":{}},{\"id\":\"h-0024\",\"type\":\"header\",\"data\":{\"text\":\"Why a Spark-like architecture becomes interesting\",\"level\":2},\"tunes\":{}},{\"id\":\"p-0025\",\"type\":\"paragraph\",\"data\":{\"text\":\"DGX Spark is not a gaming PC, and treating it as one would hide the useful lesson. Its GB10 design combines Arm CPU cores, a Blackwell GPU, Tensor and RT hardware, and up to 128 GB of coherent unified system memory. That architecture asks a different question from a classic desktop: instead of splitting the system cleanly into CPU memory and discrete VRAM, what happens when multiple compute engines share a large memory pool?\"},\"tunes\":{}},{\"id\":\"t-0026\",\"type\":\"table\",\"data\":{\"withHeadings\":true,\"stretched\":false,\"content\":[[\"Property\",\"DGX Spark reference\",\"Gaming implication if the principle were adapted\"],[\"Memory\",\"Up to 128 GB coherent unified LPDDR5X\",\"Large shared pool could reduce duplication and give local AI + graphics access to common data\"],[\"Memory bandwidth\",\"273 GB\u002Fs\",\"Capacity is not enough; high-end graphics still needs aggressive bandwidth and latency management\"],[\"Compute mix\",\"CPU + Blackwell GPU + Tensor + RT\",\"A future game workload could be split across more specialized engines\"],[\"Form factor\u002Fpower\",\"Compact GB10 system\",\"Integration can improve efficiency but may reduce modular upgradeability\"]]},\"tunes\":{}},{\"id\":\"c-0027\",\"type\":\"callout\",\"data\":{\"variant\":\"warning\",\"title\":\"Capacity is not bandwidth\",\"body\":\"A large unified pool does \u003Cstrong>not\u003C\u002Fstrong> automatically replace the bandwidth of a high-end discrete GPU. DGX Spark lists 273 GB\u002Fs of memory bandwidth, while NVIDIA's RTX PRO 6000 Blackwell Workstation Edition lists 1,792 GB\u002Fs. These are very different products and the comparison is not a gaming benchmark; it simply makes the architectural constraint visible. A Spark-like gaming design would need to solve contention, latency and bandwidth — not just provide more shared gigabytes.\"},\"tunes\":{}},{\"id\":\"h-0028\",\"type\":\"header\",\"data\":{\"text\":\"What would have to be true for this transition to work?\",\"level\":2},\"tunes\":{}},{\"id\":\"flow-0029\",\"type\":\"processFlow\",\"data\":{\"title\":\"The Neural Gaming Viability Test\",\"orientation\":\"auto\",\"steps\":[{\"label\":\"1. Inference must save more than it costs\",\"description\":\"Neural work has to remove enough storage, bandwidth or conventional rendering cost to justify its own compute and latency.\"},{\"label\":\"2. The engine must expose useful state\",\"description\":\"Prediction works better when the system knows camera direction, asset dependencies, movement and likely visibility.\"},{\"label\":\"3. Memory movement must become deliberate\",\"description\":\"Compression without good residency and prefetch policy can still stall.\"},{\"label\":\"4. Quality must remain stable\",\"description\":\"Generated detail cannot introduce unacceptable artifacts, temporal instability or incorrect game information.\"},{\"label\":\"5. Latency must stay bounded\",\"description\":\"A prettier frame is not automatically a more responsive game.\"},{\"label\":\"6. The stack must survive vendor differences\",\"description\":\"Useful techniques need APIs, drivers and engine paths that do not fragment into one-off demos.\"}]},\"tunes\":{}},{\"id\":\"v3-reader-spark-expect\",\"type\":\"header\",\"data\":{\"text\":\"What should a buyer expect from a Spark-like transition?\",\"level\":2},\"tunes\":{}},{\"id\":\"v3-buyer-transition\",\"type\":\"comparison\",\"data\":{\"title\":\"Potential benefit vs real cost\",\"layout\":\"table\",\"columns\":[{\"id\":\"gain\",\"label\":\"Potential gain\"},{\"id\":\"cost\",\"label\":\"Cost \u002F risk\"}],\"rows\":[{\"id\":\"pool\",\"label\":\"Large coherent pool\",\"values\":{\"gain\":\"More flexible sharing between CPU, graphics and local AI\",\"cost\":\"Shared clients can contend for the same bandwidth\"}},{\"id\":\"movement\",\"label\":\"Less duplication\",\"values\":{\"gain\":\"Fewer copies between separate system-RAM and VRAM pools\",\"cost\":\"Coherence and scheduling add complexity\"}},{\"id\":\"ai\",\"label\":\"More local AI headroom\",\"values\":{\"gain\":\"Larger models, agents or neural assets can coexist locally\",\"cost\":\"AI can compete with graphics for compute and memory\"}},{\"id\":\"efficiency\",\"label\":\"System-level optimization\",\"values\":{\"gain\":\"The platform can optimize data placement and accelerators together\",\"cost\":\"Tighter integration can reduce upgradeability and increase lock-in\"}},{\"id\":\"longevity\",\"label\":\"New workload flexibility\",\"values\":{\"gain\":\"A machine may age better if future games shift toward neural workloads\",\"cost\":\"Only if engines and runtimes actually use those capabilities\"}}]},\"tunes\":{}},{\"id\":\"h-0030\",\"type\":\"header\",\"data\":{\"text\":\"What would falsify this hypothesis?\",\"level\":2},\"tunes\":{}},{\"id\":\"p-0031\",\"type\":\"paragraph\",\"data\":{\"text\":\"This thesis is deliberately falsifiable. The table below defines the market, software, hardware and game-adoption signals that would make the argument stronger or weaker over time.\"},\"tunes\":{}},{\"id\":\"cmp-0032\",\"type\":\"comparison\",\"data\":{\"title\":\"Evidence that strengthens or weakens the thesis\",\"layout\":\"table\",\"columns\":[{\"id\":\"strengthens\",\"label\":\"Would strengthen it\"},{\"id\":\"weakens\",\"label\":\"Would weaken it\"}],\"rows\":[{\"id\":\"memory\",\"label\":\"Memory economics\",\"values\":{\"strengthens\":\"Server\u002FHBM priority continues to squeeze consumer allocation\",\"weakens\":\"Consumer capacity expands and price pressure normalizes\"}},{\"id\":\"graphics\",\"label\":\"Graphics stack\",\"values\":{\"strengthens\":\"ML primitives become normal engine\u002FDirectX features\",\"weakens\":\"Neural graphics remains vendor-specific optional post-processing\"}},{\"id\":\"hardware\",\"label\":\"Hardware design\",\"values\":{\"strengthens\":\"More integrated heterogeneous compute and larger coherent memory pools\",\"weakens\":\"Discrete GPU + separate RAM remains clearly superior on cost, flexibility and latency\"}},{\"id\":\"games\",\"label\":\"Game adoption\",\"values\":{\"strengthens\":\"Studios ship neural compression, generation and predictive data paths broadly\",\"weakens\":\"Tech remains limited to demos or a few flagship games\"}}]},\"tunes\":{}},{\"id\":\"h-0033\",\"type\":\"header\",\"data\":{\"text\":\"The practical conclusion for a gamer today\",\"level\":2},\"tunes\":{}},{\"id\":\"p-0034\",\"type\":\"paragraph\",\"data\":{\"text\":\"Do not buy hardware because a speculative architecture might arrive. Buy for the games, resolution, latency target and workloads you actually have. But do not assume that future-proofing can still be reduced to one number such as raster performance or VRAM capacity. The direction of the software stack suggests that memory behavior, neural compute and engine support may become increasingly important parts of the value equation.\"},\"tunes\":{}},{\"id\":\"ref-0035\",\"type\":\"referralArticle\",\"data\":{\"url\":\"https:\u002F\u002Ffigure.rocks\u002Fblog\u002Fvram-usage-is-not-vram-requirement-why-a-full-memory-meter-does-not-tell-the-whole-story\",\"title\":\"VRAM Usage Is Not VRAM Requirement: Why a Full Memory Meter Does Not Tell the Whole Story\",\"excerpt\":\"A practical explanation of working sets, residency and why an allocated VRAM number is not the same thing as a hard requirement.\",\"ctaLabel\":\"Read the VRAM guide\"},\"tunes\":{}},{\"id\":\"h-0036\",\"type\":\"header\",\"data\":{\"text\":\"Primary sources\",\"level\":2},\"tunes\":{}},{\"id\":\"v3-src-newzoo\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fnewzoo.com\u002Farticles\u002F2026-global-games-market-key-numbers\",\"meta\":{\"title\":\"Newzoo — 2026 Global Games Market Key Numbers\",\"description\":\"2026 games-market forecast used for the scale comparison.\",\"image\":{\"url\":\"\"}}},\"tunes\":{}},{\"id\":\"src-0038\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fwww.semiconductors.org\u002Fglobal-annual-semiconductor-sales-increase-25-6-to-791-7-billion-in-2025\u002F\",\"meta\":{\"title\":\"Semiconductor Industry Association — 2025 Global Sales\",\"description\":\"2025 semiconductor sales and the association's roughly $1 trillion 2026 projection.\",\"image\":{\"url\":\"\"}}},\"tunes\":{}},{\"id\":\"v3-src-trend-sep\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fwww.trendforce.com\u002Fpresscenter\u002Fnews\u002F20260930-13258.html\",\"meta\":{\"title\":\"TrendForce — AI Server Demand Sustains Memory Price Increases in 4Q26\",\"description\":\"Late-2026 evidence for ongoing server-priority capacity allocation and consumer pressure.\",\"image\":{\"url\":\"\"}}},\"tunes\":{}},{\"id\":\"v3-src-buddy\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fresearch.nvidia.com\u002Fpublication\u002F2019-04_buddy-compression-enabling-larger-memory-deep-learning-and-hpc-workloads-gpus\",\"meta\":{\"title\":\"NVIDIA Research — Buddy Compression\",\"description\":\"Research source for 2.2× HPC and 1.5× DL average compression ratios.\",\"image\":{\"url\":\"\"}}},\"tunes\":{}},{\"id\":\"v3-src-uvm\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fwww.sciencedirect.com\u002Fscience\u002Farticle\u002Fabs\u002Fpii\u002FS0743731522002490\",\"meta\":{\"title\":\"Deep-learning based CPU–GPU UVM prefetching\",\"description\":\"Research source for learned page-prefetching results.\",\"image\":{\"url\":\"\"}}},\"tunes\":{}},{\"id\":\"v3-src-aaai\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fojs.aaai.org\u002Findex.php\u002FAAAI\u002Farticle\u002Fview\u002F39224\",\"meta\":{\"title\":\"AAAI 2026 — Asynchronous KV Cache Prefetching\",\"description\":\"Research source demonstrating latency hiding through proactive GPU-cache prefetching.\",\"image\":{\"url\":\"\"}}},\"tunes\":{}},{\"id\":\"src-0039\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fwww.trendforce.com\u002Fpresscenter\u002Fnews\u002F20260331-12995.html\",\"meta\":{\"title\":\"TrendForce — AI Server Demand and 2Q26 Memory Prices\",\"description\":\"Current evidence for server\u002FHBM prioritization, GDDR constraints and consumer pressure.\",\"image\":{\"url\":\"\"}}},\"tunes\":{}},{\"id\":\"src-0040\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fdevblogs.microsoft.com\u002Fdirectx\u002Fevolving-directx-for-the-ml-era-on-windows\u002F\",\"meta\":{\"title\":\"Microsoft DirectX — Evolving DirectX for the ML Era\",\"description\":\"Official GDC 2026 overview of shader-level and model-level ML in DirectX.\",\"image\":{\"url\":\"\"}}},\"tunes\":{}},{\"id\":\"src-0041\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fresearch.nvidia.com\u002Flabs\u002Fadlr\u002FDLSS5\u002F\",\"meta\":{\"title\":\"NVIDIA Research — DLSS 5: Generative Neural Rendering\",\"description\":\"Primary research page describing DLSS 5 as a real-time generative rendering stage.\",\"image\":{\"url\":\"\"}}},\"tunes\":{}},{\"id\":\"src-0042\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fwww.nvidia.com\u002Fen-us\u002Fproducts\u002Fworkstations\u002Fdgx-spark\u002F\",\"meta\":{\"title\":\"NVIDIA — DGX Spark Specifications\",\"description\":\"Official architecture, unified memory and bandwidth specifications.\",\"image\":{\"url\":\"\"}}},\"tunes\":{}},{\"id\":\"v5-src-5090-official\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fwww.nvidia.com\u002Fen-us\u002Fgeforce\u002Fgraphics-cards\u002F50-series\u002Frtx-5090\u002F\",\"meta\":{\"title\":\"NVIDIA — GeForce RTX 5090\",\"description\":\"Official current RTX 5090 product page used to distinguish confirmed product status from production-halt rumors.\",\"image\":{\"url\":\"\"}}},\"tunes\":{}},{\"id\":\"v5-src-5090-signal\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fwww.tomshardware.com\u002Fpc-components\u002Fgpus\u002Fnvidia-rtx-5090-disappears-from-micro-center-amid-rumors-of-a-production-halt-retailer-flags-flagship-as-not-carried-as-stock-mysteriously-vanishes\",\"meta\":{\"title\":\"Tom’s Hardware — RTX 5090 Retail Disappearance and Production-Halt Rumors\",\"description\":\"Current market report used only as an industry signal; it explicitly notes that NVIDIA has not confirmed a production halt.\",\"image\":{\"url\":\"\"}}},\"tunes\":{}},{\"id\":\"v5-src-ntc\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fresearch.nvidia.com\u002Flabs\u002Frtr\u002Fneural_texture_compression\u002F\",\"meta\":{\"title\":\"NVIDIA Research — Random-Access Neural Compression of Material Textures\",\"description\":\"Primary source for the 4× linear resolution \u002F 16× texel comparison at 30% less memory versus BC High.\",\"image\":{\"url\":\"\"}}},\"tunes\":{}},{\"id\":\"v5-src-rtxpro6000\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fwww.nvidia.com\u002Fde-de\u002Fproducts\u002Fworkstations\u002Fprofessional-desktop-gpus\u002Frtx-pro-6000\u002F\",\"meta\":{\"title\":\"NVIDIA — RTX PRO 6000 Blackwell Workstation Edition\",\"description\":\"Official source for 96 GB GDDR7 and 1,792 GB\u002Fs memory bandwidth used as a bandwidth reference point.\",\"image\":{\"url\":\"\"}}},\"tunes\":{}},{\"id\":\"gloss-0044\",\"type\":\"glossary\",\"data\":{\"title\":\"Key concepts\",\"entries\":[{\"term\":\"Neural rendering\",\"definition\":\"Rendering methods that use trained models as part of producing, reconstructing or enhancing visual output.\",\"anchor\":\"neural-rendering\"},{\"term\":\"Heterogeneous compute\",\"definition\":\"A system that assigns different workloads to different compute engines such as CPU, GPU, RT and matrix\u002FTensor hardware.\",\"anchor\":\"heterogeneous-compute\"},{\"term\":\"Effective memory\",\"definition\":\"A practical description of how much useful data a memory system can keep and feed to compute after compression, residency and movement efficiency are considered. It is not a formal replacement for physical capacity.\",\"anchor\":\"effective-memory\"},{\"term\":\"Predictive memory\",\"definition\":\"Figure Rocks shorthand for combining known game state and learned prediction to prepare likely-needed data before a stall. It is a hypothesis, not an industry standard.\",\"anchor\":\"predictive-memory\"}]},\"tunes\":{}},{\"id\":\"p-0001\",\"type\":\"paragraph\",\"data\":{\"text\":\"For decades, enthusiast PC gaming followed a familiar bargain: games asked for more compute, GPU generations delivered more compute, and players upgraded when the gap became large enough. That loop never worked perfectly, but the direction was understandable. In 2026, the loop is becoming harder to read because two changes are happening at the same time: AI infrastructure is competing aggressively for silicon and memory, while AI techniques are moving directly into the way games create frames.\"},\"tunes\":{}},{\"id\":\"c-0002\",\"type\":\"callout\",\"data\":{\"variant\":\"info\",\"title\":\"Direct answer\",\"body\":\"\u003Cstrong>The old gaming upgrade cycle is under pressure, but it is not simply being replaced by more expensive GPUs.\u003C\u002Fstrong> AI is changing both sides of the equation: it increases competition for semiconductor and memory resources, and it also enables neural rendering techniques that can trade stored data or conventional rendering work for inference. The important possibility is that gaming may respond to tighter hardware economics by becoming more computationally selective, not merely by demanding larger GPUs.\"},\"tunes\":{}},{\"id\":\"c-0003\",\"type\":\"callout\",\"data\":{\"variant\":\"note\",\"title\":\"How to read this article\",\"body\":\"This is an \u003Cstrong>evidence-backed hypothesis\u003C\u002Fstrong>. Documented market data and shipping or announced technologies are labeled as facts or industry signals. The connections between them are technical inference. The future architecture described later is a Figure Rocks hypothesis, not a claim about an unannounced NVIDIA, AMD, Intel, Microsoft or console roadmap.\"},\"tunes\":{}},{\"id\":\"toc-0004\",\"type\":\"tableOfContents\",\"data\":{\"title\":\"Contents\",\"minLevel\":2,\"maxLevel\":3},\"tunes\":{}},{\"id\":\"h-0005\",\"type\":\"header\",\"data\":{\"text\":\"The scale mismatch is already visible\",\"level\":2},\"tunes\":{}},{\"id\":\"p-0006\",\"type\":\"paragraph\",\"data\":{\"text\":\"Newzoo forecasts the global games market at $213.9 billion in 2026, up 6.1% year over year. The Semiconductor Industry Association reported $791.7 billion in global semiconductor sales for 2025, up 25.6%, and projected roughly $1 trillion for 2026. The games business and semiconductor business still depend on each other, but their incentives are becoming less symmetric: the same advanced manufacturing ecosystem that enables premium gaming now serves AI and data-center buyers with much larger infrastructure budgets. These are not directly comparable product categories; they establish the scale of the upstream compute economy around gaming.\"},\"tunes\":{}},{\"id\":\"t-0007\",\"type\":\"table\",\"data\":{\"withHeadings\":true,\"stretched\":false,\"content\":[[\"Reference point\",\"Reported or forecast value\",\"Why it matters here\"],[\"Global games market, 2026 forecast\",\"$213.9B, +6.1% YoY\",\"Large consumer entertainment market, but downstream of advanced hardware supply\"],[\"PC games, 2026 forecast\",\"$45.9B, +5.3% YoY\",\"The gaming segment most directly exposed to PC component economics\"],[\"Global semiconductor sales, 2025\",\"$791.7B, +25.6% YoY\",\"Shows the scale of the upstream industry serving gaming, AI, mobile, automotive and cloud\"],[\"Global semiconductor industry, 2026 projection\",\"~$1T\",\"Raises the opportunity cost of dedicating premium silicon only to consumer gaming\"]]},\"tunes\":{}},{\"id\":\"c-0008\",\"type\":\"callout\",\"data\":{\"variant\":\"warning\",\"title\":\"What these market numbers do — and do not — prove\",\"body\":\"The figures establish \u003Cstrong>economic context, not direct causation\u003C\u002Fstrong>. They do not prove that a specific wafer, GPU or memory package was taken from a gamer and reassigned to AI. They also do not mean gaming and semiconductors are the two largest industries in the world. The narrower, defensible point is that a $200B+ games market depends on an upstream semiconductor economy approaching trillion-dollar annual scale, where AI infrastructure now competes for advanced compute and memory under very different budget and margin conditions.\"},\"tunes\":{}},{\"id\":\"h-0009\",\"type\":\"header\",\"data\":{\"text\":\"Memory is where the pressure becomes concrete\",\"level\":2},\"tunes\":{}},{\"id\":\"p-0010\",\"type\":\"paragraph\",\"data\":{\"text\":\"TrendForce has repeatedly reported that DRAM suppliers are shifting capacity toward server and HBM products as AI demand grows. In March 2026 it forecast conventional DRAM contract prices rising 58–63% quarter over quarter in Q2 and said limited GDDR allocation was constraining graphics memory supply. By Q3, TrendForce said consumer customers were reaching an affordability limit even as AI and server demand remained strong.\"},\"tunes\":{}},{\"id\":\"flow-0011\",\"type\":\"processFlow\",\"data\":{\"title\":\"The resource-allocation pressure\",\"orientation\":\"auto\",\"steps\":[{\"label\":\"1. AI infrastructure expands\",\"description\":\"Cloud providers and AI builders secure server memory, HBM, enterprise SSDs and advanced compute.\"},{\"label\":\"2. Suppliers optimize product mix\",\"description\":\"Capacity moves toward products with stronger demand, margins or long-term contracts.\"},{\"label\":\"3. Consumer allocation tightens\",\"description\":\"PC, graphics and consumer memory buyers face higher prices or constrained availability.\"},{\"label\":\"4. Gaming vendors absorb or pass through cost\",\"description\":\"Board design, VRAM capacity and system price become harder trade-offs.\"},{\"label\":\"5. The old upgrade cycle weakens\",\"description\":\"A faster GPU is still useful, but brute-force scaling becomes economically harder to repeat at the same consumer price points.\"}]},\"tunes\":{}},{\"id\":\"v3-sep-signal\",\"type\":\"callout\",\"data\":{\"variant\":\"warning\",\"title\":\"A September 2026 signal, not an old shortage story\",\"body\":\"TrendForce reported on September 30, 2026 that suppliers still prioritize advanced-process capacity for high-performance server products, with conventional DRAM contract prices projected to rise another 10–15% QoQ in Q4. Its DRAM bulletin also says HBM continues to crowd out general DRAM capacity. The pressure described here therefore remains visible late in 2026; it is not based only on a temporary early-year spike.\"},\"tunes\":{}},{\"id\":\"h-0012\",\"type\":\"header\",\"data\":{\"text\":\"At exactly the same time, the frame itself is changing\",\"level\":2},\"tunes\":{}},{\"id\":\"p-0013\",\"type\":\"paragraph\",\"data\":{\"text\":\"Microsoft now describes machine learning as becoming foundational to real-time graphics. DirectX is adding shader-level and model-level ML paths so neural texture compression, denoising, super resolution and other learned workloads can sit inside the graphics stack instead of being bolted on as isolated features. NVIDIA's DLSS 5 research pushes the idea further by introducing a generative neural rendering stage that complements conventional rendering with learned appearance priors. Neural rendering does not mean conventional rendering disappears. The emerging model is hybrid: render some information directly, infer or reconstruct other information when doing so is cheaper or more useful.\"},\"tunes\":{}},{\"id\":\"ref-0014\",\"type\":\"referralArticle\",\"data\":{\"url\":\"https:\u002F\u002Ffigure.rocks\u002Fblog\u002Fdlss-5-is-not-just-upscaling-what-3d-guided-neural-rendering-actually-changes\",\"title\":\"DLSS 5 Is Not Just Upscaling: What Generative Neural Rendering Actually Changes\",\"excerpt\":\"Why the important change is not another quality mode, but a new stage in how visual information can be produced.\",\"ctaLabel\":\"Read the DLSS 5 analysis\"},\"tunes\":{}},{\"id\":\"cmp-0015\",\"type\":\"comparison\",\"data\":{\"title\":\"Traditional scaling vs neural-era scaling\",\"layout\":\"table\",\"columns\":[{\"id\":\"old\",\"label\":\"Traditional path\"},{\"id\":\"new\",\"label\":\"Neural-era path\"}],\"rows\":[{\"id\":\"fidelity\",\"label\":\"More visual fidelity\",\"values\":{\"old\":\"Store more detail and render more work\",\"new\":\"Store\u002Frender selectively, reconstruct or infer some detail\"}},{\"id\":\"memory\",\"label\":\"Memory response\",\"values\":{\"old\":\"Add capacity and bandwidth\",\"new\":\"Add capacity where needed, but also compress and reduce movement\"}},{\"id\":\"performance\",\"label\":\"Performance response\",\"values\":{\"old\":\"More shader\u002FRT throughput\",\"new\":\"Mix shader, RT and matrix\u002FTensor-style compute\"}},{\"id\":\"software\",\"label\":\"Software role\",\"values\":{\"old\":\"Engine schedules mostly conventional graphics work\",\"new\":\"Engine increasingly schedules rendering, data movement and ML workloads\"}}]},\"tunes\":{}},{\"id\":\"c-0016\",\"type\":\"callout\",\"data\":{\"variant\":\"success\",\"title\":\"The central Figure Rocks hypothesis\",\"body\":\"\u003Cstrong>AI may be weakening the economics of brute-force gaming hardware while simultaneously supplying techniques that make brute-force scaling less necessary.\u003C\u002Fstrong> That does not mean future games need less hardware. It means more of the result may come from deciding what to store, move, render, reconstruct or infer. Figure Rocks shorthand: \u003Cstrong>store less → move less → predict more → render less → infer more.\u003C\u002Fstrong> This is a hypothesis about where performance gains may increasingly come from, not a claim that every future game will use one pipeline.\"},\"tunes\":{}},{\"id\":\"d-0017\",\"type\":\"delimiter\",\"data\":{},\"tunes\":{}},{\"id\":\"h-0018\",\"type\":\"header\",\"data\":{\"text\":\"The 5090 question is useful only if we phrase it correctly\",\"level\":2},\"tunes\":{}},{\"id\":\"p-0019\",\"type\":\"paragraph\",\"data\":{\"text\":\"The RTX 5090 is a useful example precisely because the evidence is mixed. As of October 18, 2026, NVIDIA still lists the GeForce RTX 5090 as a current product. At the same time, retailer behavior and press reports have documented unusually poor availability and unconfirmed claims of a production halt. That is a market signal, not an official end-of-production announcement. The defensible question is therefore not whether NVIDIA secretly killed one card, but whether the opportunity cost of allocating top-tier Blackwell-class silicon to enthusiast gaming rises when similar architecture can serve professional visualization and AI markets with very different economics.\"},\"tunes\":{}},{\"id\":\"c-0020\",\"type\":\"callout\",\"data\":{\"variant\":\"note\",\"title\":\"FACT \u002F SIGNAL \u002F HYPOTHESIS\",\"body\":\"\u003Cstrong>FACT:\u003C\u002Fstrong> NVIDIA still publicly lists the RTX 5090, and Blackwell architecture spans gaming and professional products.\u003Cbr>\u003Cstrong>INDUSTRY SIGNAL:\u003C\u002Fstrong> retailer disappearance, extreme scarcity and reports of a possible production halt exist, but NVIDIA has not confirmed an RTX 5090 EOL or production stop.\u003Cbr>\u003Cstrong>TECHNICAL INFERENCE:\u003C\u002Fstrong> a high-end die has multiple economically valuable destinations.\u003Cbr>\u003Cstrong>FIGURE ROCKS HYPOTHESIS:\u003C\u002Fstrong> AI and professional demand can make the traditional enthusiast expectation — that the largest practical die remains broadly attainable as a gaming flagship — harder to sustain.\"},\"tunes\":{}},{\"id\":\"v3-memory-compute\",\"type\":\"header\",\"data\":{\"text\":\"Memory is starting to become a compute problem, not only a capacity problem\",\"level\":2},\"tunes\":{}},{\"id\":\"v3-memory-compute-p1\",\"type\":\"paragraph\",\"data\":{\"text\":\"There are at least four ways to attack a memory bottleneck without simply increasing raw DRAM speed: compress the representation, reconstruct detail with compute, predict what data will be needed next, and keep the right data resident close to the execution unit. None of these makes a physical memory chip magically faster. They change how much useful work the system obtains from a finite capacity and bandwidth budget.\"},\"tunes\":{}},{\"id\":\"v3-four-memory-levers\",\"type\":\"table\",\"data\":{\"withHeadings\":true,\"stretched\":false,\"content\":[[\"Technique\",\"What changes\",\"What does not change automatically\"],[\"Compression\",\"Fewer physical bytes represent the same logical data\",\"The DRAM clock and bus width\"],[\"Neural reconstruction\",\"Some detail is generated from a compact representation\",\"The cost disappears; inference has its own compute\u002Flatency cost\"],[\"Predictive prefetch\",\"Data arrives before the requesting workload stalls\",\"Total physical bandwidth\"],[\"Residency \u002F unified access\",\"Less unnecessary migration or duplication\",\"Contention and latency constraints\"]]},\"tunes\":{}},{\"id\":\"v3-buddy-data\",\"type\":\"callout\",\"data\":{\"variant\":\"info\",\"title\":\"Research anchor: effective capacity is not raw speed\",\"body\":\"NVIDIA Research's Buddy Compression reported an average \u003Cstrong>2.2× compression ratio for HPC workloads and 1.5× for deep-learning workloads\u003C\u002Fstrong>, with 1–2% slowdown in that study. That result is not a gaming benchmark. Its relevance here is architectural: compressed memory can increase effective capacity and reduce some traffic without changing the physical DRAM clock.\"},\"tunes\":{}},{\"id\":\"v5-ntc-evidence\",\"type\":\"callout\",\"data\":{\"variant\":\"success\",\"title\":\"Gaming-specific evidence: neural texture compression\",\"body\":\"NVIDIA Research's material-texture NTC work demonstrated \u003Cstrong>4× higher linear texture resolution — 16× more texels — while using 30% less memory\u003C\u002Fstrong> than its BC High comparison. That does not mean every game gets a 16× memory gain. It demonstrates the mechanism that matters here: some visual detail can be represented more compactly and reconstructed with compute instead of being stored conventionally at the same fidelity.\"},\"tunes\":{}},{\"id\":\"v3-460-correction\",\"type\":\"callout\",\"data\":{\"variant\":\"warning\",\"title\":\"What the “460×” number does not mean\",\"body\":\"A 460× figure sometimes associated with NVIDIA systems refers to a comparison in GPU-accessible memory capacity across DGX generations, \u003Cstrong>not a 460× memory-compression ratio\u003C\u002Fstrong>. This article deliberately does not use it as evidence for compression.\"},\"tunes\":{}},{\"id\":\"h-0021\",\"type\":\"header\",\"data\":{\"text\":\"How the technical response could work\",\"level\":2},\"tunes\":{}},{\"id\":\"flow-0022\",\"type\":\"processFlow\",\"data\":{\"title\":\"One possible future frame pipeline\",\"orientation\":\"auto\",\"steps\":[{\"label\":\"Compress\",\"description\":\"Keep textures or learned representations smaller where quality permits.\"},{\"label\":\"Predict\",\"description\":\"Use deterministic game state and learned models to estimate what data will be needed next.\"},{\"label\":\"Prefetch\",\"description\":\"Move useful data before the GPU stalls waiting for it.\"},{\"label\":\"Render\",\"description\":\"Spend conventional graphics work where it changes the visible result most.\"},{\"label\":\"Infer\",\"description\":\"Use neural reconstruction or generation where inference is cheaper than storing or simulating everything directly.\"},{\"label\":\"Present\",\"description\":\"Deliver the frame while keeping latency, artifacts and frame pacing inside acceptable limits.\"}]},\"tunes\":{}},{\"id\":\"p-0023\",\"type\":\"paragraph\",\"data\":{\"text\":\"This pipeline is not one announced product. Every component already exists in some form, but the integrated chain is the hypothesis. Neural texture compression has shown that learned representations can exchange compute for memory. GPU prefetch research has shown that learned prediction can improve memory-page hit rates. DirectX is formalizing ML inside the graphics programming model. DLSS 5 shows a generative stage entering real-time rendering.\"},\"tunes\":{}},{\"id\":\"v3-predictive-loop\",\"type\":\"header\",\"data\":{\"text\":\"A game has one advantage general workloads often do not: it already knows part of its future\",\"level\":2},\"tunes\":{}},{\"id\":\"v3-predictive-loop-p1\",\"type\":\"paragraph\",\"data\":{\"text\":\"A game engine knows the camera transform, player velocity, current world cell, neighboring streaming cells, animation state, active entities, LOD transitions, shader dependencies and often likely traversal paths. That deterministic context does not reveal the future perfectly, but it gives a predictor a richer starting point than a generic memory trace alone.\"},\"tunes\":{}},{\"id\":\"v3-predictive-door\",\"type\":\"processFlow\",\"data\":{\"title\":\"A predictive gaming-memory example\",\"orientation\":\"auto\",\"steps\":[{\"label\":\"1. Observe deterministic state\",\"description\":\"Camera direction, movement vector, current room, visible portals and known asset dependencies already exist in engine state.\"},{\"label\":\"2. Estimate the near-future working set\",\"description\":\"A learned predictor can rank likely next rooms, textures, geometry, shaders or neural representations.\"},{\"label\":\"3. Prefetch during available bandwidth\",\"description\":\"Move likely resources while the memory system has headroom instead of waiting for a hard demand miss.\"},{\"label\":\"4. Keep deterministic guarantees\",\"description\":\"Critical assets still use authored rules; prediction supplements rather than replaces correctness.\"},{\"label\":\"5. Measure the miss cost\",\"description\":\"If prediction wastes bandwidth or evicts useful data, confidence or prefetch depth must fall.\"}]},\"tunes\":{}},{\"id\":\"v3-learned-evidence\",\"type\":\"callout\",\"data\":{\"variant\":\"note\",\"title\":\"This idea has a research precedent, but not yet a gaming proof\",\"body\":\"A published CPU–GPU unified-virtual-memory study reported a \u003Cstrong>10.89% performance improvement\u003C\u002Fstrong>, page-hit rate rising from \u003Cstrong>76.10% to 89.02%\u003C\u002Fstrong>, and \u003Cstrong>11.05% lower CPU–GPU interconnect traffic\u003C\u002Fstrong> using learned prefetching. A 2026 AAAI paper on asynchronous KV-cache prefetching reported \u003Cstrong>2.15× attention-kernel efficiency and up to 1.97× end-to-end throughput\u003C\u002Fstrong> on its LLM workload. Neither result should be copied onto games; both demonstrate that prediction and prefetch can hide memory cost in real GPU workloads.\"},\"tunes\":{}},{\"id\":\"h-0024\",\"type\":\"header\",\"data\":{\"text\":\"Why a Spark-like architecture becomes interesting\",\"level\":2},\"tunes\":{}},{\"id\":\"p-0025\",\"type\":\"paragraph\",\"data\":{\"text\":\"DGX Spark is not a gaming PC, and treating it as one would hide the useful lesson. Its GB10 design combines Arm CPU cores, a Blackwell GPU, Tensor and RT hardware, and up to 128 GB of coherent unified system memory. That architecture asks a different question from a classic desktop: instead of splitting the system cleanly into CPU memory and discrete VRAM, what happens when multiple compute engines share a large memory pool?\"},\"tunes\":{}},{\"id\":\"t-0026\",\"type\":\"table\",\"data\":{\"withHeadings\":true,\"stretched\":false,\"content\":[[\"Property\",\"DGX Spark reference\",\"Gaming implication if the principle were adapted\"],[\"Memory\",\"Up to 128 GB coherent unified LPDDR5X\",\"Large shared pool could reduce duplication and give local AI + graphics access to common data\"],[\"Memory bandwidth\",\"273 GB\u002Fs\",\"Capacity is not enough; high-end graphics still needs aggressive bandwidth and latency management\"],[\"Compute mix\",\"CPU + Blackwell GPU + Tensor + RT\",\"A future game workload could be split across more specialized engines\"],[\"Form factor\u002Fpower\",\"Compact GB10 system\",\"Integration can improve efficiency but may reduce modular upgradeability\"]]},\"tunes\":{}},{\"id\":\"c-0027\",\"type\":\"callout\",\"data\":{\"variant\":\"warning\",\"title\":\"Capacity is not bandwidth\",\"body\":\"A large unified pool does \u003Cstrong>not\u003C\u002Fstrong> automatically replace the bandwidth of a high-end discrete GPU. DGX Spark lists 273 GB\u002Fs of memory bandwidth, while NVIDIA's RTX PRO 6000 Blackwell Workstation Edition lists 1,792 GB\u002Fs. These are very different products and the comparison is not a gaming benchmark; it simply makes the architectural constraint visible. A Spark-like gaming design would need to solve contention, latency and bandwidth — not just provide more shared gigabytes.\"},\"tunes\":{}},{\"id\":\"h-0028\",\"type\":\"header\",\"data\":{\"text\":\"What would have to be true for this transition to work?\",\"level\":2},\"tunes\":{}},{\"id\":\"flow-0029\",\"type\":\"processFlow\",\"data\":{\"title\":\"The Neural Gaming Viability Test\",\"orientation\":\"auto\",\"steps\":[{\"label\":\"1. Inference must save more than it costs\",\"description\":\"Neural work has to remove enough storage, bandwidth or conventional rendering cost to justify its own compute and latency.\"},{\"label\":\"2. The engine must expose useful state\",\"description\":\"Prediction works better when the system knows camera direction, asset dependencies, movement and likely visibility.\"},{\"label\":\"3. Memory movement must become deliberate\",\"description\":\"Compression without good residency and prefetch policy can still stall.\"},{\"label\":\"4. Quality must remain stable\",\"description\":\"Generated detail cannot introduce unacceptable artifacts, temporal instability or incorrect game information.\"},{\"label\":\"5. Latency must stay bounded\",\"description\":\"A prettier frame is not automatically a more responsive game.\"},{\"label\":\"6. The stack must survive vendor differences\",\"description\":\"Useful techniques need APIs, drivers and engine paths that do not fragment into one-off demos.\"}]},\"tunes\":{}},{\"id\":\"v3-reader-spark-expect\",\"type\":\"header\",\"data\":{\"text\":\"What should a buyer expect from a Spark-like transition?\",\"level\":2},\"tunes\":{}},{\"id\":\"v3-buyer-transition\",\"type\":\"comparison\",\"data\":{\"title\":\"Potential benefit vs real cost\",\"layout\":\"table\",\"columns\":[{\"id\":\"gain\",\"label\":\"Potential gain\"},{\"id\":\"cost\",\"label\":\"Cost \u002F risk\"}],\"rows\":[{\"id\":\"pool\",\"label\":\"Large coherent pool\",\"values\":{\"gain\":\"More flexible sharing between CPU, graphics and local AI\",\"cost\":\"Shared clients can contend for the same bandwidth\"}},{\"id\":\"movement\",\"label\":\"Less duplication\",\"values\":{\"gain\":\"Fewer copies between separate system-RAM and VRAM pools\",\"cost\":\"Coherence and scheduling add complexity\"}},{\"id\":\"ai\",\"label\":\"More local AI headroom\",\"values\":{\"gain\":\"Larger models, agents or neural assets can coexist locally\",\"cost\":\"AI can compete with graphics for compute and memory\"}},{\"id\":\"efficiency\",\"label\":\"System-level optimization\",\"values\":{\"gain\":\"The platform can optimize data placement and accelerators together\",\"cost\":\"Tighter integration can reduce upgradeability and increase lock-in\"}},{\"id\":\"longevity\",\"label\":\"New workload flexibility\",\"values\":{\"gain\":\"A machine may age better if future games shift toward neural workloads\",\"cost\":\"Only if engines and runtimes actually use those capabilities\"}}]},\"tunes\":{}},{\"id\":\"h-0030\",\"type\":\"header\",\"data\":{\"text\":\"What would falsify this hypothesis?\",\"level\":2},\"tunes\":{}},{\"id\":\"p-0031\",\"type\":\"paragraph\",\"data\":{\"text\":\"This thesis is deliberately falsifiable. The table below defines the market, software, hardware and game-adoption signals that would make the argument stronger or weaker over time.\"},\"tunes\":{}},{\"id\":\"cmp-0032\",\"type\":\"comparison\",\"data\":{\"title\":\"Evidence that strengthens or weakens the thesis\",\"layout\":\"table\",\"columns\":[{\"id\":\"strengthens\",\"label\":\"Would strengthen it\"},{\"id\":\"weakens\",\"label\":\"Would weaken it\"}],\"rows\":[{\"id\":\"memory\",\"label\":\"Memory economics\",\"values\":{\"strengthens\":\"Server\u002FHBM priority continues to squeeze consumer allocation\",\"weakens\":\"Consumer capacity expands and price pressure normalizes\"}},{\"id\":\"graphics\",\"label\":\"Graphics stack\",\"values\":{\"strengthens\":\"ML primitives become normal engine\u002FDirectX features\",\"weakens\":\"Neural graphics remains vendor-specific optional post-processing\"}},{\"id\":\"hardware\",\"label\":\"Hardware design\",\"values\":{\"strengthens\":\"More integrated heterogeneous compute and larger coherent memory pools\",\"weakens\":\"Discrete GPU + separate RAM remains clearly superior on cost, flexibility and latency\"}},{\"id\":\"games\",\"label\":\"Game adoption\",\"values\":{\"strengthens\":\"Studios ship neural compression, generation and predictive data paths broadly\",\"weakens\":\"Tech remains limited to demos or a few flagship games\"}}]},\"tunes\":{}},{\"id\":\"h-0033\",\"type\":\"header\",\"data\":{\"text\":\"The practical conclusion for a gamer today\",\"level\":2},\"tunes\":{}},{\"id\":\"p-0034\",\"type\":\"paragraph\",\"data\":{\"text\":\"Do not buy hardware because a speculative architecture might arrive. Buy for the games, resolution, latency target and workloads you actually have. But do not assume that future-proofing can still be reduced to one number such as raster performance or VRAM capacity. The direction of the software stack suggests that memory behavior, neural compute and engine support may become increasingly important parts of the value equation.\"},\"tunes\":{}},{\"id\":\"ref-0035\",\"type\":\"referralArticle\",\"data\":{\"url\":\"https:\u002F\u002Ffigure.rocks\u002Fblog\u002Fvram-usage-is-not-vram-requirement-why-a-full-memory-meter-does-not-tell-the-whole-story\",\"title\":\"VRAM Usage Is Not VRAM Requirement: Why a Full Memory Meter Does Not Tell the Whole Story\",\"excerpt\":\"A practical explanation of working sets, residency and why an allocated VRAM number is not the same thing as a hard requirement.\",\"ctaLabel\":\"Read the VRAM guide\"},\"tunes\":{}},{\"id\":\"h-0036\",\"type\":\"header\",\"data\":{\"text\":\"Primary sources\",\"level\":2},\"tunes\":{}},{\"id\":\"v3-src-newzoo\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fnewzoo.com\u002Farticles\u002F2026-global-games-market-key-numbers\",\"meta\":{\"title\":\"Newzoo — 2026 Global Games Market Key Numbers\",\"description\":\"2026 games-market forecast used for the scale comparison.\",\"image\":{\"url\":\"\"}}},\"tunes\":{}},{\"id\":\"src-0038\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fwww.semiconductors.org\u002Fglobal-annual-semiconductor-sales-increase-25-6-to-791-7-billion-in-2025\u002F\",\"meta\":{\"title\":\"Semiconductor Industry Association — 2025 Global Sales\",\"description\":\"2025 semiconductor sales and the association's roughly $1 trillion 2026 projection.\",\"image\":{\"url\":\"\"}}},\"tunes\":{}},{\"id\":\"v3-src-trend-sep\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fwww.trendforce.com\u002Fpresscenter\u002Fnews\u002F20260930-13258.html\",\"meta\":{\"title\":\"TrendForce — AI Server Demand Sustains Memory Price Increases in 4Q26\",\"description\":\"Late-2026 evidence for ongoing server-priority capacity allocation and consumer pressure.\",\"image\":{\"url\":\"\"}}},\"tunes\":{}},{\"id\":\"v3-src-buddy\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fresearch.nvidia.com\u002Fpublication\u002F2019-04_buddy-compression-enabling-larger-memory-deep-learning-and-hpc-workloads-gpus\",\"meta\":{\"title\":\"NVIDIA Research — Buddy Compression\",\"description\":\"Research source for 2.2× HPC and 1.5× DL average compression ratios.\",\"image\":{\"url\":\"\"}}},\"tunes\":{}},{\"id\":\"v3-src-uvm\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fwww.sciencedirect.com\u002Fscience\u002Farticle\u002Fabs\u002Fpii\u002FS0743731522002490\",\"meta\":{\"title\":\"Deep-learning based CPU–GPU UVM prefetching\",\"description\":\"Research source for learned page-prefetching results.\",\"image\":{\"url\":\"\"}}},\"tunes\":{}},{\"id\":\"v3-src-aaai\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fojs.aaai.org\u002Findex.php\u002FAAAI\u002Farticle\u002Fview\u002F39224\",\"meta\":{\"title\":\"AAAI 2026 — Asynchronous KV Cache Prefetching\",\"description\":\"Research source demonstrating latency hiding through proactive GPU-cache prefetching.\",\"image\":{\"url\":\"\"}}},\"tunes\":{}},{\"id\":\"src-0039\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fwww.trendforce.com\u002Fpresscenter\u002Fnews\u002F20260331-12995.html\",\"meta\":{\"title\":\"TrendForce — AI Server Demand and 2Q26 Memory Prices\",\"description\":\"Current evidence for server\u002FHBM prioritization, GDDR constraints and consumer pressure.\",\"image\":{\"url\":\"\"}}},\"tunes\":{}},{\"id\":\"src-0040\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fdevblogs.microsoft.com\u002Fdirectx\u002Fevolving-directx-for-the-ml-era-on-windows\u002F\",\"meta\":{\"title\":\"Microsoft DirectX — Evolving DirectX for the ML Era\",\"description\":\"Official GDC 2026 overview of shader-level and model-level ML in DirectX.\",\"image\":{\"url\":\"\"}}},\"tunes\":{}},{\"id\":\"src-0041\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fresearch.nvidia.com\u002Flabs\u002Fadlr\u002FDLSS5\u002F\",\"meta\":{\"title\":\"NVIDIA Research — DLSS 5: Generative Neural Rendering\",\"description\":\"Primary research page describing DLSS 5 as a real-time generative rendering stage.\",\"image\":{\"url\":\"\"}}},\"tunes\":{}},{\"id\":\"src-0042\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fwww.nvidia.com\u002Fen-us\u002Fproducts\u002Fworkstations\u002Fdgx-spark\u002F\",\"meta\":{\"title\":\"NVIDIA — DGX Spark Specifications\",\"description\":\"Official architecture, unified memory and bandwidth specifications.\",\"image\":{\"url\":\"\"}}},\"tunes\":{}},{\"id\":\"v5-src-5090-official\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fwww.nvidia.com\u002Fen-us\u002Fgeforce\u002Fgraphics-cards\u002F50-series\u002Frtx-5090\u002F\",\"meta\":{\"title\":\"NVIDIA — GeForce RTX 5090\",\"description\":\"Official current RTX 5090 product page used to distinguish confirmed product status from production-halt rumors.\",\"image\":{\"url\":\"\"}}},\"tunes\":{}},{\"id\":\"v5-src-5090-signal\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fwww.tomshardware.com\u002Fpc-components\u002Fgpus\u002Fnvidia-rtx-5090-disappears-from-micro-center-amid-rumors-of-a-production-halt-retailer-flags-flagship-as-not-carried-as-stock-mysteriously-vanishes\",\"meta\":{\"title\":\"Tom’s Hardware — RTX 5090 Retail Disappearance and Production-Halt Rumors\",\"description\":\"Current market report used only as an industry signal; it explicitly notes that NVIDIA has not confirmed a production halt.\",\"image\":{\"url\":\"\"}}},\"tunes\":{}},{\"id\":\"v5-src-ntc\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fresearch.nvidia.com\u002Flabs\u002Frtr\u002Fneural_texture_compression\u002F\",\"meta\":{\"title\":\"NVIDIA Research — Random-Access Neural Compression of Material Textures\",\"description\":\"Primary source for the 4× linear resolution \u002F 16× texel comparison at 30% less memory versus BC High.\",\"image\":{\"url\":\"\"}}},\"tunes\":{}},{\"id\":\"v5-src-rtxpro6000\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fwww.nvidia.com\u002Fde-de\u002Fproducts\u002Fworkstations\u002Fprofessional-desktop-gpus\u002Frtx-pro-6000\u002F\",\"meta\":{\"title\":\"NVIDIA — RTX PRO 6000 Blackwell Workstation Edition\",\"description\":\"Official source for 96 GB GDDR7 and 1,792 GB\u002Fs memory bandwidth used as a bandwidth reference point.\",\"image\":{\"url\":\"\"}}},\"tunes\":{}},{\"id\":\"gloss-0044\",\"type\":\"glossary\",\"data\":{\"title\":\"Key concepts\",\"entries\":[{\"term\":\"Neural rendering\",\"definition\":\"Rendering methods that use trained models as part of producing, reconstructing or enhancing visual output.\",\"anchor\":\"neural-rendering\"},{\"term\":\"Heterogeneous compute\",\"definition\":\"A system that assigns different workloads to different compute engines such as CPU, GPU, RT and matrix\u002FTensor hardware.\",\"anchor\":\"heterogeneous-compute\"},{\"term\":\"Effective memory\",\"definition\":\"A practical description of how much useful data a memory system can keep and feed to compute after compression, residency and movement efficiency are considered. It is not a formal replacement for physical capacity.\",\"anchor\":\"effective-memory\"},{\"term\":\"Predictive memory\",\"definition\":\"Figure Rocks shorthand for combining known game state and learned prediction to prepare likely-needed data before a stall. It is a hypothesis, not an industry standard.\",\"anchor\":\"predictive-memory\"}]},\"tunes\":{}}],\"version\":\"2.31.6\"}",{"time":1670,"blocks":1671,"version":2515},1792336677504,[1672,1676,1681,1686,1690,1694,1698,1722,1727,1731,1735,1755,1760,1764,1768,1775,1805,1810,1813,1817,1821,1826,1830,1834,1858,1863,1868,1873,1877,1900,1904,1908,1912,1932,1937,1941,1945,1967,1972,1976,1999,2003,2038,2042,2046,2076,2080,2084,2091,2095,2102,2109,2116,2122,2129,2136,2143,2150,2157,2164,2170,2177,2184,2191,2208,2211,2214,2217,2220,2223,2226,2235,2238,2241,2244,2253,2256,2259,2262,2265,2280,2283,2286,2289,2292,2295,2298,2301,2310,2313,2316,2319,2322,2332,2335,2338,2341,2350,2353,2356,2359,2368,2371,2374,2384,2387,2404,2407,2410,2425,2428,2431,2434,2437,2442,2447,2452,2457,2462,2467,2472,2477,2482,2487,2492,2497,2502,2507],{"id":541,"data":1673,"type":544,"tunes":1675},{"text":1674},"For decades, enthusiast PC gaming followed a familiar bargain: games asked for more compute, GPU generations delivered more compute, and players upgraded when the gap became large enough. That loop never worked perfectly, but the direction was understandable. In 2026, the loop is becoming harder to read because two changes are happening at the same time: AI infrastructure is competing aggressively for silicon and memory, while AI techniques are moving directly into the way games create frames.",{},{"id":547,"data":1677,"type":552,"tunes":1680},{"body":1678,"title":1679,"variant":551},"\u003Cstrong>The old gaming upgrade cycle is under pressure, but it is not simply being replaced by more expensive GPUs.\u003C\u002Fstrong> AI is changing both sides of the equation: it increases competition for semiconductor and memory resources, and it also enables neural rendering techniques that can trade stored data or conventional rendering work for inference. The important possibility is that gaming may respond to tighter hardware economics by becoming more computationally selective, not merely by demanding larger GPUs.","Direct answer",{},{"id":555,"data":1682,"type":552,"tunes":1685},{"body":1683,"title":1684,"variant":559},"This is an \u003Cstrong>evidence-backed hypothesis\u003C\u002Fstrong>. Documented market data and shipping or announced technologies are labeled as facts or industry signals. The connections between them are technical inference. The future architecture described later is a Figure Rocks hypothesis, not a claim about an unannounced NVIDIA, AMD, Intel, Microsoft or console roadmap.","How to read this article",{},{"id":562,"data":1687,"type":566,"tunes":1689},{"title":1688,"maxLevel":565,"minLevel":47},"Contents",{},{"id":569,"data":1691,"type":572,"tunes":1693},{"text":1692,"level":47},"The scale mismatch is already visible",{},{"id":575,"data":1695,"type":544,"tunes":1697},{"text":1696},"Newzoo forecasts the global games market at $213.9 billion in 2026, up 6.1% year over year. The Semiconductor Industry Association reported $791.7 billion in global semiconductor sales for 2025, up 25.6%, and projected roughly $1 trillion for 2026. The games business and semiconductor business still depend on each other, but their incentives are becoming less symmetric: the same advanced manufacturing ecosystem that enables premium gaming now serves AI and data-center buyers with much larger infrastructure budgets. These are not directly comparable product categories; they establish the scale of the upstream compute economy around gaming.",{},{"id":580,"data":1699,"type":604,"tunes":1721},{"content":1700,"stretched":603,"withHeadings":15},[1701,1705,1709,1713,1717],[1702,1703,1704],"Reference point","Reported or forecast value","Why it matters here",[1706,1707,1708],"Global games market, 2026 forecast","$213.9B, +6.1% YoY","Large consumer entertainment market, but downstream of advanced hardware supply",[1710,1711,1712],"PC games, 2026 forecast","$45.9B, +5.3% YoY","The gaming segment most directly exposed to PC component economics",[1714,1715,1716],"Global semiconductor sales, 2025","$791.7B, +25.6% YoY","Shows the scale of the upstream industry serving gaming, AI, mobile, automotive and cloud",[1718,1719,1720],"Global semiconductor industry, 2026 projection","~$1T","Raises the opportunity cost of dedicating premium silicon only to consumer gaming",{},{"id":607,"data":1723,"type":552,"tunes":1726},{"body":1724,"title":1725,"variant":611},"The figures establish \u003Cstrong>economic context, not direct causation\u003C\u002Fstrong>. They do not prove that a specific wafer, GPU or memory package was taken from a gamer and reassigned to AI. They also do not mean gaming and semiconductors are the two largest industries in the world. The narrower, defensible point is that a $200B+ games market depends on an upstream semiconductor economy approaching trillion-dollar annual scale, where AI infrastructure now competes for advanced compute and memory under very different budget and margin conditions.","What these market numbers do — and do not — prove",{},{"id":614,"data":1728,"type":572,"tunes":1730},{"text":1729,"level":47},"Memory is where the pressure becomes concrete",{},{"id":619,"data":1732,"type":544,"tunes":1734},{"text":1733},"TrendForce has repeatedly reported that DRAM suppliers are shifting capacity toward server and HBM products as AI demand grows. In March 2026 it forecast conventional DRAM contract prices rising 58–63% quarter over quarter in Q2 and said limited GDDR allocation was constraining graphics memory supply. By Q3, TrendForce said consumer customers were reaching an affordability limit even as AI and server demand remained strong.",{},{"id":624,"data":1736,"type":644,"tunes":1754},{"steps":1737,"title":1753,"orientation":643},[1738,1741,1744,1747,1750],{"label":1739,"description":1740},"1. AI infrastructure expands","Cloud providers and AI builders secure server memory, HBM, enterprise SSDs and advanced compute.",{"label":1742,"description":1743},"2. Suppliers optimize product mix","Capacity moves toward products with stronger demand, margins or long-term contracts.",{"label":1745,"description":1746},"3. Consumer allocation tightens","PC, graphics and consumer memory buyers face higher prices or constrained availability.",{"label":1748,"description":1749},"4. Gaming vendors absorb or pass through cost","Board design, VRAM capacity and system price become harder trade-offs.",{"label":1751,"description":1752},"5. The old upgrade cycle weakens","A faster GPU is still useful, but brute-force scaling becomes economically harder to repeat at the same consumer price points.","The resource-allocation pressure",{},{"id":647,"data":1756,"type":552,"tunes":1759},{"body":1757,"title":1758,"variant":611},"TrendForce reported on September 30, 2026 that suppliers still prioritize advanced-process capacity for high-performance server products, with conventional DRAM contract prices projected to rise another 10–15% QoQ in Q4. Its DRAM bulletin also says HBM continues to crowd out general DRAM capacity. The pressure described here therefore remains visible late in 2026; it is not based only on a temporary early-year spike.","A September 2026 signal, not an old shortage story",{},{"id":653,"data":1761,"type":572,"tunes":1763},{"text":1762,"level":47},"At exactly the same time, the frame itself is changing",{},{"id":658,"data":1765,"type":544,"tunes":1767},{"text":1766},"Microsoft now describes machine learning as becoming foundational to real-time graphics. DirectX is adding shader-level and model-level ML paths so neural texture compression, denoising, super resolution and other learned workloads can sit inside the graphics stack instead of being bolted on as isolated features. NVIDIA's DLSS 5 research pushes the idea further by introducing a generative neural rendering stage that complements conventional rendering with learned appearance priors. Neural rendering does not mean conventional rendering disappears. The emerging model is hybrid: render some information directly, infer or reconstruct other information when doing so is cheaper or more useful.",{},{"id":663,"data":1769,"type":669,"tunes":1774},{"url":1770,"title":1771,"excerpt":1772,"ctaLabel":1773},"https:\u002F\u002Ffigure.rocks\u002Fblog\u002Fdlss-5-is-not-just-upscaling-what-3d-guided-neural-rendering-actually-changes","DLSS 5 Is Not Just Upscaling: What Generative Neural Rendering Actually Changes","Why the important change is not another quality mode, but a new stage in how visual information can be produced.","Read the DLSS 5 analysis",{},{"id":672,"data":1776,"type":707,"tunes":1804},{"rows":1777,"title":1798,"layout":604,"columns":1799},[1778,1783,1788,1793],{"id":676,"label":1779,"values":1780},"More visual fidelity",{"new":1781,"old":1782},"Store\u002Frender selectively, reconstruct or infer some detail","Store more detail and render more work",{"id":682,"label":1784,"values":1785},"Memory response",{"new":1786,"old":1787},"Add capacity where needed, but also compress and reduce movement","Add capacity and bandwidth",{"id":688,"label":1789,"values":1790},"Performance response",{"new":1791,"old":1792},"Mix shader, RT and matrix\u002FTensor-style compute","More shader\u002FRT throughput",{"id":694,"label":1794,"values":1795},"Software role",{"new":1796,"old":1797},"Engine increasingly schedules rendering, data movement and ML workloads","Engine schedules mostly conventional graphics work","Traditional scaling vs neural-era scaling",[1800,1802],{"id":702,"label":1801},"Traditional path",{"id":705,"label":1803},"Neural-era path",{},{"id":710,"data":1806,"type":552,"tunes":1809},{"body":1807,"title":1808,"variant":714},"\u003Cstrong>AI may be weakening the economics of brute-force gaming hardware while simultaneously supplying techniques that make brute-force scaling less necessary.\u003C\u002Fstrong> That does not mean future games need less hardware. It means more of the result may come from deciding what to store, move, render, reconstruct or infer. Figure Rocks shorthand: \u003Cstrong>store less → move less → predict more → render less → infer more.\u003C\u002Fstrong> This is a hypothesis about where performance gains may increasingly come from, not a claim that every future game will use one pipeline.","The central Figure Rocks hypothesis",{},{"id":717,"data":1811,"type":719,"tunes":1812},{},{},{"id":722,"data":1814,"type":572,"tunes":1816},{"text":1815,"level":47},"The 5090 question is useful only if we phrase it correctly",{},{"id":727,"data":1818,"type":544,"tunes":1820},{"text":1819},"The RTX 5090 is a useful example precisely because the evidence is mixed. As of October 18, 2026, NVIDIA still lists the GeForce RTX 5090 as a current product. At the same time, retailer behavior and press reports have documented unusually poor availability and unconfirmed claims of a production halt. That is a market signal, not an official end-of-production announcement. The defensible question is therefore not whether NVIDIA secretly killed one card, but whether the opportunity cost of allocating top-tier Blackwell-class silicon to enthusiast gaming rises when similar architecture can serve professional visualization and AI markets with very different economics.",{},{"id":732,"data":1822,"type":552,"tunes":1825},{"body":1823,"title":1824,"variant":559},"\u003Cstrong>FACT:\u003C\u002Fstrong> NVIDIA still publicly lists the RTX 5090, and Blackwell architecture spans gaming and professional products.\u003Cbr>\u003Cstrong>INDUSTRY SIGNAL:\u003C\u002Fstrong> retailer disappearance, extreme scarcity and reports of a possible production halt exist, but NVIDIA has not confirmed an RTX 5090 EOL or production stop.\u003Cbr>\u003Cstrong>TECHNICAL INFERENCE:\u003C\u002Fstrong> a high-end die has multiple economically valuable destinations.\u003Cbr>\u003Cstrong>FIGURE ROCKS HYPOTHESIS:\u003C\u002Fstrong> AI and professional demand can make the traditional enthusiast expectation — that the largest practical die remains broadly attainable as a gaming flagship — harder to sustain.","FACT \u002F SIGNAL \u002F HYPOTHESIS",{},{"id":738,"data":1827,"type":572,"tunes":1829},{"text":1828,"level":47},"Memory is starting to become a compute problem, not only a capacity problem",{},{"id":743,"data":1831,"type":544,"tunes":1833},{"text":1832},"There are at least four ways to attack a memory bottleneck without simply increasing raw DRAM speed: compress the representation, reconstruct detail with compute, predict what data will be needed next, and keep the right data resident close to the execution unit. None of these makes a physical memory chip magically faster. They change how much useful work the system obtains from a finite capacity and bandwidth budget.",{},{"id":748,"data":1835,"type":604,"tunes":1857},{"content":1836,"stretched":603,"withHeadings":15},[1837,1841,1845,1849,1853],[1838,1839,1840],"Technique","What changes","What does not change automatically",[1842,1843,1844],"Compression","Fewer physical bytes represent the same logical data","The DRAM clock and bus width",[1846,1847,1848],"Neural reconstruction","Some detail is generated from a compact representation","The cost disappears; inference has its own compute\u002Flatency cost",[1850,1851,1852],"Predictive prefetch","Data arrives before the requesting workload stalls","Total physical bandwidth",[1854,1855,1856],"Residency \u002F unified access","Less unnecessary migration or duplication","Contention and latency constraints",{},{"id":773,"data":1859,"type":552,"tunes":1862},{"body":1860,"title":1861,"variant":551},"NVIDIA Research's Buddy Compression reported an average \u003Cstrong>2.2× compression ratio for HPC workloads and 1.5× for deep-learning workloads\u003C\u002Fstrong>, with 1–2% slowdown in that study. That result is not a gaming benchmark. Its relevance here is architectural: compressed memory can increase effective capacity and reduce some traffic without changing the physical DRAM clock.","Research anchor: effective capacity is not raw speed",{},{"id":779,"data":1864,"type":552,"tunes":1867},{"body":1865,"title":1866,"variant":714},"NVIDIA Research's material-texture NTC work demonstrated \u003Cstrong>4× higher linear texture resolution — 16× more texels — while using 30% less memory\u003C\u002Fstrong> than its BC High comparison. That does not mean every game gets a 16× memory gain. It demonstrates the mechanism that matters here: some visual detail can be represented more compactly and reconstructed with compute instead of being stored conventionally at the same fidelity.","Gaming-specific evidence: neural texture compression",{},{"id":785,"data":1869,"type":552,"tunes":1872},{"body":1870,"title":1871,"variant":611},"A 460× figure sometimes associated with NVIDIA systems refers to a comparison in GPU-accessible memory capacity across DGX generations, \u003Cstrong>not a 460× memory-compression ratio\u003C\u002Fstrong>. This article deliberately does not use it as evidence for compression.","What the “460×” number does not mean",{},{"id":791,"data":1874,"type":572,"tunes":1876},{"text":1875,"level":47},"How the technical response could work",{},{"id":796,"data":1878,"type":644,"tunes":1899},{"steps":1879,"title":1898,"orientation":643},[1880,1883,1886,1889,1892,1895],{"label":1881,"description":1882},"Compress","Keep textures or learned representations smaller where quality permits.",{"label":1884,"description":1885},"Predict","Use deterministic game state and learned models to estimate what data will be needed next.",{"label":1887,"description":1888},"Prefetch","Move useful data before the GPU stalls waiting for it.",{"label":1890,"description":1891},"Render","Spend conventional graphics work where it changes the visible result most.",{"label":1893,"description":1894},"Infer","Use neural reconstruction or generation where inference is cheaper than storing or simulating everything directly.",{"label":1896,"description":1897},"Present","Deliver the frame while keeping latency, artifacts and frame pacing inside acceptable limits.","One possible future frame pipeline",{},{"id":819,"data":1901,"type":544,"tunes":1903},{"text":1902},"This pipeline is not one announced product. Every component already exists in some form, but the integrated chain is the hypothesis. Neural texture compression has shown that learned representations can exchange compute for memory. GPU prefetch research has shown that learned prediction can improve memory-page hit rates. DirectX is formalizing ML inside the graphics programming model. DLSS 5 shows a generative stage entering real-time rendering.",{},{"id":824,"data":1905,"type":572,"tunes":1907},{"text":1906,"level":47},"A game has one advantage general workloads often do not: it already knows part of its future",{},{"id":829,"data":1909,"type":544,"tunes":1911},{"text":1910},"A game engine knows the camera transform, player velocity, current world cell, neighboring streaming cells, animation state, active entities, LOD transitions, shader dependencies and often likely traversal paths. That deterministic context does not reveal the future perfectly, but it gives a predictor a richer starting point than a generic memory trace alone.",{},{"id":834,"data":1913,"type":644,"tunes":1931},{"steps":1914,"title":1930,"orientation":643},[1915,1918,1921,1924,1927],{"label":1916,"description":1917},"1. Observe deterministic state","Camera direction, movement vector, current room, visible portals and known asset dependencies already exist in engine state.",{"label":1919,"description":1920},"2. Estimate the near-future working set","A learned predictor can rank likely next rooms, textures, geometry, shaders or neural representations.",{"label":1922,"description":1923},"3. Prefetch during available bandwidth","Move likely resources while the memory system has headroom instead of waiting for a hard demand miss.",{"label":1925,"description":1926},"4. Keep deterministic guarantees","Critical assets still use authored rules; prediction supplements rather than replaces correctness.",{"label":1928,"description":1929},"5. Measure the miss cost","If prediction wastes bandwidth or evicts useful data, confidence or prefetch depth must fall.","A predictive gaming-memory example",{},{"id":855,"data":1933,"type":552,"tunes":1936},{"body":1934,"title":1935,"variant":559},"A published CPU–GPU unified-virtual-memory study reported a \u003Cstrong>10.89% performance improvement\u003C\u002Fstrong>, page-hit rate rising from \u003Cstrong>76.10% to 89.02%\u003C\u002Fstrong>, and \u003Cstrong>11.05% lower CPU–GPU interconnect traffic\u003C\u002Fstrong> using learned prefetching. A 2026 AAAI paper on asynchronous KV-cache prefetching reported \u003Cstrong>2.15× attention-kernel efficiency and up to 1.97× end-to-end throughput\u003C\u002Fstrong> on its LLM workload. Neither result should be copied onto games; both demonstrate that prediction and prefetch can hide memory cost in real GPU workloads.","This idea has a research precedent, but not yet a gaming proof",{},{"id":861,"data":1938,"type":572,"tunes":1940},{"text":1939,"level":47},"Why a Spark-like architecture becomes interesting",{},{"id":866,"data":1942,"type":544,"tunes":1944},{"text":1943},"DGX Spark is not a gaming PC, and treating it as one would hide the useful lesson. Its GB10 design combines Arm CPU cores, a Blackwell GPU, Tensor and RT hardware, and up to 128 GB of coherent unified system memory. That architecture asks a different question from a classic desktop: instead of splitting the system cleanly into CPU memory and discrete VRAM, what happens when multiple compute engines share a large memory pool?",{},{"id":871,"data":1946,"type":604,"tunes":1966},{"content":1947,"stretched":603,"withHeadings":15},[1948,1952,1956,1959,1962],[1949,1950,1951],"Property","DGX Spark reference","Gaming implication if the principle were adapted",[1953,1954,1955],"Memory","Up to 128 GB coherent unified LPDDR5X","Large shared pool could reduce duplication and give local AI + graphics access to common data",[1957,884,1958],"Memory bandwidth","Capacity is not enough; high-end graphics still needs aggressive bandwidth and latency management",[1960,888,1961],"Compute mix","A future game workload could be split across more specialized engines",[1963,1964,1965],"Form factor\u002Fpower","Compact GB10 system","Integration can improve efficiency but may reduce modular upgradeability",{},{"id":896,"data":1968,"type":552,"tunes":1971},{"body":1969,"title":1970,"variant":611},"A large unified pool does \u003Cstrong>not\u003C\u002Fstrong> automatically replace the bandwidth of a high-end discrete GPU. DGX Spark lists 273 GB\u002Fs of memory bandwidth, while NVIDIA's RTX PRO 6000 Blackwell Workstation Edition lists 1,792 GB\u002Fs. These are very different products and the comparison is not a gaming benchmark; it simply makes the architectural constraint visible. A Spark-like gaming design would need to solve contention, latency and bandwidth — not just provide more shared gigabytes.","Capacity is not bandwidth",{},{"id":902,"data":1973,"type":572,"tunes":1975},{"text":1974,"level":47},"What would have to be true for this transition to work?",{},{"id":907,"data":1977,"type":644,"tunes":1998},{"steps":1978,"title":1997,"orientation":643},[1979,1982,1985,1988,1991,1994],{"label":1980,"description":1981},"1. Inference must save more than it costs","Neural work has to remove enough storage, bandwidth or conventional rendering cost to justify its own compute and latency.",{"label":1983,"description":1984},"2. The engine must expose useful state","Prediction works better when the system knows camera direction, asset dependencies, movement and likely visibility.",{"label":1986,"description":1987},"3. Memory movement must become deliberate","Compression without good residency and prefetch policy can still stall.",{"label":1989,"description":1990},"4. Quality must remain stable","Generated detail cannot introduce unacceptable artifacts, temporal instability or incorrect game information.",{"label":1992,"description":1993},"5. Latency must stay bounded","A prettier frame is not automatically a more responsive game.",{"label":1995,"description":1996},"6. The stack must survive vendor differences","Useful techniques need APIs, drivers and engine paths that do not fragment into one-off demos.","The Neural Gaming Viability Test",{},{"id":931,"data":2000,"type":572,"tunes":2002},{"text":2001,"level":47},"What should a buyer expect from a Spark-like transition?",{},{"id":936,"data":2004,"type":707,"tunes":2037},{"rows":2005,"title":2031,"layout":604,"columns":2032},[2006,2011,2016,2021,2026],{"id":940,"label":2007,"values":2008},"Large coherent pool",{"cost":2009,"gain":2010},"Shared clients can contend for the same bandwidth","More flexible sharing between CPU, graphics and local AI",{"id":946,"label":2012,"values":2013},"Less duplication",{"cost":2014,"gain":2015},"Coherence and scheduling add complexity","Fewer copies between separate system-RAM and VRAM pools",{"id":952,"label":2017,"values":2018},"More local AI headroom",{"cost":2019,"gain":2020},"AI can compete with graphics for compute and memory","Larger models, agents or neural assets can coexist locally",{"id":958,"label":2022,"values":2023},"System-level optimization",{"cost":2024,"gain":2025},"Tighter integration can reduce upgradeability and increase lock-in","The platform can optimize data placement and accelerators together",{"id":964,"label":2027,"values":2028},"New workload flexibility",{"cost":2029,"gain":2030},"Only if engines and runtimes actually use those capabilities","A machine may age better if future games shift toward neural workloads","Potential benefit vs real cost",[2033,2035],{"id":972,"label":2034},"Potential gain",{"id":975,"label":2036},"Cost \u002F risk",{},{"id":979,"data":2039,"type":572,"tunes":2041},{"text":2040,"level":47},"What would falsify this hypothesis?",{},{"id":984,"data":2043,"type":544,"tunes":2045},{"text":2044},"This thesis is deliberately falsifiable. The table below defines the market, software, hardware and game-adoption signals that would make the argument stronger or weaker over time.",{},{"id":989,"data":2047,"type":707,"tunes":2075},{"rows":2048,"title":2069,"layout":604,"columns":2070},[2049,2054,2059,2064],{"id":682,"label":2050,"values":2051},"Memory economics",{"weakens":2052,"strengthens":2053},"Consumer capacity expands and price pressure normalizes","Server\u002FHBM priority continues to squeeze consumer allocation",{"id":998,"label":2055,"values":2056},"Graphics stack",{"weakens":2057,"strengthens":2058},"Neural graphics remains vendor-specific optional post-processing","ML primitives become normal engine\u002FDirectX features",{"id":1004,"label":2060,"values":2061},"Hardware design",{"weakens":2062,"strengthens":2063},"Discrete GPU + separate RAM remains clearly superior on cost, flexibility and latency","More integrated heterogeneous compute and larger coherent memory pools",{"id":1010,"label":2065,"values":2066},"Game adoption",{"weakens":2067,"strengthens":2068},"Tech remains limited to demos or a few flagship games","Studios ship neural compression, generation and predictive data paths broadly","Evidence that strengthens or weakens the thesis",[2071,2073],{"id":1018,"label":2072},"Would strengthen it",{"id":1021,"label":2074},"Would weaken it",{},{"id":1025,"data":2077,"type":572,"tunes":2079},{"text":2078,"level":47},"The practical conclusion for a gamer today",{},{"id":1030,"data":2081,"type":544,"tunes":2083},{"text":2082},"Do not buy hardware because a speculative architecture might arrive. Buy for the games, resolution, latency target and workloads you actually have. But do not assume that future-proofing can still be reduced to one number such as raster performance or VRAM capacity. The direction of the software stack suggests that memory behavior, neural compute and engine support may become increasingly important parts of the value equation.",{},{"id":1035,"data":2085,"type":669,"tunes":2090},{"url":2086,"title":2087,"excerpt":2088,"ctaLabel":2089},"https:\u002F\u002Ffigure.rocks\u002Fblog\u002Fvram-usage-is-not-vram-requirement-why-a-full-memory-meter-does-not-tell-the-whole-story","VRAM Usage Is Not VRAM Requirement: Why a Full Memory Meter Does Not Tell the Whole Story","A practical explanation of working sets, residency and why an allocated VRAM number is not the same thing as a hard requirement.","Read the VRAM guide",{},{"id":1043,"data":2092,"type":572,"tunes":2094},{"text":2093,"level":47},"Primary sources",{},{"id":1048,"data":2096,"type":1055,"tunes":2101},{"link":1050,"meta":2097},{"image":2098,"title":2099,"description":2100},{"url":13},"Newzoo — 2026 Global Games Market Key Numbers","2026 games-market forecast used for the scale comparison.",{},{"id":1058,"data":2103,"type":1055,"tunes":2108},{"link":1060,"meta":2104},{"image":2105,"title":2106,"description":2107},{"url":13},"Semiconductor Industry Association — 2025 Global Sales","2025 semiconductor sales and the association's roughly $1 trillion 2026 projection.",{},{"id":1067,"data":2110,"type":1055,"tunes":2115},{"link":1069,"meta":2111},{"image":2112,"title":2113,"description":2114},{"url":13},"TrendForce — AI Server Demand Sustains Memory Price Increases in 4Q26","Late-2026 evidence for ongoing server-priority capacity allocation and consumer pressure.",{},{"id":1076,"data":2117,"type":1055,"tunes":2121},{"link":1078,"meta":2118},{"image":2119,"title":1554,"description":2120},{"url":13},"Research source for 2.2× HPC and 1.5× DL average compression ratios.",{},{"id":1085,"data":2123,"type":1055,"tunes":2128},{"link":1087,"meta":2124},{"image":2125,"title":2126,"description":2127},{"url":13},"Deep-learning based CPU–GPU UVM prefetching","Research source for learned page-prefetching results.",{},{"id":1094,"data":2130,"type":1055,"tunes":2135},{"link":1096,"meta":2131},{"image":2132,"title":2133,"description":2134},{"url":13},"AAAI 2026 — Asynchronous KV Cache Prefetching","Research source demonstrating latency hiding through proactive GPU-cache prefetching.",{},{"id":1103,"data":2137,"type":1055,"tunes":2142},{"link":1105,"meta":2138},{"image":2139,"title":2140,"description":2141},{"url":13},"TrendForce — AI Server Demand and 2Q26 Memory Prices","Current evidence for server\u002FHBM prioritization, GDDR constraints and consumer pressure.",{},{"id":1112,"data":2144,"type":1055,"tunes":2149},{"link":1114,"meta":2145},{"image":2146,"title":2147,"description":2148},{"url":13},"Microsoft DirectX — Evolving DirectX for the ML Era","Official GDC 2026 overview of shader-level and model-level ML in DirectX.",{},{"id":1121,"data":2151,"type":1055,"tunes":2156},{"link":1123,"meta":2152},{"image":2153,"title":2154,"description":2155},{"url":13},"NVIDIA Research — DLSS 5: Generative Neural Rendering","Primary research page describing DLSS 5 as a real-time generative rendering stage.",{},{"id":1130,"data":2158,"type":1055,"tunes":2163},{"link":1132,"meta":2159},{"image":2160,"title":2161,"description":2162},{"url":13},"NVIDIA — DGX Spark Specifications","Official architecture, unified memory and bandwidth specifications.",{},{"id":1139,"data":2165,"type":1055,"tunes":2169},{"link":1141,"meta":2166},{"image":2167,"title":1144,"description":2168},{"url":13},"Official current RTX 5090 product page used to distinguish confirmed product status from production-halt rumors.",{},{"id":1148,"data":2171,"type":1055,"tunes":2176},{"link":1150,"meta":2172},{"image":2173,"title":2174,"description":2175},{"url":13},"Tom’s Hardware — RTX 5090 Retail Disappearance and Production-Halt Rumors","Current market report used only as an industry signal; it explicitly notes that NVIDIA has not confirmed a production halt.",{},{"id":1157,"data":2178,"type":1055,"tunes":2183},{"link":1159,"meta":2179},{"image":2180,"title":2181,"description":2182},{"url":13},"NVIDIA Research — Random-Access Neural Compression of Material Textures","Primary source for the 4× linear resolution \u002F 16× texel comparison at 30% less memory versus BC High.",{},{"id":1166,"data":2185,"type":1055,"tunes":2190},{"link":1168,"meta":2186},{"image":2187,"title":2188,"description":2189},{"url":13},"NVIDIA — RTX PRO 6000 Blackwell Workstation Edition","Official source for 96 GB GDDR7 and 1,792 GB\u002Fs memory bandwidth used as a bandwidth reference point.",{},{"id":1175,"data":2192,"type":1195,"tunes":2207},{"title":2193,"entries":2194},"Key concepts",[2195,2198,2201,2204],{"term":2196,"anchor":1181,"definition":2197},"Neural rendering","Rendering methods that use trained models as part of producing, reconstructing or enhancing visual output.",{"term":2199,"anchor":1185,"definition":2200},"Heterogeneous compute","A system that assigns different workloads to different compute engines such as CPU, GPU, RT and matrix\u002FTensor hardware.",{"term":2202,"anchor":1189,"definition":2203},"Effective memory","A practical description of how much useful data a memory system can keep and feed to compute after compression, residency and movement efficiency are considered. It is not a formal replacement for physical capacity.",{"term":2205,"anchor":1193,"definition":2206},"Predictive memory","Figure Rocks shorthand for combining known game state and learned prediction to prepare likely-needed data before a stall. It is a hypothesis, not an industry standard.",{},{"id":541,"data":2209,"type":544,"tunes":2210},{"text":1674},{},{"id":547,"data":2212,"type":552,"tunes":2213},{"body":1678,"title":1679,"variant":551},{},{"id":555,"data":2215,"type":552,"tunes":2216},{"body":1683,"title":1684,"variant":559},{},{"id":562,"data":2218,"type":566,"tunes":2219},{"title":1688,"maxLevel":565,"minLevel":47},{},{"id":569,"data":2221,"type":572,"tunes":2222},{"text":1692,"level":47},{},{"id":575,"data":2224,"type":544,"tunes":2225},{"text":1696},{},{"id":580,"data":2227,"type":604,"tunes":2234},{"content":2228,"stretched":603,"withHeadings":15},[2229,2230,2231,2232,2233],[1702,1703,1704],[1706,1707,1708],[1710,1711,1712],[1714,1715,1716],[1718,1719,1720],{},{"id":607,"data":2236,"type":552,"tunes":2237},{"body":1724,"title":1725,"variant":611},{},{"id":614,"data":2239,"type":572,"tunes":2240},{"text":1729,"level":47},{},{"id":619,"data":2242,"type":544,"tunes":2243},{"text":1733},{},{"id":624,"data":2245,"type":644,"tunes":2252},{"steps":2246,"title":1753,"orientation":643},[2247,2248,2249,2250,2251],{"label":1739,"description":1740},{"label":1742,"description":1743},{"label":1745,"description":1746},{"label":1748,"description":1749},{"label":1751,"description":1752},{},{"id":647,"data":2254,"type":552,"tunes":2255},{"body":1757,"title":1758,"variant":611},{},{"id":653,"data":2257,"type":572,"tunes":2258},{"text":1762,"level":47},{},{"id":658,"data":2260,"type":544,"tunes":2261},{"text":1766},{},{"id":663,"data":2263,"type":669,"tunes":2264},{"url":1770,"title":1771,"excerpt":1772,"ctaLabel":1773},{},{"id":672,"data":2266,"type":707,"tunes":2279},{"rows":2267,"title":1798,"layout":604,"columns":2276},[2268,2270,2272,2274],{"id":676,"label":1779,"values":2269},{"new":1781,"old":1782},{"id":682,"label":1784,"values":2271},{"new":1786,"old":1787},{"id":688,"label":1789,"values":2273},{"new":1791,"old":1792},{"id":694,"label":1794,"values":2275},{"new":1796,"old":1797},[2277,2278],{"id":702,"label":1801},{"id":705,"label":1803},{},{"id":710,"data":2281,"type":552,"tunes":2282},{"body":1807,"title":1808,"variant":714},{},{"id":717,"data":2284,"type":719,"tunes":2285},{},{},{"id":722,"data":2287,"type":572,"tunes":2288},{"text":1815,"level":47},{},{"id":727,"data":2290,"type":544,"tunes":2291},{"text":1819},{},{"id":732,"data":2293,"type":552,"tunes":2294},{"body":1823,"title":1824,"variant":559},{},{"id":738,"data":2296,"type":572,"tunes":2297},{"text":1828,"level":47},{},{"id":743,"data":2299,"type":544,"tunes":2300},{"text":1832},{},{"id":748,"data":2302,"type":604,"tunes":2309},{"content":2303,"stretched":603,"withHeadings":15},[2304,2305,2306,2307,2308],[1838,1839,1840],[1842,1843,1844],[1846,1847,1848],[1850,1851,1852],[1854,1855,1856],{},{"id":773,"data":2311,"type":552,"tunes":2312},{"body":1860,"title":1861,"variant":551},{},{"id":779,"data":2314,"type":552,"tunes":2315},{"body":1865,"title":1866,"variant":714},{},{"id":785,"data":2317,"type":552,"tunes":2318},{"body":1870,"title":1871,"variant":611},{},{"id":791,"data":2320,"type":572,"tunes":2321},{"text":1875,"level":47},{},{"id":796,"data":2323,"type":644,"tunes":2331},{"steps":2324,"title":1898,"orientation":643},[2325,2326,2327,2328,2329,2330],{"label":1881,"description":1882},{"label":1884,"description":1885},{"label":1887,"description":1888},{"label":1890,"description":1891},{"label":1893,"description":1894},{"label":1896,"description":1897},{},{"id":819,"data":2333,"type":544,"tunes":2334},{"text":1902},{},{"id":824,"data":2336,"type":572,"tunes":2337},{"text":1906,"level":47},{},{"id":829,"data":2339,"type":544,"tunes":2340},{"text":1910},{},{"id":834,"data":2342,"type":644,"tunes":2349},{"steps":2343,"title":1930,"orientation":643},[2344,2345,2346,2347,2348],{"label":1916,"description":1917},{"label":1919,"description":1920},{"label":1922,"description":1923},{"label":1925,"description":1926},{"label":1928,"description":1929},{},{"id":855,"data":2351,"type":552,"tunes":2352},{"body":1934,"title":1935,"variant":559},{},{"id":861,"data":2354,"type":572,"tunes":2355},{"text":1939,"level":47},{},{"id":866,"data":2357,"type":544,"tunes":2358},{"text":1943},{},{"id":871,"data":2360,"type":604,"tunes":2367},{"content":2361,"stretched":603,"withHeadings":15},[2362,2363,2364,2365,2366],[1949,1950,1951],[1953,1954,1955],[1957,884,1958],[1960,888,1961],[1963,1964,1965],{},{"id":896,"data":2369,"type":552,"tunes":2370},{"body":1969,"title":1970,"variant":611},{},{"id":902,"data":2372,"type":572,"tunes":2373},{"text":1974,"level":47},{},{"id":907,"data":2375,"type":644,"tunes":2383},{"steps":2376,"title":1997,"orientation":643},[2377,2378,2379,2380,2381,2382],{"label":1980,"description":1981},{"label":1983,"description":1984},{"label":1986,"description":1987},{"label":1989,"description":1990},{"label":1992,"description":1993},{"label":1995,"description":1996},{},{"id":931,"data":2385,"type":572,"tunes":2386},{"text":2001,"level":47},{},{"id":936,"data":2388,"type":707,"tunes":2403},{"rows":2389,"title":2031,"layout":604,"columns":2400},[2390,2392,2394,2396,2398],{"id":940,"label":2007,"values":2391},{"cost":2009,"gain":2010},{"id":946,"label":2012,"values":2393},{"cost":2014,"gain":2015},{"id":952,"label":2017,"values":2395},{"cost":2019,"gain":2020},{"id":958,"label":2022,"values":2397},{"cost":2024,"gain":2025},{"id":964,"label":2027,"values":2399},{"cost":2029,"gain":2030},[2401,2402],{"id":972,"label":2034},{"id":975,"label":2036},{},{"id":979,"data":2405,"type":572,"tunes":2406},{"text":2040,"level":47},{},{"id":984,"data":2408,"type":544,"tunes":2409},{"text":2044},{},{"id":989,"data":2411,"type":707,"tunes":2424},{"rows":2412,"title":2069,"layout":604,"columns":2421},[2413,2415,2417,2419],{"id":682,"label":2050,"values":2414},{"weakens":2052,"strengthens":2053},{"id":998,"label":2055,"values":2416},{"weakens":2057,"strengthens":2058},{"id":1004,"label":2060,"values":2418},{"weakens":2062,"strengthens":2063},{"id":1010,"label":2065,"values":2420},{"weakens":2067,"strengthens":2068},[2422,2423],{"id":1018,"label":2072},{"id":1021,"label":2074},{},{"id":1025,"data":2426,"type":572,"tunes":2427},{"text":2078,"level":47},{},{"id":1030,"data":2429,"type":544,"tunes":2430},{"text":2082},{},{"id":1035,"data":2432,"type":669,"tunes":2433},{"url":2086,"title":2087,"excerpt":2088,"ctaLabel":2089},{},{"id":1043,"data":2435,"type":572,"tunes":2436},{"text":2093,"level":47},{},{"id":1048,"data":2438,"type":1055,"tunes":2441},{"link":1050,"meta":2439},{"image":2440,"title":2099,"description":2100},{"url":13},{},{"id":1058,"data":2443,"type":1055,"tunes":2446},{"link":1060,"meta":2444},{"image":2445,"title":2106,"description":2107},{"url":13},{},{"id":1067,"data":2448,"type":1055,"tunes":2451},{"link":1069,"meta":2449},{"image":2450,"title":2113,"description":2114},{"url":13},{},{"id":1076,"data":2453,"type":1055,"tunes":2456},{"link":1078,"meta":2454},{"image":2455,"title":1554,"description":2120},{"url":13},{},{"id":1085,"data":2458,"type":1055,"tunes":2461},{"link":1087,"meta":2459},{"image":2460,"title":2126,"description":2127},{"url":13},{},{"id":1094,"data":2463,"type":1055,"tunes":2466},{"link":1096,"meta":2464},{"image":2465,"title":2133,"description":2134},{"url":13},{},{"id":1103,"data":2468,"type":1055,"tunes":2471},{"link":1105,"meta":2469},{"image":2470,"title":2140,"description":2141},{"url":13},{},{"id":1112,"data":2473,"type":1055,"tunes":2476},{"link":1114,"meta":2474},{"image":2475,"title":2147,"description":2148},{"url":13},{},{"id":1121,"data":2478,"type":1055,"tunes":2481},{"link":1123,"meta":2479},{"image":2480,"title":2154,"description":2155},{"url":13},{},{"id":1130,"data":2483,"type":1055,"tunes":2486},{"link":1132,"meta":2484},{"image":2485,"title":2161,"description":2162},{"url":13},{},{"id":1139,"data":2488,"type":1055,"tunes":2491},{"link":1141,"meta":2489},{"image":2490,"title":1144,"description":2168},{"url":13},{},{"id":1148,"data":2493,"type":1055,"tunes":2496},{"link":1150,"meta":2494},{"image":2495,"title":2174,"description":2175},{"url":13},{},{"id":1157,"data":2498,"type":1055,"tunes":2501},{"link":1159,"meta":2499},{"image":2500,"title":2181,"description":2182},{"url":13},{},{"id":1166,"data":2503,"type":1055,"tunes":2506},{"link":1168,"meta":2504},{"image":2505,"title":2188,"description":2189},{"url":13},{},{"id":1175,"data":2508,"type":1195,"tunes":2514},{"title":2193,"entries":2509},[2510,2511,2512,2513],{"term":2196,"anchor":1181,"definition":2197},{"term":2199,"anchor":1185,"definition":2200},{"term":2202,"anchor":1189,"definition":2203},{"term":2205,"anchor":1193,"definition":2206},{},"2.31.6","AI is making high-performance silicon and memory more valuable outside gaming while also giving games new ways to render, compress and predict. This analysis asks whether those two trends are converging into a new hardware model.",{"lang":7,"title":534,"content":536,"contentJson":2518,"excerpt":1627},{"time":538,"blocks":2519,"version":1626},[2520,2523,2526,2529,2532,2535,2538,2547,2550,2553,2556,2565,2568,2571,2574,2577,2592,2595,2598,2601,2604,2607,2610,2613,2622,2625,2628,2631,2634,2644,2647,2650,2653,2662,2665,2668,2671,2680,2683,2686,2696,2699,2716,2719,2722,2737,2740,2743,2746,2749,2754,2759,2764,2769,2774,2779,2784,2789,2794,2799,2804,2809,2814,2819,2827,2830,2833,2836,2839,2842,2845,2854,2857,2860,2863,2872,2875,2878,2881,2884,2899,2902,2905,2908,2911,2914,2917,2920,2929,2932,2935,2938,2941,2951,2954,2957,2960,2969,2972,2975,2978,2987,2990,2993,3003,3006,3023,3026,3029,3044,3047,3050,3053,3056,3061,3066,3071,3076,3081,3086,3091,3096,3101,3106,3111,3116,3121,3126],{"id":541,"data":2521,"type":544,"tunes":2522},{"text":543},{},{"id":547,"data":2524,"type":552,"tunes":2525},{"body":549,"title":550,"variant":551},{},{"id":555,"data":2527,"type":552,"tunes":2528},{"body":557,"title":558,"variant":559},{},{"id":562,"data":2530,"type":566,"tunes":2531},{"title":564,"maxLevel":565,"minLevel":47},{},{"id":569,"data":2533,"type":572,"tunes":2534},{"text":571,"level":47},{},{"id":575,"data":2536,"type":544,"tunes":2537},{"text":577},{},{"id":580,"data":2539,"type":604,"tunes":2546},{"content":2540,"stretched":603,"withHeadings":15},[2541,2542,2543,2544,2545],[584,585,586],[588,589,590],[592,593,594],[596,597,598],[600,601,602],{},{"id":607,"data":2548,"type":552,"tunes":2549},{"body":609,"title":610,"variant":611},{},{"id":614,"data":2551,"type":572,"tunes":2552},{"text":616,"level":47},{},{"id":619,"data":2554,"type":544,"tunes":2555},{"text":621},{},{"id":624,"data":2557,"type":644,"tunes":2564},{"steps":2558,"title":642,"orientation":643},[2559,2560,2561,2562,2563],{"label":628,"description":629},{"label":631,"description":632},{"label":634,"description":635},{"label":637,"description":638},{"label":640,"description":641},{},{"id":647,"data":2566,"type":552,"tunes":2567},{"body":649,"title":650,"variant":611},{},{"id":653,"data":2569,"type":572,"tunes":2570},{"text":655,"level":47},{},{"id":658,"data":2572,"type":544,"tunes":2573},{"text":660},{},{"id":663,"data":2575,"type":669,"tunes":2576},{"url":665,"title":666,"excerpt":667,"ctaLabel":668},{},{"id":672,"data":2578,"type":707,"tunes":2591},{"rows":2579,"title":699,"layout":604,"columns":2588},[2580,2582,2584,2586],{"id":676,"label":677,"values":2581},{"new":679,"old":680},{"id":682,"label":683,"values":2583},{"new":685,"old":686},{"id":688,"label":689,"values":2585},{"new":691,"old":692},{"id":694,"label":695,"values":2587},{"new":697,"old":698},[2589,2590],{"id":702,"label":703},{"id":705,"label":706},{},{"id":710,"data":2593,"type":552,"tunes":2594},{"body":712,"title":713,"variant":714},{},{"id":717,"data":2596,"type":719,"tunes":2597},{},{},{"id":722,"data":2599,"type":572,"tunes":2600},{"text":724,"level":47},{},{"id":727,"data":2602,"type":544,"tunes":2603},{"text":729},{},{"id":732,"data":2605,"type":552,"tunes":2606},{"body":734,"title":735,"variant":559},{},{"id":738,"data":2608,"type":572,"tunes":2609},{"text":740,"level":47},{},{"id":743,"data":2611,"type":544,"tunes":2612},{"text":745},{},{"id":748,"data":2614,"type":604,"tunes":2621},{"content":2615,"stretched":603,"withHeadings":15},[2616,2617,2618,2619,2620],[752,753,754],[756,757,758],[760,761,762],[764,765,766],[768,769,770],{},{"id":773,"data":2623,"type":552,"tunes":2624},{"body":775,"title":776,"variant":551},{},{"id":779,"data":2626,"type":552,"tunes":2627},{"body":781,"title":782,"variant":714},{},{"id":785,"data":2629,"type":552,"tunes":2630},{"body":787,"title":788,"variant":611},{},{"id":791,"data":2632,"type":572,"tunes":2633},{"text":793,"level":47},{},{"id":796,"data":2635,"type":644,"tunes":2643},{"steps":2636,"title":816,"orientation":643},[2637,2638,2639,2640,2641,2642],{"label":756,"description":800},{"label":802,"description":803},{"label":805,"description":806},{"label":808,"description":809},{"label":811,"description":812},{"label":814,"description":815},{},{"id":819,"data":2645,"type":544,"tunes":2646},{"text":821},{},{"id":824,"data":2648,"type":572,"tunes":2649},{"text":826,"level":47},{},{"id":829,"data":2651,"type":544,"tunes":2652},{"text":831},{},{"id":834,"data":2654,"type":644,"tunes":2661},{"steps":2655,"title":852,"orientation":643},[2656,2657,2658,2659,2660],{"label":838,"description":839},{"label":841,"description":842},{"label":844,"description":845},{"label":847,"description":848},{"label":850,"description":851},{},{"id":855,"data":2663,"type":552,"tunes":2664},{"body":857,"title":858,"variant":559},{},{"id":861,"data":2666,"type":572,"tunes":2667},{"text":863,"level":47},{},{"id":866,"data":2669,"type":544,"tunes":2670},{"text":868},{},{"id":871,"data":2672,"type":604,"tunes":2679},{"content":2673,"stretched":603,"withHeadings":15},[2674,2675,2676,2677,2678],[875,876,877],[879,880,881],[883,884,885],[887,888,889],[891,892,893],{},{"id":896,"data":2681,"type":552,"tunes":2682},{"body":898,"title":899,"variant":611},{},{"id":902,"data":2684,"type":572,"tunes":2685},{"text":904,"level":47},{},{"id":907,"data":2687,"type":644,"tunes":2695},{"steps":2688,"title":928,"orientation":643},[2689,2690,2691,2692,2693,2694],{"label":911,"description":912},{"label":914,"description":915},{"label":917,"description":918},{"label":920,"description":921},{"label":923,"description":924},{"label":926,"description":927},{},{"id":931,"data":2697,"type":572,"tunes":2698},{"text":933,"level":47},{},{"id":936,"data":2700,"type":707,"tunes":2715},{"rows":2701,"title":969,"layout":604,"columns":2712},[2702,2704,2706,2708,2710],{"id":940,"label":941,"values":2703},{"cost":943,"gain":944},{"id":946,"label":947,"values":2705},{"cost":949,"gain":950},{"id":952,"label":953,"values":2707},{"cost":955,"gain":956},{"id":958,"label":959,"values":2709},{"cost":961,"gain":962},{"id":964,"label":965,"values":2711},{"cost":967,"gain":968},[2713,2714],{"id":972,"label":973},{"id":975,"label":976},{},{"id":979,"data":2717,"type":572,"tunes":2718},{"text":981,"level":47},{},{"id":984,"data":2720,"type":544,"tunes":2721},{"text":986},{},{"id":989,"data":2723,"type":707,"tunes":2736},{"rows":2724,"title":1015,"layout":604,"columns":2733},[2725,2727,2729,2731],{"id":682,"label":993,"values":2726},{"weakens":995,"strengthens":996},{"id":998,"label":999,"values":2728},{"weakens":1001,"strengthens":1002},{"id":1004,"label":1005,"values":2730},{"weakens":1007,"strengthens":1008},{"id":1010,"label":1011,"values":2732},{"weakens":1013,"strengthens":1014},[2734,2735],{"id":1018,"label":1019},{"id":1021,"label":1022},{},{"id":1025,"data":2738,"type":572,"tunes":2739},{"text":1027,"level":47},{},{"id":1030,"data":2741,"type":544,"tunes":2742},{"text":1032},{},{"id":1035,"data":2744,"type":669,"tunes":2745},{"url":1037,"title":1038,"excerpt":1039,"ctaLabel":1040},{},{"id":1043,"data":2747,"type":572,"tunes":2748},{"text":1045,"level":47},{},{"id":1048,"data":2750,"type":1055,"tunes":2753},{"link":1050,"meta":2751},{"image":2752,"title":1053,"description":1054},{"url":13},{},{"id":1058,"data":2755,"type":1055,"tunes":2758},{"link":1060,"meta":2756},{"image":2757,"title":1063,"description":1064},{"url":13},{},{"id":1067,"data":2760,"type":1055,"tunes":2763},{"link":1069,"meta":2761},{"image":2762,"title":1072,"description":1073},{"url":13},{},{"id":1076,"data":2765,"type":1055,"tunes":2768},{"link":1078,"meta":2766},{"image":2767,"title":1081,"description":1082},{"url":13},{},{"id":1085,"data":2770,"type":1055,"tunes":2773},{"link":1087,"meta":2771},{"image":2772,"title":1090,"description":1091},{"url":13},{},{"id":1094,"data":2775,"type":1055,"tunes":2778},{"link":1096,"meta":2776},{"image":2777,"title":1099,"description":1100},{"url":13},{},{"id":1103,"data":2780,"type":1055,"tunes":2783},{"link":1105,"meta":2781},{"image":2782,"title":1108,"description":1109},{"url":13},{},{"id":1112,"data":2785,"type":1055,"tunes":2788},{"link":1114,"meta":2786},{"image":2787,"title":1117,"description":1118},{"url":13},{},{"id":1121,"data":2790,"type":1055,"tunes":2793},{"link":1123,"meta":2791},{"image":2792,"title":1126,"description":1127},{"url":13},{},{"id":1130,"data":2795,"type":1055,"tunes":2798},{"link":1132,"meta":2796},{"image":2797,"title":1135,"description":1136},{"url":13},{},{"id":1139,"data":2800,"type":1055,"tunes":2803},{"link":1141,"meta":2801},{"image":2802,"title":1144,"description":1145},{"url":13},{},{"id":1148,"data":2805,"type":1055,"tunes":2808},{"link":1150,"meta":2806},{"image":2807,"title":1153,"description":1154},{"url":13},{},{"id":1157,"data":2810,"type":1055,"tunes":2813},{"link":1159,"meta":2811},{"image":2812,"title":1162,"description":1163},{"url":13},{},{"id":1166,"data":2815,"type":1055,"tunes":2818},{"link":1168,"meta":2816},{"image":2817,"title":1171,"description":1172},{"url":13},{},{"id":1175,"data":2820,"type":1195,"tunes":2826},{"title":1177,"entries":2821},[2822,2823,2824,2825],{"term":1180,"anchor":1181,"definition":1182},{"term":1184,"anchor":1185,"definition":1186},{"term":1188,"anchor":1189,"definition":1190},{"term":1192,"anchor":1193,"definition":1194},{},{"id":541,"data":2828,"type":544,"tunes":2829},{"text":1199},{},{"id":547,"data":2831,"type":552,"tunes":2832},{"body":1203,"title":550,"variant":551},{},{"id":555,"data":2834,"type":552,"tunes":2835},{"body":1207,"title":558,"variant":559},{},{"id":562,"data":2837,"type":566,"tunes":2838},{"title":564,"maxLevel":565,"minLevel":47},{},{"id":569,"data":2840,"type":572,"tunes":2841},{"text":1214,"level":47},{},{"id":575,"data":2843,"type":544,"tunes":2844},{"text":1218},{},{"id":580,"data":2846,"type":604,"tunes":2853},{"content":2847,"stretched":603,"withHeadings":15},[2848,2849,2850,2851,2852],[584,585,586],[1225,1226,1227],[1229,1230,1231],[1233,1234,1235],[1237,1238,1239],{},{"id":607,"data":2855,"type":552,"tunes":2856},{"body":1243,"title":610,"variant":611},{},{"id":614,"data":2858,"type":572,"tunes":2859},{"text":616,"level":47},{},{"id":619,"data":2861,"type":544,"tunes":2862},{"text":1250},{},{"id":624,"data":2864,"type":644,"tunes":2871},{"steps":2865,"title":642,"orientation":643},[2866,2867,2868,2869,2870],{"label":628,"description":1256},{"label":631,"description":1258},{"label":634,"description":635},{"label":637,"description":638},{"label":640,"description":1262},{},{"id":647,"data":2873,"type":552,"tunes":2874},{"body":1266,"title":1267,"variant":611},{},{"id":653,"data":2876,"type":572,"tunes":2877},{"text":1271,"level":47},{},{"id":658,"data":2879,"type":544,"tunes":2880},{"text":1275},{},{"id":663,"data":2882,"type":669,"tunes":2883},{"url":665,"title":1279,"excerpt":667,"ctaLabel":668},{},{"id":672,"data":2885,"type":707,"tunes":2898},{"rows":2886,"title":1294,"layout":604,"columns":2895},[2887,2889,2891,2893],{"id":676,"label":677,"values":2888},{"new":1286,"old":680},{"id":682,"label":683,"values":2890},{"new":1289,"old":686},{"id":688,"label":689,"values":2892},{"new":691,"old":692},{"id":694,"label":695,"values":2894},{"new":697,"old":698},[2896,2897],{"id":702,"label":703},{"id":705,"label":706},{},{"id":710,"data":2900,"type":552,"tunes":2901},{"body":1301,"title":713,"variant":714},{},{"id":717,"data":2903,"type":719,"tunes":2904},{},{},{"id":722,"data":2906,"type":572,"tunes":2907},{"text":1308,"level":47},{},{"id":727,"data":2909,"type":544,"tunes":2910},{"text":1312},{},{"id":732,"data":2912,"type":552,"tunes":2913},{"body":1316,"title":735,"variant":559},{},{"id":738,"data":2915,"type":572,"tunes":2916},{"text":1320,"level":47},{},{"id":743,"data":2918,"type":544,"tunes":2919},{"text":1324},{},{"id":748,"data":2921,"type":604,"tunes":2928},{"content":2922,"stretched":603,"withHeadings":15},[2923,2924,2925,2926,2927],[752,753,754],[756,1331,1332],[760,1334,1335],[764,765,766],[1338,769,770],{},{"id":773,"data":2930,"type":552,"tunes":2931},{"body":1342,"title":1343,"variant":551},{},{"id":779,"data":2933,"type":552,"tunes":2934},{"body":1347,"title":782,"variant":714},{},{"id":785,"data":2936,"type":552,"tunes":2937},{"body":1351,"title":1352,"variant":611},{},{"id":791,"data":2939,"type":572,"tunes":2940},{"text":1356,"level":47},{},{"id":796,"data":2942,"type":644,"tunes":2950},{"steps":2943,"title":816,"orientation":643},[2944,2945,2946,2947,2948,2949],{"label":756,"description":1362},{"label":802,"description":1364},{"label":805,"description":1366},{"label":808,"description":1368},{"label":811,"description":812},{"label":814,"description":1371},{},{"id":819,"data":2952,"type":544,"tunes":2953},{"text":1375},{},{"id":824,"data":2955,"type":572,"tunes":2956},{"text":1379,"level":47},{},{"id":829,"data":2958,"type":544,"tunes":2959},{"text":1383},{},{"id":834,"data":2961,"type":644,"tunes":2968},{"steps":2962,"title":852,"orientation":643},[2963,2964,2965,2966,2967],{"label":838,"description":1389},{"label":841,"description":1391},{"label":844,"description":1393},{"label":847,"description":1395},{"label":1397,"description":851},{},{"id":855,"data":2970,"type":552,"tunes":2971},{"body":1401,"title":1402,"variant":559},{},{"id":861,"data":2973,"type":572,"tunes":2974},{"text":1406,"level":47},{},{"id":866,"data":2976,"type":544,"tunes":2977},{"text":1410},{},{"id":871,"data":2979,"type":604,"tunes":2986},{"content":2980,"stretched":603,"withHeadings":15},[2981,2982,2983,2984,2985],[875,1416,1417],[879,1419,1420],[883,884,885],[887,888,889],[1424,1425,1426],{},{"id":896,"data":2988,"type":552,"tunes":2989},{"body":1430,"title":899,"variant":611},{},{"id":902,"data":2991,"type":572,"tunes":2992},{"text":1434,"level":47},{},{"id":907,"data":2994,"type":644,"tunes":3002},{"steps":2995,"title":928,"orientation":643},[2996,2997,2998,2999,3000,3001],{"label":911,"description":912},{"label":914,"description":1441},{"label":1443,"description":1444},{"label":920,"description":921},{"label":923,"description":1447},{"label":1449,"description":1450},{},{"id":931,"data":3004,"type":572,"tunes":3005},{"text":1454,"level":47},{},{"id":936,"data":3007,"type":707,"tunes":3022},{"rows":3008,"title":969,"layout":604,"columns":3019},[3009,3011,3013,3015,3017],{"id":940,"label":941,"values":3010},{"cost":943,"gain":1461},{"id":946,"label":947,"values":3012},{"cost":1464,"gain":1465},{"id":952,"label":1467,"values":3014},{"cost":1469,"gain":956},{"id":958,"label":959,"values":3016},{"cost":961,"gain":962},{"id":964,"label":965,"values":3018},{"cost":967,"gain":1474},[3020,3021],{"id":972,"label":973},{"id":975,"label":1478},{},{"id":979,"data":3024,"type":572,"tunes":3025},{"text":981,"level":47},{},{"id":984,"data":3027,"type":544,"tunes":3028},{"text":1485},{},{"id":989,"data":3030,"type":707,"tunes":3043},{"rows":3031,"title":1508,"layout":604,"columns":3040},[3032,3034,3036,3038],{"id":682,"label":1491,"values":3033},{"weakens":1493,"strengthens":1494},{"id":998,"label":1496,"values":3035},{"weakens":1498,"strengthens":1499},{"id":1004,"label":1005,"values":3037},{"weakens":1502,"strengthens":1503},{"id":1010,"label":1011,"values":3039},{"weakens":1506,"strengthens":1507},[3041,3042],{"id":1018,"label":1511},{"id":1021,"label":1513},{},{"id":1025,"data":3045,"type":572,"tunes":3046},{"text":1517,"level":47},{},{"id":1030,"data":3048,"type":544,"tunes":3049},{"text":1521},{},{"id":1035,"data":3051,"type":669,"tunes":3052},{"url":1037,"title":1525,"excerpt":1526,"ctaLabel":1040},{},{"id":1043,"data":3054,"type":572,"tunes":3055},{"text":1045,"level":47},{},{"id":1048,"data":3057,"type":1055,"tunes":3060},{"link":1050,"meta":3058},{"image":3059,"title":1535,"description":1054},{"url":13},{},{"id":1058,"data":3062,"type":1055,"tunes":3065},{"link":1060,"meta":3063},{"image":3064,"title":1063,"description":1541},{"url":13},{},{"id":1067,"data":3067,"type":1055,"tunes":3070},{"link":1069,"meta":3068},{"image":3069,"title":1547,"description":1548},{"url":13},{},{"id":1076,"data":3072,"type":1055,"tunes":3075},{"link":1078,"meta":3073},{"image":3074,"title":1554,"description":1555},{"url":13},{},{"id":1085,"data":3077,"type":1055,"tunes":3080},{"link":1087,"meta":3078},{"image":3079,"title":1090,"description":1561},{"url":13},{},{"id":1094,"data":3082,"type":1055,"tunes":3085},{"link":1096,"meta":3083},{"image":3084,"title":1099,"description":1567},{"url":13},{},{"id":1103,"data":3087,"type":1055,"tunes":3090},{"link":1105,"meta":3088},{"image":3089,"title":1108,"description":1573},{"url":13},{},{"id":1112,"data":3092,"type":1055,"tunes":3095},{"link":1114,"meta":3093},{"image":3094,"title":1117,"description":1579},{"url":13},{},{"id":1121,"data":3097,"type":1055,"tunes":3100},{"link":1123,"meta":3098},{"image":3099,"title":1126,"description":1585},{"url":13},{},{"id":1130,"data":3102,"type":1055,"tunes":3105},{"link":1132,"meta":3103},{"image":3104,"title":1135,"description":1136},{"url":13},{},{"id":1139,"data":3107,"type":1055,"tunes":3110},{"link":1141,"meta":3108},{"image":3109,"title":1144,"description":1145},{"url":13},{},{"id":1148,"data":3112,"type":1055,"tunes":3115},{"link":1150,"meta":3113},{"image":3114,"title":1153,"description":1601},{"url":13},{},{"id":1157,"data":3117,"type":1055,"tunes":3120},{"link":1159,"meta":3118},{"image":3119,"title":1162,"description":1607},{"url":13},{},{"id":1166,"data":3122,"type":1055,"tunes":3125},{"link":1168,"meta":3123},{"image":3124,"title":1171,"description":1613},{"url":13},{},{"id":1175,"data":3127,"type":1195,"tunes":3133},{"title":1177,"entries":3128},[3129,3130,3131,3132],{"term":1180,"anchor":1181,"definition":1619},{"term":1184,"anchor":1185,"definition":1186},{"term":1188,"anchor":1189,"definition":1622},{"term":1192,"anchor":1193,"definition":1624},{},"Post erfolgreich abgerufen",{"items":3136,"source":3208,"manualIds":3209,"manualMatchedIds":3210},[3137,3143,3149,3155,3160,3165,3172,3178,3185,3190,3196,3202],{"id":3138,"slug":3139,"title":3140,"excerpt":3141,"featuredImage":14,"publishedAt":3142},"118","queue-management-for-gaming-stop-bufferbloat-spikes-under-load","游戏队列管理：负载下消除缓冲区膨胀峰值","Bufferbloat（缓冲膨胀）是导致有人在后台上传或直播时，你的游戏体验变得卡顿的元凶。掌握这个实用步骤：先在高负载下测量，再管理队列，最后才升级硬件。","2026-02-19T11:00:00.000Z",{"id":3144,"slug":3145,"title":3146,"excerpt":3147,"featuredImage":14,"publishedAt":3148},"399","beyond-scanning-creative-uses-and-modifications-of-amiibo","\"超越扫描——amiibo的创意用途与改造\"","amiibo是任天堂自2014年起推出的搭载NFC技术的角色手办。从技术层面看，它们是封装在塑胶模型内的微型数据载体。在实际应用中，这些产品介于玩具、收藏品与交互设备之间。将其扫描接入游戏主机仅是它们生命周期中的一环。多年观察可见，许多拥有者将其视为具备独立创作潜能的实体物件。","2026-02-27T16:36:00.000Z",{"id":3150,"slug":3151,"title":3152,"excerpt":3153,"featuredImage":14,"publishedAt":3154},"20","console-vs-pc-display-setup-the-clean-stable-way","主机与PC显示器设置：简洁稳定的方法","主机与PC需要不同的显示优先级设置。本指南旨在规避常见误区，确保动态画面与延迟表现保持稳定一致。","2026-02-19T08:00:00.000Z",{"id":3156,"slug":3157,"title":3158,"excerpt":3159,"featuredImage":14,"publishedAt":3142},"37","the-gaming-feel-chain-why-one-upgrade-often-does-nothing","游戏体验链：为何单一升级往往收效甚微","感觉是一条链：输入、系统、帧率同步、同步、显示、网络。若其中一环不稳，升级便无法落地。",{"id":3161,"slug":3162,"title":3163,"excerpt":3164,"featuredImage":14,"publishedAt":3142},"66","experience-first-why-gaming-feel-beats-specs-a-value-standard","体验至上：为何游戏感受胜过硬件参数（一种价值标准）","互联网售卖的是数字，玩家感受的是时机。这是一个实用的价值标准：体验为先，稳定其次，随后才是装备与价格。",{"id":3166,"slug":3167,"title":3168,"excerpt":3169,"featuredImage":3170,"publishedAt":3171},"457","what-does-future-proof-gaming-hardware-mean-in-the-neural-rendering-era","在神经渲染时代，“面向未来”的游戏硬件意味着什么？","过去，面向未来意味着购买比当下所需更多的 CPU、GPU 和 VRAM。神经渲染让这条规则变得不那么可靠。这里有一个更好的决策框架。","\u002Fuploads\u002F2026\u002F10\u002Fwhat-does-future-proof-gaming-hardware-mean-in-the-neural-rendering-era-1792340303825-012uhb.webp","2026-10-18T12:16:00.000Z",{"id":3173,"slug":3174,"title":3175,"excerpt":3176,"featuredImage":14,"publishedAt":3177},"198","mtu-myths-for-gaming-when-changing-mtu-helps-rare-vs-breaks-things","游戏中的MTU误区：何时调整MTU有益（罕见）与何时会引发问题","MTU调整通常是安慰剂效应，且可能导致数据包分片问题。了解MTU调整的实际适用场景，以及必须测试时应采取的安全方法。","2026-02-20T18:00:00.000Z",{"id":3179,"slug":3180,"title":3181,"excerpt":3182,"featuredImage":3183,"publishedAt":3184},"454","why-ai-is-reshaping-gaming-memory-hbm-gddr-vram-and-the-fight-for-capacity","为什么AI正在重塑游戏内存：HBM、GDDR、VRAM与容量之争","现代游戏背后的内存问题不再仅仅是GPU需要多少千兆字节。AI基础设施正在改变供应商优先考虑哪些内存产品，以及消费者买家能负担多少。","\u002Fuploads\u002F2026\u002F10\u002Fwhy-ai-is-reshaping-gaming-memory-hbm-gddr-vram-and-the-fight-for-capacity-1792335495408-ymehaw.webp","2026-10-18T10:56:00.000Z",{"id":3186,"slug":3187,"title":3188,"excerpt":3189,"featuredImage":14,"publishedAt":3142},"122","ergonomics-for-long-sessions-comfort-that-prevents-fatigue","长时间使用的工效学设计：舒适度预防疲劳","舒适即是性能。运用这一简单的人体工学基准，可有效减轻疲劳感，保持操控一致性，助您更持久畅玩而无痛感。",{"id":3191,"slug":3192,"title":3193,"excerpt":14,"featuredImage":3194,"publishedAt":3195},"456","what-would-a-spark-like-gaming-pc-actually-change-unified-memory-ai-compute-and-the-classic-gpu-split","一台类似Spark的游戏PC究竟会带来什么改变？统一内存、AI计算与经典的GPU分工","\u002Fuploads\u002F2026\u002F10\u002Fwhat-would-a-spark-like-gaming-pc-actually-change-unified-memory-ai-compute-and-the-classic-gpu-split-1792340021752-bia7n6.webp","2026-10-18T12:12:00.000Z",{"id":3197,"slug":3198,"title":3199,"excerpt":3200,"featuredImage":14,"publishedAt":3201},"402","ll-amiibo-compatible-games-complete-overview-of-bonuses-and-gameplay-integration","所有兼容amiibo的游戏——奖励与游戏玩法整合完整概览","amiibo手办可与Wii U、任天堂3DS及任天堂Switch平台上的众多任天堂游戏互动。不同游戏中的附加价值各有差异：有时奖励仅为外观装饰，有时则会直接影响游戏玩法。以下将按游戏平台分类，系统梳理兼容作品及扫描手办可解锁的核心内容。","2026-02-27T18:13:00.000Z",{"id":3203,"slug":3204,"title":3205,"excerpt":3206,"featuredImage":14,"publishedAt":3207},"40","experience-first-gaming-the-anti-race-handbook","体验至上游戏：《反竞速手册》","厌倦了规格竞赛的玩家指南。构建手感、稳定性与品味：更佳的游戏体验，更从容的升级，更长远的价值。","2026-02-19T12:00:00.000Z","fallback",[],[]]