[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"portal-settings:figure:zh":3,"public-menus:all":45,"post:directx-is-becoming-an-ml-platform-what-linear-algebra-and-neural-shaders-mean-for-future-games:zh":531,"related:post:directx-is-becoming-an-ml-platform-what-linear-algebra-and-neural-shaders-mean-for-future-games:zh:1":2067},{"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":2066},{"id":533,"title":534,"slug":535,"content":536,"contentJson":537,"excerpt":1178,"featuredImage":1179,"featuredImageAlt":1180,"featuredImageCaption":14,"featuredImageTitle":14,"featuredImageCopyright":14,"featuredImageAuthor":14,"featuredImageSourceUrl":14,"featuredImageLicense":14,"featuredImageIsAiGenerated":720,"status":1181,"publishedAt":1182,"createdAt":1183,"updatedAt":1184,"seoLocalePaths":1185,"categories":1194,"author":1215,"translations":1219},"446","DirectX 正在成为一个机器学习平台：线性代数和神经着色器对未来游戏意味着什么","directx-is-becoming-an-ml-platform-what-linear-algebra-and-neural-shaders-mean-for-future-games","\u003Cp>DirectX 不再只是向 GPU 发送传统着色器的图形 API。微软正在将机器学习原语直接添加到 HLSL 中，并为在 DirectX 生态系统中运行更大的 ML 图开辟了第二条路径。这改变了神经渲染在未来 PC 游戏中的位置。\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>DirectX 正在成为一个具备 ML 能力的图形平台。\u003C\u002Fstrong>小型神经工作负载可以通过 DX Linear Algebra 直接在着色器内运行，而更大的模型则由 DirectX Compute Graph Compiler 来处理。目标是让游戏引擎能够使用 GPU AI 硬件，而无需为每种技术构建单独的厂商特定路径。\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\">当前状态\u003C\u002Fstrong>\u003Cdiv class=\"text-gray-700 dark:text-gray-200\">截至 2026 年 9 月，\u003Cstrong>DX Linear Algebra 仍是一项预览技术\u003C\u002Fstrong>，位于 Shader Model 6.10 \u002F Agility SDK 预览路径中。微软宣布 DirectX Compute Graph Compiler 为私人预览，而非广泛发布的零售功能。本文解释的是架构和方向，并不声称当前每款游戏今天都能使用它。\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\">为什么神经渲染需要新的 DirectX 原语\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-9\" class=\"editorjs-toc__link\">两级 DirectX ML 模型\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-12\" class=\"editorjs-toc__link\">DX Linear Algebra 实际上为着色器提供了什么\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-17\" class=\"editorjs-toc__link\">为什么 Cooperative Vector 只是开始\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-21\" class=\"editorjs-toc__link\">着色器内部实际可以运行什么？\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-24\" class=\"editorjs-toc__link\">为什么这对纹理内存很重要\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-29\" class=\"editorjs-toc__link\">为什么完整模型需要不同的路径\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-33\" class=\"editorjs-toc__link\">着色器或图的边界\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-35\" class=\"editorjs-toc__link\">为什么跨供应商支持比另一个 AI 功能更重要\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-40\" class=\"editorjs-toc__link\">当前预览版支持哪些硬件？\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-44\" class=\"editorjs-toc__link\">神经渲染正在成为基础设施\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-49\" class=\"editorjs-toc__link\">神经渲染堆栈正在变得分层\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-51\" class=\"editorjs-toc__link\">为什么统一性能分析很重要\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-55\" class=\"editorjs-toc__link\">这并不意味着每个着色器都会变成神经网络\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-59\" class=\"editorjs-toc__link\">神经着色器价值测试\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-61\" class=\"editorjs-toc__link\">这对游戏玩家意味着什么\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-65\" class=\"editorjs-toc__link\">什么会改变这个答案？\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-69\" class=\"editorjs-toc__link\">局限性\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-72\" class=\"editorjs-toc__link\">结论\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-76\" class=\"editorjs-toc__link\">常见问题\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-78\" class=\"editorjs-toc__link\">术语表\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-80\" class=\"editorjs-toc__link\">主要来源\u003C\u002Fa>\u003C\u002Fli>\u003C\u002Fol>\u003C\u002Fnav>\n\u003Ch2 id=\"section-5\">为什么神经渲染需要新的 DirectX 原语\u003C\u002Fh2>\n\u003Cp>现代 GPU 已经包含用于机器学习矩阵运算的专用硬件。对于游戏开发者来说，问题不仅在于该硬件是否存在，还在于如何从实时图形管线中高效地访问它。\u003C\u002Fp>\n\u003Cp>微软最初通过 Cooperative Vector 支持对此进行了探索。2026 年，这项工作演变为更广泛的 DirectX Linear Algebra 设计，同时支持向量-矩阵和矩阵-矩阵运算。\u003C\u002Fp>\n\u003Cp>这很重要，因为不同的神经渲染任务具有不同的形态。逐像素评估材质行为的小型模型与大型超分辨率或去噪图的工作负载并不相同。\u003C\u002Fp>\n\u003Ch2 id=\"section-9\">两级 DirectX ML 模型\u003C\u002Fh2>\n\u003Csection class=\"editorjs-process my-6\">\u003Ch3 class=\"mb-3 text-lg font-semibold\">ML 进入 DirectX 图形管线的两种方式\u003C\u002Fh3>\u003Cdiv class=\"flex flex-col sm:flex-row gap-3\">\u003Cdiv class=\"editorjs-process__step min-w-0 flex-1 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. 着色器级 ML\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">小型神经或线性代数工作负载直接从 HLSL 中与传统着色器代码一起执行。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__arrow shrink-0 self-center text-xl text-gray-400 rotate-90 sm:rotate-0\" aria-hidden=\"true\">→\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0 flex-1 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. DX Linear Algebra\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">着色器可以请求硬件加速的向量和矩阵运算，而无需手动实现每个 ML 原语。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__arrow shrink-0 self-center text-xl text-gray-400 rotate-90 sm:rotate-0\" aria-hidden=\"true\">→\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0 flex-1 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. 模型级 ML\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">更大的神经网络被表示为完整的计算图，而不是手写的着色器片段。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__arrow shrink-0 self-center text-xl text-gray-400 rotate-90 sm:rotate-0\" aria-hidden=\"true\">→\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0 flex-1 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. DirectX Compute Graph Compiler\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">微软计划的编译器路径会分析这些图并将其降低为与 D3D12 队列和命令列表集成的优化 GPU 工作负载。\u003C\u002Fdiv>\u003C\u002Fdiv>\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\">简单区别\u003C\u002Fstrong>\u003Cdiv class=\"text-gray-700 dark:text-gray-200\">\u003Cstrong>DX Linear Algebra：\u003C\u002Fstrong>将小型 ML 数学放入着色器内。\u003Cbr>\u003Cstrong>Compute Graph Compiler：\u003C\u002Fstrong>将更大的 ML 模型作为图引入引擎。\u003C\u002Fdiv>\u003C\u002Faside>\n\u003Ch2 id=\"section-12\">DX Linear Algebra 实际上为着色器提供了什么\u003C\u002Fh2>\n\u003Cp>传统 HLSL 围绕图形和计算操作构建。神经工作负载严重依赖线性代数：向量、矩阵、乘法、累加以及针对这些操作优化的数据布局。\u003C\u002Fp>\n\u003Cp>Shader Model 6.10 预览版添加了一流的矩阵 API，使开发者能够更直接地表达这些工作负载，并让驱动程序将它们映射到专用硬件。\u003C\u002Fp>\n\u003Cp>微软 2026 年 4 月的预览版明确将其描述为神经渲染、ML 和图像处理工作负载的统一路径，而非仅限图形的功能。\u003C\u002Fp>\n\u003Csection class=\"editorjs-comparison my-6\">\u003Ch3 class=\"mb-3 text-lg font-semibold\">传统着色器数学与面向 ML 的着色器数学\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\">传统着色器重点\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\">面向 ML 的重点\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\">典型工作\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">硬件路径\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">开发者表达\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\u003C\u002Fsection>\n\u003Ch2 id=\"section-17\">为什么 Cooperative Vector 只是开始\u003C\u002Fh2>\n\u003Cp>Cooperative Vector 允许着色器线程请求向量-矩阵工作，驱动程序可以将其映射到专用硬件。这对于高度并行的逐像素工作负载非常有用。\u003C\u002Fp>\n\u003Cp>微软后来得出结论，许多重要的机器学习工作负载需要的不仅仅是向量-矩阵运算。超分辨率、去噪、时间重建、更大的图像模型和通用推理可能需要矩阵-矩阵运算以及跨多个线程的共享工作。\u003C\u002Fp>\n\u003Cp>因此，DX 线性代数扩展了模型，而不是将协作向量视为最终抽象。\u003C\u002Fp>\n\u003Ch2 id=\"section-21\">着色器内部实际可以运行什么？\u003C\u002Fh2>\n\u003Cp>最有趣的着色器级机器学习工作负载足够小，可以在其操作的图形数据附近执行。\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\">工作负载\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">为什么着色器级机器学习适合\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">神经纹理压缩\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">小型网络可以在着色器需要纹理信息的点附近重建纹理信息\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">神经材质评估\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">学习函数可以替代或增强昂贵的手工编写材质数学\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">神经辐射缓存\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">逐像素或局部推理可以从学习到的场景行为中估计光照信息\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">小型去噪\u002F重建内核\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">机器学习操作可以直接位于其改进的渲染阶段旁边\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">图像处理推理\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">矩阵运算可以嵌入 GPU 处理中，而无需单独的外部运行时\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\n\u003Ch2 id=\"section-24\">为什么这对纹理内存很重要\u003C\u002Fh2>\n\u003Cp>微软反复提到的一个例子是神经纹理压缩。\u003C\u002Fp>\n\u003Cp>游戏可以存储更紧凑的表示，并在渲染期间使用小型神经网络重建细节，而不是以传统保真度存储每个纹理通道。\u003C\u002Fp>\n\u003Cp>这用一些 GPU 推理工作换取了更低的存储或内存压力。具体收益取决于技术和硬件，但架构变化很重要：一些视觉细节可以变成计算而不是存储数据。\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\">VRAM 使用量不是 VRAM 需求：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\">关于 VRAM 容量、驻留预算、工作集以及为什么内存压力比一个使用量数字更复杂的实用指南。\u003C\u002Fp>\u003Cspan class=\"mt-3 inline-flex text-sm font-medium text-primary-600 dark:text-primary-400\">阅读 VRAM 指南 →\u003C\u002Fspan>\u003C\u002Fdiv>\u003C\u002Fa>\u003C\u002Faside>\n\u003Ch2 id=\"section-29\">为什么完整模型需要不同的路径\u003C\u002Fh2>\n\u003Cp>在 HLSL 中手写一些矩阵运算对于小型神经函数是可行的。当工作负载是具有许多层、依赖关系和中间张量的完整现代模型时，这就变得非常不切实际。\u003C\u002Fp>\n\u003Cp>微软的 DirectX 计算图编译器就是为这类更大的工作负载设计的。\u003C\u002Fp>\n\u003Cp>编译器可以接受计算图、分析整个图、规划内存、融合操作，并将结果降低为与 DirectX 12 集成的 GPU 工作，而不是将模型重写为自定义着色器代码。\u003C\u002Fp>\n\u003Ch2 id=\"section-33\">着色器或图的边界\u003C\u002Fh2>\n\u003Csection class=\"editorjs-comparison my-6\">\u003Ch3 class=\"mb-3 text-lg font-semibold\">机器学习工作负载何时属于 HLSL 与模型编译器\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\">着色器级路径\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\">模型级路径\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\">模型大小\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">执行风格\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">创作\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">优化范围\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">典型用途\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\u003C\u002Fsection>\n\u003Ch2 id=\"section-35\">为什么跨供应商支持比另一个 AI 功能更重要\u003C\u002Fh2>\n\u003Cp>AMD、Intel、NVIDIA 和高通展示了不同的 GPU 架构和不同形式的专用矩阵加速。\u003C\u002Fp>\n\u003Cp>DirectX 抽象层为微软和驱动程序供应商提供了一个平台，将通用的 HLSL 或图级别 ML 转换为正确的硬件路径。\u003C\u002Fp>\n\u003Cp>这并不会让每个 GPU 都同样快，但它可以减少游戏引擎为每个供应商实现完全不同的神经渲染 API 的需求。\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\">可移植性才是真正的平台故事\u003C\u002Fstrong>\u003Cdiv class=\"text-gray-700 dark:text-gray-200\">有趣的部分不是 DirectX 可以“运行 AI”。GPU 本来就可以。平台的变化在于 \u003Cstrong>ML 可以通过通用的图形 API 和工具链来表达\u003C\u002Fstrong>。\u003C\u002Fdiv>\u003C\u002Faside>\n\u003Ch2 id=\"section-40\">当前预览版支持哪些硬件？\u003C\u002Fh2>\n\u003Cp>答案取决于具体的线性代数操作和预览版驱动程序。\u003C\u002Fp>\n\u003Cp>微软的 Agility SDK 1.721 预览版支持表列出了 LinAlg VectorAccumulate 对 AMD Radeon RX 9000 系列硬件、通过即将推出的驱动程序支持的 Intel Xe2 或更新硬件，以及通过支持的预览路径的 NVIDIA RTX 硬件的支持。\u003C\u002Fp>\n\u003Cp>这是预览时代的支持，不是通用的零售保证。硬件和驱动程序支持在最终确定之前可能会发生变化。\u003C\u002Fp>\n\u003Ch2 id=\"section-44\">神经渲染正在成为基础设施\u003C\u002Fh2>\n\u003Cp>DLSS、FSR 和其他神经图形技术通常被讨论为游戏设置菜单中可见的品牌功能。\u003C\u002Fp>\n\u003Cp>DirectX 线性代数指向更深层次的变化。神经操作可以成为渲染器内部的实现细节，而不是单一的可选后处理功能。\u003C\u002Fp>\n\u003Cp>开发者可以将 ML 用于纹理、材质、光照、重建或其他局部功能，而无需将每个功能都暴露为面向消费者的 AI 开关。\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：3D 引导的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\">神经渲染如何超越超分辨率和生成帧，进入图形管线内的材质和光照重建。\u003C\u002Fp>\u003Cspan class=\"mt-3 inline-flex text-sm font-medium text-primary-600 dark:text-primary-400\">阅读 DLSS 5 神经渲染指南 →\u003C\u002Fspan>\u003C\u002Fdiv>\u003C\u002Fa>\u003C\u002Faside>\n\u003Ch2 id=\"section-49\">神经渲染堆栈正在变得分层\u003C\u002Fh2>\n\u003Csection class=\"editorjs-process my-6\">\u003Ch3 class=\"mb-3 text-lg font-semibold\">未来可能的 DirectX 游戏管线\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\">传统引擎工作\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">模拟、几何、可见性和基础渲染仍然是标准引擎的职责。\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\">内联神经着色器\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">小型学习函数在 HLSL 内重建纹理、材质或光照。\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\">更大的 ML 图\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">完整的重建或推理模型通过图级别的 DirectX 路径执行。\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\">供应商硬件映射\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">驱动程序将常见的 DirectX 操作映射到 GPU 的专用 AI 和矩阵单元上。\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\">PIX 可见性\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">图形和 ML 工作可以一起进行性能分析，而不是存在于单独的不透明运行时中。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003C\u002Fdiv>\u003C\u002Fsection>\n\u003Ch2 id=\"section-51\">为什么统一性能分析很重要\u003C\u002Fh2>\n\u003Cp>一个改善图像质量的神经工作负载如果意外消耗帧时间、内存带宽或 VRAM，仍然可能损害游戏。\u003C\u002Fp>\n\u003Cp>微软明确将统一的 PIX 可见性作为 Compute Graph Compiler 方向的一部分。这很重要，因为开发者需要在同一帧捕获中看到图形和 ML 工作。\u003C\u002Fp>\n\u003Cp>如果神经渲染成为基础设施，它必须像任何其他渲染阶段一样可测量。\u003C\u002Fp>\n\u003Ch2 id=\"section-55\">这并不意味着每个着色器都会变成神经网络\u003C\u002Fh2>\n\u003Cp>对于许多工作负载，传统着色器数学仍然高效、确定性强且易于推理。\u003C\u002Fp>\n\u003Cp>当神经函数能够更高效地近似或重建昂贵的东西、压缩数据，或者产生原本需要过多传统计算或内存才能达到的质量时，它才有意义。\u003C\u002Fp>\n\u003Cp>正确的架构将保持混合模式。\u003C\u002Fp>\n\u003Ch2 id=\"section-59\">神经着色器价值测试\u003C\u002Fh2>\n\u003Csection class=\"editorjs-process my-6\">\u003Ch3 class=\"mb-3 text-lg font-semibold\">何时将机器学习放入渲染管线才真正有意义？\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. 识别昂贵的传统操作\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">你试图替代的是什么计算、带宽、内存或存储成本？\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. 定义神经替代方案\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">一个小模型可以重建、预测或压缩什么？\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. 测量推理成本\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">机器学习模型本身会消耗 GPU 时间、内存和带宽。\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. 测量质量稳定性\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">寻找时间伪影、重建误差和失败案例。\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. 测量净节省\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">只有当避免的传统成本值得增加的机器学习成本时，该技术才有用。\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. 跨供应商测试\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">DirectX 抽象有助于可移植性，但实际硬件性能仍然存在差异。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003C\u002Fdiv>\u003C\u002Fsection>\n\u003Ch2 id=\"section-61\">这对游戏玩家意味着什么\u003C\u002Fh2>\n\u003Cp>游戏玩家可能永远不会在图形菜单中看到“DX 线性代数”开关。\u003C\u002Fp>\n\u003Cp>可能的影响是间接的：更小的纹理占用、更好的光照重建、更高效的神经图形，或者因为引擎可以通过标准路径访问矩阵加速而变得实用的新视觉技术。\u003C\u002Fp>\n\u003Cp>该功能作为基础设施比作为品牌更重要。\u003C\u002Fp>\n\u003Ch2 id=\"section-65\">什么会改变这个答案？\u003C\u002Fh2>\n\u003Cp>最大的不确定性是这些 API 的最终零售形态。DX 线性代数仍处于预览阶段，而计算图编译器尚未达到广泛的零售可用性。\u003C\u002Fp>\n\u003Cp>在这些系统成为正常的发行游戏基础设施之前，最终的硬件支持、API 细节、编译器行为和引擎采用可能会发生变化。\u003C\u002Fp>\n\u003Cp>如果主要引擎直接采用这些抽象，神经着色器可能会变得更加普遍，而无需各个游戏团队从头实现每一项技术。\u003C\u002Fp>\n\u003Ch2 id=\"section-69\">局限性\u003C\u002Fh2>\n\u003Cp>本文解释了微软记录的 DirectX 架构和预览 API。它并不声称 DX 线性代数目前能提高每款游戏的性能，也不声称计算图编译器是成品零售产品。\u003C\u002Fp>\n\u003Cp>微软的示例描述了能力和预期用途。实际收益取决于模型、引擎集成、GPU 架构、驱动程序和负载。\u003C\u002Fp>\n\u003Ch2 id=\"section-72\">结论\u003C\u002Fh2>\n\u003Cp>DirectX 的重要变化并不是又一个名为 AI 的复选框。\u003C\u002Fp>\n\u003Cp>而是机器学习数学正在进入图形编程模型本身。小型神经函数可以驻留在着色器内，更大的模型可以走向图级编译，而同一套 DirectX 工具链可以在多个 GPU 厂商之间暴露这些工作负载。\u003C\u002Fp>\n\u003Cp>这就是神经渲染如何不再只是一个品牌化功能，而开始成为渲染基础设施的一部分。\u003C\u002Fp>\n\u003Ch2 id=\"section-76\">常见问题\u003C\u002Fh2>\n\u003Csection class=\"editorjs-faq my-6 rounded-xl border border-gray-200 p-5 dark:border-gray-700\">\u003Ch3 class=\"mb-3 text-lg font-semibold\">DirectX 线性代数与神经着色器\u003C\u002Fh3>\u003Cdiv id=\"faq1\" class=\"border-t border-gray-200 py-4 first:border-t-0 dark:border-gray-700\">\u003Ch4 class=\"font-semibold text-gray-900 dark:text-gray-100\">什么是 DirectX 线性代数？\u003C\u002Fh4>\u003Cdiv class=\"mt-2 text-gray-600 dark:text-gray-300\">它是一组 DirectX\u002FHLSL 功能，用于机器学习、神经渲染和图像处理工作负载所使用的硬件加速向量和矩阵运算。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv id=\"faq2\" class=\"border-t border-gray-200 py-4 first:border-t-0 dark:border-gray-700\">\u003Ch4 class=\"font-semibold text-gray-900 dark:text-gray-100\">什么是神经着色器？\u003C\u002Fh4>\u003Cdiv class=\"mt-2 text-gray-600 dark:text-gray-300\">一个有用的通俗描述是：一种将学习到的神经函数或 ML 推理步骤作为其图形工作一部分的着色器。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv id=\"faq3\" class=\"border-t border-gray-200 py-4 first:border-t-0 dark:border-gray-700\">\u003Ch4 class=\"font-semibold text-gray-900 dark:text-gray-100\">DX 线性代数已经是普通的零售版 DirectX 功能了吗？\u003C\u002Fh4>\u003Cdiv class=\"mt-2 text-gray-600 dark:text-gray-300\">截至 2026 年 9 月，它仍处于 Shader Model 6.10 \u002F Agility SDK 预览路径中。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv id=\"faq4\" class=\"border-t border-gray-200 py-4 first:border-t-0 dark:border-gray-700\">\u003Ch4 class=\"font-semibold text-gray-900 dark:text-gray-100\">什么是 DirectX 计算图编译器？\u003C\u002Fh4>\u003Cdiv class=\"mt-2 text-gray-600 dark:text-gray-300\">它是微软宣布的模型级 ML 编译器 API，旨在接收更大的计算图，并将其降低为与 D3D12 集成的优化 GPU 工作负载。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv id=\"faq5\" class=\"border-t border-gray-200 py-4 first:border-t-0 dark:border-gray-700\">\u003Ch4 class=\"font-semibold text-gray-900 dark:text-gray-100\">为什么不直接在 HLSL 中运行每个 ML 模型？\u003C\u002Fh4>\u003Cdiv class=\"mt-2 text-gray-600 dark:text-gray-300\">小型函数很适合着色器级编写，而大型模型则受益于全图优化、内存规划和算子融合。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv id=\"faq6\" class=\"border-t border-gray-200 py-4 first:border-t-0 dark:border-gray-700\">\u003Ch4 class=\"font-semibold text-gray-900 dark:text-gray-100\">这会取代 DLSS 或 FSR 吗？\u003C\u002Fh4>\u003Cdiv class=\"mt-2 text-gray-600 dark:text-gray-300\">不会直接取代。DirectX 提供的是更底层的基础设施，厂商和开发者可将其用于神经图形。品牌化技术仍然可以实现自己的模型和集成策略。\u003C\u002Fdiv>\u003C\u002Fdiv>\u003C\u002Fsection>\n\u003Ch2 id=\"section-78\">术语表\u003C\u002Fh2>\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 DirectX ML 术语\u003C\u002Fh3>\u003Cdl>\u003Cdiv id=\"dx-linear-algebra\" 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\">DX 线性代数\u003C\u002Fdt>\u003Cdd class=\"mt-1 text-gray-600 dark:text-gray-300\">用于加速向量和矩阵运算的 DirectX\u002FHLSL API，面向神经渲染、ML 和图像处理工作负载。\u003C\u002Fdd>\u003C\u002Fdiv>\u003Cdiv id=\"cooperative-vector\" 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\">协作向量\u003C\u002Fdt>\u003Cdd class=\"mt-1 text-gray-600 dark:text-gray-300\">一种较早的 DirectX 方法，用于在着色器内加速向量-矩阵运算，帮助确立了着色器级神经渲染路径。\u003C\u002Fdd>\u003C\u002Fdiv>\u003Cdiv id=\"shader-model-6-10\" 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\">着色器模型 6.10\u003C\u002Fdt>\u003Cdd class=\"mt-1 text-gray-600 dark:text-gray-300\">预览版着色器模型代际，包含当前的 DirectX 线性代数矩阵 API。\u003C\u002Fdd>\u003C\u002Fdiv>\u003Cdiv id=\"compute-graph-compiler\" 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\">DirectX 计算图编译器\u003C\u002Fdt>\u003Cdd class=\"mt-1 text-gray-600 dark:text-gray-300\">微软宣布的编译器 API，用于将更大的 ML 计算图优化并执行为原生 DirectX GPU 工作负载。\u003C\u002Fdd>\u003C\u002Fdiv>\u003Cdiv id=\"neural-shader\" 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\">神经着色器\u003C\u002Fdt>\u003Cdd class=\"mt-1 text-gray-600 dark:text-gray-300\">一个实用术语，指着色器在图形处理过程中执行学习到的推理或神经计算。\u003C\u002Fdd>\u003C\u002Fdiv>\u003Cdiv id=\"shader-or-graph-boundary\" 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\">着色器或图边界\u003C\u002Fdt>\u003Cdd class=\"mt-1 text-gray-600 dark:text-gray-300\">Figure Rocks 的一种模型，用于判断 ML 工作负载应作为小型内联着色器数学，还是作为更大的模型级计算图。\u003C\u002Fdd>\u003C\u002Fdiv>\u003Cdiv id=\"neural-shader-value-test\" 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\">神经着色器价值测试\u003C\u002Fdt>\u003Cdd class=\"mt-1 text-gray-600 dark:text-gray-300\">Figure Rocks 的一种工作流程，用于判断神经技术所避免的传统计算或内存成本是否值得其推理成本和画质权衡。\u003C\u002Fdd>\u003C\u002Fdiv>\u003C\u002Fdl>\u003C\u002Fsection>\n\u003Ch2 id=\"section-80\">主要来源\u003C\u002Fh2>\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 Windows 上的 ML 时代演进 DirectX\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">GDC 2026 官方架构概述，涵盖着色器级 ML、DX 线性代数和 DirectX 计算图编译器。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fdevblogs.microsoft.com\u002Fdirectx\u002Fd3d12-linalg-preview\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 — D3D12 LinAlg 矩阵预览\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">2026 年 4 月官方预览，解释统一的线性代数 API、矩阵运算和神经渲染动机。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fdevblogs.microsoft.com\u002Fdirectx\u002Fannouncing-agilitysdk-721-preview-and-more-shader-model-6-10-features\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 — Agility SDK 1.721 预览\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">2026 年 5 月官方发布，记录 Shader Model 6.10 线性代数更新和预览硬件支持。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fdeveloper.microsoft.com\u002Fen-us\u002Fgames\u002Farticles\u002F2026\u002F03\u002Fgdc-2026-evolving-directx-for-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 Game Dev — GDC 2026：为 ML 时代演进 DirectX\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">微软游戏开发官方摘要，解释着色器级和模型级 ML 以及计算图编译器的作用。\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fdevblogs.microsoft.com\u002Fdirectx\u002Fcooperative-vector\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 — D3D12 协作向量\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">关于直接从着色器线程进行硬件加速向量\u002F矩阵运算和神经渲染的官方背景资料。\u003C\u002Fp>\u003C\u002Fa>",{"time":538,"blocks":539,"version":1177},1790378238937,[540,546,554,561,568,574,579,584,589,594,614,621,626,631,636,641,668,673,678,683,688,693,698,722,727,732,737,742,751,756,761,766,771,776,809,814,819,824,829,835,840,845,850,855,860,865,870,875,883,888,909,914,919,924,929,934,939,944,949,954,978,983,988,993,998,1003,1008,1013,1018,1023,1028,1033,1038,1043,1048,1053,1058,1088,1093,1126,1131,1141,1150,1159,1168],{"id":541,"data":542,"type":544,"tunes":545},"intro",{"text":543},"DirectX 不再只是向 GPU 发送传统着色器的图形 API。微软正在将机器学习原语直接添加到 HLSL 中，并为在 DirectX 生态系统中运行更大的 ML 图开辟了第二条路径。这改变了神经渲染在未来 PC 游戏中的位置。","paragraph",{},{"id":547,"data":548,"type":552,"tunes":553},"direct",{"body":549,"title":550,"variant":551},"\u003Cstrong>DirectX 正在成为一个具备 ML 能力的图形平台。\u003C\u002Fstrong>小型神经工作负载可以通过 DX Linear Algebra 直接在着色器内运行，而更大的模型则由 DirectX Compute Graph Compiler 来处理。目标是让游戏引擎能够使用 GPU AI 硬件，而无需为每种技术构建单独的厂商特定路径。","直接回答","info","callout",{},{"id":555,"data":556,"type":552,"tunes":560},"status-note",{"body":557,"title":558,"variant":559},"截至 2026 年 9 月，\u003Cstrong>DX Linear Algebra 仍是一项预览技术\u003C\u002Fstrong>，位于 Shader Model 6.10 \u002F Agility SDK 预览路径中。微软宣布 DirectX Compute Graph Compiler 为私人预览，而非广泛发布的零售功能。本文解释的是架构和方向，并不声称当前每款游戏今天都能使用它。","当前状态","note",{},{"id":562,"data":563,"type":566,"tunes":567},"toc",{"title":564,"maxLevel":565,"minLevel":47},"目录",3,"tableOfContents",{},{"id":569,"data":570,"type":572,"tunes":573},"h-why",{"text":571,"level":47},"为什么神经渲染需要新的 DirectX 原语","header",{},{"id":575,"data":576,"type":544,"tunes":578},"p-why-1",{"text":577},"现代 GPU 已经包含用于机器学习矩阵运算的专用硬件。对于游戏开发者来说，问题不仅在于该硬件是否存在，还在于如何从实时图形管线中高效地访问它。",{},{"id":580,"data":581,"type":544,"tunes":583},"p-why-2",{"text":582},"微软最初通过 Cooperative Vector 支持对此进行了探索。2026 年，这项工作演变为更广泛的 DirectX Linear Algebra 设计，同时支持向量-矩阵和矩阵-矩阵运算。",{},{"id":585,"data":586,"type":544,"tunes":588},"p-why-3",{"text":587},"这很重要，因为不同的神经渲染任务具有不同的形态。逐像素评估材质行为的小型模型与大型超分辨率或去噪图的工作负载并不相同。",{},{"id":590,"data":591,"type":572,"tunes":593},"h-two-level",{"text":592,"level":47},"两级 DirectX ML 模型",{},{"id":595,"data":596,"type":612,"tunes":613},"two-level-flow",{"steps":597,"title":610,"orientation":611},[598,601,604,607],{"label":599,"description":600},"1. 着色器级 ML","小型神经或线性代数工作负载直接从 HLSL 中与传统着色器代码一起执行。",{"label":602,"description":603},"2. DX Linear Algebra","着色器可以请求硬件加速的向量和矩阵运算，而无需手动实现每个 ML 原语。",{"label":605,"description":606},"3. 模型级 ML","更大的神经网络被表示为完整的计算图，而不是手写的着色器片段。",{"label":608,"description":609},"4. DirectX Compute Graph Compiler","微软计划的编译器路径会分析这些图并将其降低为与 D3D12 队列和命令列表集成的优化 GPU 工作负载。","ML 进入 DirectX 图形管线的两种方式","auto","processFlow",{},{"id":615,"data":616,"type":552,"tunes":620},"simple-diff",{"body":617,"title":618,"variant":619},"\u003Cstrong>DX Linear Algebra：\u003C\u002Fstrong>将小型 ML 数学放入着色器内。\u003Cbr>\u003Cstrong>Compute Graph Compiler：\u003C\u002Fstrong>将更大的 ML 模型作为图引入引擎。","简单区别","success",{},{"id":622,"data":623,"type":572,"tunes":625},"h-linalg",{"text":624,"level":47},"DX Linear Algebra 实际上为着色器提供了什么",{},{"id":627,"data":628,"type":544,"tunes":630},"p-linalg-1",{"text":629},"传统 HLSL 围绕图形和计算操作构建。神经工作负载严重依赖线性代数：向量、矩阵、乘法、累加以及针对这些操作优化的数据布局。",{},{"id":632,"data":633,"type":544,"tunes":635},"p-linalg-2",{"text":634},"Shader Model 6.10 预览版添加了一流的矩阵 API，使开发者能够更直接地表达这些工作负载，并让驱动程序将它们映射到专用硬件。",{},{"id":637,"data":638,"type":544,"tunes":640},"p-linalg-3",{"text":639},"微软 2026 年 4 月的预览版明确将其描述为神经渲染、ML 和图像处理工作负载的统一路径，而非仅限图形的功能。",{},{"id":642,"data":643,"type":666,"tunes":667},"math-table",{"rows":644,"title":657,"layout":658,"columns":659},[645,649,653],{"id":646,"label":647,"values":648},"work","典型工作",[13,13],{"id":650,"label":651,"values":652},"hardware","硬件路径",[13,13],{"id":654,"label":655,"values":656},"code","开发者表达",[13,13],"传统着色器数学与面向 ML 的着色器数学","table",[660,663],{"id":661,"label":662},"traditional","传统着色器重点",{"id":664,"label":665},"ml","面向 ML 的重点","comparison",{},{"id":669,"data":670,"type":572,"tunes":672},"h-coop",{"text":671,"level":47},"为什么 Cooperative Vector 只是开始",{},{"id":674,"data":675,"type":544,"tunes":677},"p-coop-1",{"text":676},"Cooperative Vector 允许着色器线程请求向量-矩阵工作，驱动程序可以将其映射到专用硬件。这对于高度并行的逐像素工作负载非常有用。",{},{"id":679,"data":680,"type":544,"tunes":682},"p-coop-2",{"text":681},"微软后来得出结论，许多重要的机器学习工作负载需要的不仅仅是向量-矩阵运算。超分辨率、去噪、时间重建、更大的图像模型和通用推理可能需要矩阵-矩阵运算以及跨多个线程的共享工作。",{},{"id":684,"data":685,"type":544,"tunes":687},"p-coop-3",{"text":686},"因此，DX 线性代数扩展了模型，而不是将协作向量视为最终抽象。",{},{"id":689,"data":690,"type":572,"tunes":692},"h-usecases",{"text":691,"level":47},"着色器内部实际可以运行什么？",{},{"id":694,"data":695,"type":544,"tunes":697},"p-use-1",{"text":696},"最有趣的着色器级机器学习工作负载足够小，可以在其操作的图形数据附近执行。",{},{"id":699,"data":700,"type":658,"tunes":721},"usecases-table",{"content":701,"stretched":720,"withHeadings":15},[702,705,708,711,714,717],[703,704],"工作负载","为什么着色器级机器学习适合",[706,707],"神经纹理压缩","小型网络可以在着色器需要纹理信息的点附近重建纹理信息",[709,710],"神经材质评估","学习函数可以替代或增强昂贵的手工编写材质数学",[712,713],"神经辐射缓存","逐像素或局部推理可以从学习到的场景行为中估计光照信息",[715,716],"小型去噪\u002F重建内核","机器学习操作可以直接位于其改进的渲染阶段旁边",[718,719],"图像处理推理","矩阵运算可以嵌入 GPU 处理中，而无需单独的外部运行时",false,{},{"id":723,"data":724,"type":572,"tunes":726},"h-texture",{"text":725,"level":47},"为什么这对纹理内存很重要",{},{"id":728,"data":729,"type":544,"tunes":731},"p-tex-1",{"text":730},"微软反复提到的一个例子是神经纹理压缩。",{},{"id":733,"data":734,"type":544,"tunes":736},"p-tex-2",{"text":735},"游戏可以存储更紧凑的表示，并在渲染期间使用小型神经网络重建细节，而不是以传统保真度存储每个纹理通道。",{},{"id":738,"data":739,"type":544,"tunes":741},"p-tex-3",{"text":740},"这用一些 GPU 推理工作换取了更低的存储或内存压力。具体收益取决于技术和硬件，但架构变化很重要：一些视觉细节可以变成计算而不是存储数据。",{},{"id":743,"data":744,"type":749,"tunes":750},"ref-vram",{"url":745,"title":746,"excerpt":747,"ctaLabel":748},"https:\u002F\u002Ffigure.rocks\u002Fzh\u002Fblog\u002Fvram-usage-is-not-vram-requirement-why-a-full-memory-meter-does-not-tell-the-whole-story","VRAM 使用量不是 VRAM 需求：为什么完整的内存计量器不能说明全部情况","关于 VRAM 容量、驻留预算、工作集以及为什么内存压力比一个使用量数字更复杂的实用指南。","阅读 VRAM 指南","referralArticle",{},{"id":752,"data":753,"type":572,"tunes":755},"h-models",{"text":754,"level":47},"为什么完整模型需要不同的路径",{},{"id":757,"data":758,"type":544,"tunes":760},"p-models-1",{"text":759},"在 HLSL 中手写一些矩阵运算对于小型神经函数是可行的。当工作负载是具有许多层、依赖关系和中间张量的完整现代模型时，这就变得非常不切实际。",{},{"id":762,"data":763,"type":544,"tunes":765},"p-models-2",{"text":764},"微软的 DirectX 计算图编译器就是为这类更大的工作负载设计的。",{},{"id":767,"data":768,"type":544,"tunes":770},"p-models-3",{"text":769},"编译器可以接受计算图、分析整个图、规划内存、融合操作，并将结果降低为与 DirectX 12 集成的 GPU 工作，而不是将模型重写为自定义着色器代码。",{},{"id":772,"data":773,"type":572,"tunes":775},"h-boundary",{"text":774,"level":47},"着色器或图的边界",{},{"id":777,"data":778,"type":666,"tunes":808},"boundary-table",{"rows":779,"title":800,"layout":658,"columns":801},[780,784,788,792,796],{"id":781,"label":782,"values":783},"size","模型大小",[13,13],{"id":785,"label":786,"values":787},"location","执行风格",[13,13],{"id":789,"label":790,"values":791},"authoring","创作",[13,13],{"id":793,"label":794,"values":795},"optimization","优化范围",[13,13],{"id":797,"label":798,"values":799},"use","典型用途",[13,13],"机器学习工作负载何时属于 HLSL 与模型编译器",[802,805],{"id":803,"label":804},"shader","着色器级路径",{"id":806,"label":807},"graph","模型级路径",{},{"id":810,"data":811,"type":572,"tunes":813},"h-crossvendor",{"text":812,"level":47},"为什么跨供应商支持比另一个 AI 功能更重要",{},{"id":815,"data":816,"type":544,"tunes":818},"p-cross-1",{"text":817},"AMD、Intel、NVIDIA 和高通展示了不同的 GPU 架构和不同形式的专用矩阵加速。",{},{"id":820,"data":821,"type":544,"tunes":823},"p-cross-2",{"text":822},"DirectX 抽象层为微软和驱动程序供应商提供了一个平台，将通用的 HLSL 或图级别 ML 转换为正确的硬件路径。",{},{"id":825,"data":826,"type":544,"tunes":828},"p-cross-3",{"text":827},"这并不会让每个 GPU 都同样快，但它可以减少游戏引擎为每个供应商实现完全不同的神经渲染 API 的需求。",{},{"id":830,"data":831,"type":552,"tunes":834},"portability-note",{"body":832,"title":833,"variant":559},"有趣的部分不是 DirectX 可以“运行 AI”。GPU 本来就可以。平台的变化在于 \u003Cstrong>ML 可以通过通用的图形 API 和工具链来表达\u003C\u002Fstrong>。","可移植性才是真正的平台故事",{},{"id":836,"data":837,"type":572,"tunes":839},"h-support",{"text":838,"level":47},"当前预览版支持哪些硬件？",{},{"id":841,"data":842,"type":544,"tunes":844},"p-support-1",{"text":843},"答案取决于具体的线性代数操作和预览版驱动程序。",{},{"id":846,"data":847,"type":544,"tunes":849},"p-support-2",{"text":848},"微软的 Agility SDK 1.721 预览版支持表列出了 LinAlg VectorAccumulate 对 AMD Radeon RX 9000 系列硬件、通过即将推出的驱动程序支持的 Intel Xe2 或更新硬件，以及通过支持的预览路径的 NVIDIA RTX 硬件的支持。",{},{"id":851,"data":852,"type":544,"tunes":854},"p-support-3",{"text":853},"这是预览时代的支持，不是通用的零售保证。硬件和驱动程序支持在最终确定之前可能会发生变化。",{},{"id":856,"data":857,"type":572,"tunes":859},"h-infra",{"text":858,"level":47},"神经渲染正在成为基础设施",{},{"id":861,"data":862,"type":544,"tunes":864},"p-infra-1",{"text":863},"DLSS、FSR 和其他神经图形技术通常被讨论为游戏设置菜单中可见的品牌功能。",{},{"id":866,"data":867,"type":544,"tunes":869},"p-infra-2",{"text":868},"DirectX 线性代数指向更深层次的变化。神经操作可以成为渲染器内部的实现细节，而不是单一的可选后处理功能。",{},{"id":871,"data":872,"type":544,"tunes":874},"p-infra-3",{"text":873},"开发者可以将 ML 用于纹理、材质、光照、重建或其他局部功能，而无需将每个功能都暴露为面向消费者的 AI 开关。",{},{"id":876,"data":877,"type":749,"tunes":882},"ref-dlss5",{"url":878,"title":879,"excerpt":880,"ctaLabel":881},"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：3D 引导的神经渲染实际改变了什么","神经渲染如何超越超分辨率和生成帧，进入图形管线内的材质和光照重建。","阅读 DLSS 5 神经渲染指南",{},{"id":884,"data":885,"type":572,"tunes":887},"h-stack",{"text":886,"level":47},"神经渲染堆栈正在变得分层",{},{"id":889,"data":890,"type":612,"tunes":908},"stack-flow",{"steps":891,"title":907,"orientation":611},[892,895,898,901,904],{"label":893,"description":894},"传统引擎工作","模拟、几何、可见性和基础渲染仍然是标准引擎的职责。",{"label":896,"description":897},"内联神经着色器","小型学习函数在 HLSL 内重建纹理、材质或光照。",{"label":899,"description":900},"更大的 ML 图","完整的重建或推理模型通过图级别的 DirectX 路径执行。",{"label":902,"description":903},"供应商硬件映射","驱动程序将常见的 DirectX 操作映射到 GPU 的专用 AI 和矩阵单元上。",{"label":905,"description":906},"PIX 可见性","图形和 ML 工作可以一起进行性能分析，而不是存在于单独的不透明运行时中。","未来可能的 DirectX 游戏管线",{},{"id":910,"data":911,"type":572,"tunes":913},"h-pix",{"text":912,"level":47},"为什么统一性能分析很重要",{},{"id":915,"data":916,"type":544,"tunes":918},"p-pix-1",{"text":917},"一个改善图像质量的神经工作负载如果意外消耗帧时间、内存带宽或 VRAM，仍然可能损害游戏。",{},{"id":920,"data":921,"type":544,"tunes":923},"p-pix-2",{"text":922},"微软明确将统一的 PIX 可见性作为 Compute Graph Compiler 方向的一部分。这很重要，因为开发者需要在同一帧捕获中看到图形和 ML 工作。",{},{"id":925,"data":926,"type":544,"tunes":928},"p-pix-3",{"text":927},"如果神经渲染成为基础设施，它必须像任何其他渲染阶段一样可测量。",{},{"id":930,"data":931,"type":572,"tunes":933},"h-not-everything",{"text":932,"level":47},"这并不意味着每个着色器都会变成神经网络",{},{"id":935,"data":936,"type":544,"tunes":938},"p-not-1",{"text":937},"对于许多工作负载，传统着色器数学仍然高效、确定性强且易于推理。",{},{"id":940,"data":941,"type":544,"tunes":943},"p-not-2",{"text":942},"当神经函数能够更高效地近似或重建昂贵的东西、压缩数据，或者产生原本需要过多传统计算或内存才能达到的质量时，它才有意义。",{},{"id":945,"data":946,"type":544,"tunes":948},"p-not-3",{"text":947},"正确的架构将保持混合模式。",{},{"id":950,"data":951,"type":572,"tunes":953},"h-test",{"text":952,"level":47},"神经着色器价值测试",{},{"id":955,"data":956,"type":612,"tunes":977},"value-test",{"steps":957,"title":976,"orientation":611},[958,961,964,967,970,973],{"label":959,"description":960},"1. 识别昂贵的传统操作","你试图替代的是什么计算、带宽、内存或存储成本？",{"label":962,"description":963},"2. 定义神经替代方案","一个小模型可以重建、预测或压缩什么？",{"label":965,"description":966},"3. 测量推理成本","机器学习模型本身会消耗 GPU 时间、内存和带宽。",{"label":968,"description":969},"4. 测量质量稳定性","寻找时间伪影、重建误差和失败案例。",{"label":971,"description":972},"5. 测量净节省","只有当避免的传统成本值得增加的机器学习成本时，该技术才有用。",{"label":974,"description":975},"6. 跨供应商测试","DirectX 抽象有助于可移植性，但实际硬件性能仍然存在差异。","何时将机器学习放入渲染管线才真正有意义？",{},{"id":979,"data":980,"type":572,"tunes":982},"h-gamers",{"text":981,"level":47},"这对游戏玩家意味着什么",{},{"id":984,"data":985,"type":544,"tunes":987},"p-gamers-1",{"text":986},"游戏玩家可能永远不会在图形菜单中看到“DX 线性代数”开关。",{},{"id":989,"data":990,"type":544,"tunes":992},"p-gamers-2",{"text":991},"可能的影响是间接的：更小的纹理占用、更好的光照重建、更高效的神经图形，或者因为引擎可以通过标准路径访问矩阵加速而变得实用的新视觉技术。",{},{"id":994,"data":995,"type":544,"tunes":997},"p-gamers-3",{"text":996},"该功能作为基础设施比作为品牌更重要。",{},{"id":999,"data":1000,"type":572,"tunes":1002},"h-change",{"text":1001,"level":47},"什么会改变这个答案？",{},{"id":1004,"data":1005,"type":544,"tunes":1007},"p-change-1",{"text":1006},"最大的不确定性是这些 API 的最终零售形态。DX 线性代数仍处于预览阶段，而计算图编译器尚未达到广泛的零售可用性。",{},{"id":1009,"data":1010,"type":544,"tunes":1012},"p-change-2",{"text":1011},"在这些系统成为正常的发行游戏基础设施之前，最终的硬件支持、API 细节、编译器行为和引擎采用可能会发生变化。",{},{"id":1014,"data":1015,"type":544,"tunes":1017},"p-change-3",{"text":1016},"如果主要引擎直接采用这些抽象，神经着色器可能会变得更加普遍，而无需各个游戏团队从头实现每一项技术。",{},{"id":1019,"data":1020,"type":572,"tunes":1022},"h-limit",{"text":1021,"level":47},"局限性",{},{"id":1024,"data":1025,"type":544,"tunes":1027},"p-limit-1",{"text":1026},"本文解释了微软记录的 DirectX 架构和预览 API。它并不声称 DX 线性代数目前能提高每款游戏的性能，也不声称计算图编译器是成品零售产品。",{},{"id":1029,"data":1030,"type":544,"tunes":1032},"p-limit-2",{"text":1031},"微软的示例描述了能力和预期用途。实际收益取决于模型、引擎集成、GPU 架构、驱动程序和负载。",{},{"id":1034,"data":1035,"type":572,"tunes":1037},"h-conclusion",{"text":1036,"level":47},"结论",{},{"id":1039,"data":1040,"type":544,"tunes":1042},"p-conc-1",{"text":1041},"DirectX 的重要变化并不是又一个名为 AI 的复选框。",{},{"id":1044,"data":1045,"type":544,"tunes":1047},"p-conc-2",{"text":1046},"而是机器学习数学正在进入图形编程模型本身。小型神经函数可以驻留在着色器内，更大的模型可以走向图级编译，而同一套 DirectX 工具链可以在多个 GPU 厂商之间暴露这些工作负载。",{},{"id":1049,"data":1050,"type":544,"tunes":1052},"p-conc-3",{"text":1051},"这就是神经渲染如何不再只是一个品牌化功能，而开始成为渲染基础设施的一部分。",{},{"id":1054,"data":1055,"type":572,"tunes":1057},"h-faq",{"text":1056,"level":47},"常见问题",{},{"id":1059,"data":1060,"type":1059,"tunes":1087},"faq",{"items":1061,"title":1086},[1062,1066,1070,1074,1078,1082],{"id":1063,"answer":1064,"question":1065},"faq1","它是一组 DirectX\u002FHLSL 功能，用于机器学习、神经渲染和图像处理工作负载所使用的硬件加速向量和矩阵运算。","什么是 DirectX 线性代数？",{"id":1067,"answer":1068,"question":1069},"faq2","一个有用的通俗描述是：一种将学习到的神经函数或 ML 推理步骤作为其图形工作一部分的着色器。","什么是神经着色器？",{"id":1071,"answer":1072,"question":1073},"faq3","截至 2026 年 9 月，它仍处于 Shader Model 6.10 \u002F Agility SDK 预览路径中。","DX 线性代数已经是普通的零售版 DirectX 功能了吗？",{"id":1075,"answer":1076,"question":1077},"faq4","它是微软宣布的模型级 ML 编译器 API，旨在接收更大的计算图，并将其降低为与 D3D12 集成的优化 GPU 工作负载。","什么是 DirectX 计算图编译器？",{"id":1079,"answer":1080,"question":1081},"faq5","小型函数很适合着色器级编写，而大型模型则受益于全图优化、内存规划和算子融合。","为什么不直接在 HLSL 中运行每个 ML 模型？",{"id":1083,"answer":1084,"question":1085},"faq6","不会直接取代。DirectX 提供的是更底层的基础设施，厂商和开发者可将其用于神经图形。品牌化技术仍然可以实现自己的模型和集成策略。","这会取代 DLSS 或 FSR 吗？","DirectX 线性代数与神经着色器",{},{"id":1089,"data":1090,"type":572,"tunes":1092},"h-glossary",{"text":1091,"level":47},"术语表",{},{"id":1094,"data":1095,"type":1094,"tunes":1125},"glossary",{"title":1096,"entries":1097},"关键 DirectX ML 术语",[1098,1102,1106,1110,1114,1118,1122],{"term":1099,"anchor":1100,"definition":1101},"DX 线性代数","dx-linear-algebra","用于加速向量和矩阵运算的 DirectX\u002FHLSL API，面向神经渲染、ML 和图像处理工作负载。",{"term":1103,"anchor":1104,"definition":1105},"协作向量","cooperative-vector","一种较早的 DirectX 方法，用于在着色器内加速向量-矩阵运算，帮助确立了着色器级神经渲染路径。",{"term":1107,"anchor":1108,"definition":1109},"着色器模型 6.10","shader-model-6-10","预览版着色器模型代际，包含当前的 DirectX 线性代数矩阵 API。",{"term":1111,"anchor":1112,"definition":1113},"DirectX 计算图编译器","compute-graph-compiler","微软宣布的编译器 API，用于将更大的 ML 计算图优化并执行为原生 DirectX GPU 工作负载。",{"term":1115,"anchor":1116,"definition":1117},"神经着色器","neural-shader","一个实用术语，指着色器在图形处理过程中执行学习到的推理或神经计算。",{"term":1119,"anchor":1120,"definition":1121},"着色器或图边界","shader-or-graph-boundary","Figure Rocks 的一种模型，用于判断 ML 工作负载应作为小型内联着色器数学，还是作为更大的模型级计算图。",{"term":952,"anchor":1123,"definition":1124},"neural-shader-value-test","Figure Rocks 的一种工作流程，用于判断神经技术所避免的传统计算或内存成本是否值得其推理成本和画质权衡。",{},{"id":1127,"data":1128,"type":572,"tunes":1130},"h-sources",{"text":1129,"level":47},"主要来源",{},{"id":1132,"data":1133,"type":1139,"tunes":1140},"src-ms-ml-era",{"link":1134,"meta":1135},"https:\u002F\u002Fdevblogs.microsoft.com\u002Fdirectx\u002Fevolving-directx-for-the-ml-era-on-windows\u002F",{"image":1136,"title":1137,"description":1138},{"url":13},"Microsoft DirectX — 为 Windows 上的 ML 时代演进 DirectX","GDC 2026 官方架构概述，涵盖着色器级 ML、DX 线性代数和 DirectX 计算图编译器。","linkTool",{},{"id":1142,"data":1143,"type":1139,"tunes":1149},"src-ms-linalg",{"link":1144,"meta":1145},"https:\u002F\u002Fdevblogs.microsoft.com\u002Fdirectx\u002Fd3d12-linalg-preview\u002F",{"image":1146,"title":1147,"description":1148},{"url":13},"Microsoft DirectX — D3D12 LinAlg 矩阵预览","2026 年 4 月官方预览，解释统一的线性代数 API、矩阵运算和神经渲染动机。",{},{"id":1151,"data":1152,"type":1139,"tunes":1158},"src-ms-agility721",{"link":1153,"meta":1154},"https:\u002F\u002Fdevblogs.microsoft.com\u002Fdirectx\u002Fannouncing-agilitysdk-721-preview-and-more-shader-model-6-10-features\u002F",{"image":1155,"title":1156,"description":1157},{"url":13},"Microsoft DirectX — Agility SDK 1.721 预览","2026 年 5 月官方发布，记录 Shader Model 6.10 线性代数更新和预览硬件支持。",{},{"id":1160,"data":1161,"type":1139,"tunes":1167},"src-ms-gdc",{"link":1162,"meta":1163},"https:\u002F\u002Fdeveloper.microsoft.com\u002Fen-us\u002Fgames\u002Farticles\u002F2026\u002F03\u002Fgdc-2026-evolving-directx-for-ml-era-on-windows\u002F",{"image":1164,"title":1165,"description":1166},{"url":13},"Microsoft Game Dev — GDC 2026：为 ML 时代演进 DirectX","微软游戏开发官方摘要，解释着色器级和模型级 ML 以及计算图编译器的作用。",{},{"id":1169,"data":1170,"type":1139,"tunes":1176},"src-ms-coop",{"link":1171,"meta":1172},"https:\u002F\u002Fdevblogs.microsoft.com\u002Fdirectx\u002Fcooperative-vector\u002F",{"image":1173,"title":1174,"description":1175},{"url":13},"Microsoft DirectX — D3D12 协作向量","关于直接从着色器线程进行硬件加速向量\u002F矩阵运算和神经渲染的官方背景资料。",{},"2.31","DirectX 正在超越传统的图形着色器。微软正在将硬件加速的线性代数直接添加到 HLSL 中，并为更大的机器学习模型提供一条独立路径，为神经纹理、学习材质、神经光照以及其他 AI 驱动的渲染技术奠定基础。","\u002Fuploads\u002F2026\u002F09\u002Fdirectx-is-becoming-an-ml-platform-what-linear-algebra-and-neural-shaders-mean-for-future-games-1790378019857-utigyk.webp","directx-is-becoming-an-ml-platform-what-linear-algebra-and-neural-shaders-mean-for-future-games-1790378019857-utigyk","PUBLISHED","2026-09-25T17:12:00.000Z","2026-09-25T23:12:37.728Z","2026-09-26T06:45:35.226Z",{"en":1186,"de":1187,"sr":1188,"es":1189,"fr":1190,"it":1191,"ru":1192,"zh":1193},"\u002Fblog\u002Fdirectx-is-becoming-an-ml-platform-what-linear-algebra-and-neural-shaders-mean-for-future-games","\u002Fde\u002Fblog\u002Fdirectx-is-becoming-an-ml-platform-what-linear-algebra-and-neural-shaders-mean-for-future-games","\u002Fsr\u002Fblog\u002Fdirectx-is-becoming-an-ml-platform-what-linear-algebra-and-neural-shaders-mean-for-future-games","\u002Fes\u002Fblog\u002Fdirectx-is-becoming-an-ml-platform-what-linear-algebra-and-neural-shaders-mean-for-future-games","\u002Ffr\u002Fblog\u002Fdirectx-is-becoming-an-ml-platform-what-linear-algebra-and-neural-shaders-mean-for-future-games","\u002Fit\u002Fblog\u002Fdirectx-is-becoming-an-ml-platform-what-linear-algebra-and-neural-shaders-mean-for-future-games","\u002Fru\u002Fblog\u002Fdirectx-is-becoming-an-ml-platform-what-linear-algebra-and-neural-shaders-mean-for-future-games","\u002Fzh\u002Fblog\u002Fdirectx-is-becoming-an-ml-platform-what-linear-algebra-and-neural-shaders-mean-for-future-games",[1195,1199,1203,1207,1211],{"id":1196,"name":1197,"slug":1198},251,"模糊与余辉","blur-and-persistence",{"id":1200,"name":1201,"slug":1202},252,"过载与拖影","overdrive-and-smearing",{"id":1204,"name":1205,"slug":1206},253,"刷新率与清晰度","refresh-and-clarity",{"id":1208,"name":1209,"slug":1210},430,"Animal Crossing","animal-crossing",{"id":1212,"name":1213,"slug":1214},220,"为实用而设计","design-that-serves-use",{"id":283,"login":1216,"email":1217,"displayName":1218},"aleksandar","aleksandar@stajic.de","Aleksandar Stajic",[1220,1734],{"lang":8,"title":1221,"content":1222,"contentJson":1223,"excerpt":1733},"DirectX Is Becoming an ML Platform: What Linear Algebra and Neural Shaders Mean for Future Games","{\"time\":1790405124627,\"blocks\":[{\"id\":\"intro\",\"type\":\"paragraph\",\"data\":{\"text\":\"DirectX is no longer only a graphics API that sends traditional shaders to the GPU. Microsoft is adding machine-learning primitives directly to HLSL and a second path for running larger ML graphs inside the DirectX ecosystem. That changes where neural rendering can live inside future PC games.\"},\"tunes\":{}},{\"id\":\"direct\",\"type\":\"callout\",\"data\":{\"variant\":\"info\",\"title\":\"Direct answer\",\"body\":\"\u003Cstrong>DirectX is becoming an ML-capable graphics platform.\u003C\u002Fstrong> Small neural workloads can run directly inside shaders through DX Linear Algebra, while larger models are being targeted by the DirectX Compute Graph Compiler. The goal is to let game engines use GPU AI hardware without building a separate vendor-specific path for every technique.\"},\"tunes\":{}},{\"id\":\"status-note\",\"type\":\"callout\",\"data\":{\"variant\":\"note\",\"title\":\"Current status\",\"body\":\"As of September 2026, \u003Cstrong>DX Linear Algebra is still a preview technology\u003C\u002Fstrong> in the Shader Model 6.10 \u002F Agility SDK preview path. Microsoft announced the DirectX Compute Graph Compiler for private preview rather than as a broadly shipping retail feature. This article explains the architecture and direction, not a claim that every current game can use it today.\"},\"tunes\":{}},{\"id\":\"toc\",\"type\":\"tableOfContents\",\"data\":{\"title\":\"Contents\",\"minLevel\":2,\"maxLevel\":3},\"tunes\":{}},{\"id\":\"h-why\",\"type\":\"header\",\"data\":{\"text\":\"Why neural rendering needs new DirectX primitives\",\"level\":2},\"tunes\":{}},{\"id\":\"p-why-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"Modern GPUs already contain specialized hardware for matrix operations used by machine learning. The problem for a game developer is not only whether that hardware exists, but how to access it efficiently from a real-time graphics pipeline.\"},\"tunes\":{}},{\"id\":\"p-why-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"Microsoft first explored this with Cooperative Vector support. In 2026, that work evolved into a broader DirectX Linear Algebra design that supports both vector-matrix and matrix-matrix operations.\"},\"tunes\":{}},{\"id\":\"p-why-3\",\"type\":\"paragraph\",\"data\":{\"text\":\"That matters because different neural-rendering jobs have different shapes. A tiny model evaluating material behavior per pixel is not the same workload as a large super-resolution or denoising graph.\"},\"tunes\":{}},{\"id\":\"h-two-level\",\"type\":\"header\",\"data\":{\"text\":\"The Two-Level DirectX ML Model\",\"level\":2},\"tunes\":{}},{\"id\":\"two-level-flow\",\"type\":\"processFlow\",\"data\":{\"title\":\"Two ways ML can enter the DirectX graphics pipeline\",\"orientation\":\"auto\",\"steps\":[{\"label\":\"1. Shader-level ML\",\"description\":\"Small neural or linear-algebra workloads execute directly from HLSL alongside traditional shader code.\"},{\"label\":\"2. DX Linear Algebra\",\"description\":\"The shader can request hardware-accelerated vector and matrix operations instead of manually implementing every ML primitive.\"},{\"label\":\"3. Model-level ML\",\"description\":\"Larger neural networks are represented as complete computation graphs rather than hand-written shader fragments.\"},{\"label\":\"4. DirectX Compute Graph Compiler\",\"description\":\"Microsoft's planned compiler path analyzes and lowers those graphs into optimized GPU workloads integrated with D3D12 queues and command lists.\"}]},\"tunes\":{}},{\"id\":\"simple-diff\",\"type\":\"callout\",\"data\":{\"variant\":\"success\",\"title\":\"The simple difference\",\"body\":\"\u003Cstrong>DX Linear Algebra:\u003C\u002Fstrong> put small ML math inside the shader.\u003Cbr>\u003Cstrong>Compute Graph Compiler:\u003C\u002Fstrong> bring a larger ML model into the engine as a graph.\"},\"tunes\":{}},{\"id\":\"h-linalg\",\"type\":\"header\",\"data\":{\"text\":\"What DX Linear Algebra actually gives a shader\",\"level\":2},\"tunes\":{}},{\"id\":\"p-linalg-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"Traditional HLSL is built around graphics and compute operations. Neural workloads rely heavily on linear algebra: vectors, matrices, multiplication, accumulation and data layouts optimized for those operations.\"},\"tunes\":{}},{\"id\":\"p-linalg-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"The Shader Model 6.10 preview adds first-class matrix APIs so developers can express those workloads more directly and let the driver map them to specialized hardware.\"},\"tunes\":{}},{\"id\":\"p-linalg-3\",\"type\":\"paragraph\",\"data\":{\"text\":\"Microsoft's April 2026 preview explicitly describes this as a unified path for neural rendering, ML and image-processing workloads rather than a graphics-only feature.\"},\"tunes\":{}},{\"id\":\"math-table\",\"type\":\"comparison\",\"data\":{\"title\":\"Traditional shader math vs ML-oriented shader math\",\"layout\":\"table\",\"columns\":[{\"id\":\"traditional\",\"label\":\"Traditional shader focus\"},{\"id\":\"ml\",\"label\":\"ML-oriented focus\"}],\"rows\":[{\"id\":\"work\",\"label\":\"Typical work\",\"values\":[\"\",\"\"]},{\"id\":\"hardware\",\"label\":\"Hardware path\",\"values\":[\"\",\"\"]},{\"id\":\"code\",\"label\":\"Developer expression\",\"values\":[\"\",\"\"]}]},\"tunes\":{}},{\"id\":\"h-coop\",\"type\":\"header\",\"data\":{\"text\":\"Why Cooperative Vector was only the beginning\",\"level\":2},\"tunes\":{}},{\"id\":\"p-coop-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"Cooperative Vector allowed shader threads to request vector-matrix work that drivers could map to specialized hardware. That was useful for highly parallel per-pixel workloads.\"},\"tunes\":{}},{\"id\":\"p-coop-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"Microsoft later concluded that many important ML workloads need more than vector-matrix operations. Super resolution, denoising, temporal reconstruction, larger image models and general inference can require matrix-matrix operations and shared work across many threads.\"},\"tunes\":{}},{\"id\":\"p-coop-3\",\"type\":\"paragraph\",\"data\":{\"text\":\"DX Linear Algebra therefore broadens the model instead of treating Cooperative Vector as the final abstraction.\"},\"tunes\":{}},{\"id\":\"h-usecases\",\"type\":\"header\",\"data\":{\"text\":\"What can actually run inside a shader?\",\"level\":2},\"tunes\":{}},{\"id\":\"p-use-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"The most interesting shader-level ML workloads are small enough to execute close to the graphics data they operate on.\"},\"tunes\":{}},{\"id\":\"usecases-table\",\"type\":\"table\",\"data\":{\"withHeadings\":true,\"stretched\":false,\"content\":[[\"Workload\",\"Why shader-level ML fits\"],[\"Neural texture compression\",\"A small network can reconstruct texture information near the point where the shader needs it\"],[\"Neural material evaluation\",\"A learned function can replace or augment expensive hand-authored material math\"],[\"Neural radiance caching\",\"Per-pixel or local inference can estimate lighting information from learned scene behavior\"],[\"Small denoising\u002Freconstruction kernels\",\"ML operations can sit directly beside the rendering stage they improve\"],[\"Image-processing inference\",\"Matrix operations can be embedded into GPU processing without a separate external runtime\"]]},\"tunes\":{}},{\"id\":\"h-texture\",\"type\":\"header\",\"data\":{\"text\":\"Why this matters for texture memory\",\"level\":2},\"tunes\":{}},{\"id\":\"p-tex-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"One of Microsoft's recurring examples is neural texture compression.\"},\"tunes\":{}},{\"id\":\"p-tex-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"Instead of storing every texture channel at conventional fidelity, a game can store a more compact representation and use a small neural network to reconstruct detail during rendering.\"},\"tunes\":{}},{\"id\":\"p-tex-3\",\"type\":\"paragraph\",\"data\":{\"text\":\"That trades some GPU inference work for lower storage or memory pressure. The exact benefit depends on the technique and hardware, but the architectural change is important: some visual detail can become computation instead of stored data.\"},\"tunes\":{}},{\"id\":\"ref-vram\",\"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 guide to VRAM capacity, residency budgets, working sets and why memory pressure is more complicated than one usage number.\",\"ctaLabel\":\"Read the VRAM guide\"},\"tunes\":{}},{\"id\":\"h-models\",\"type\":\"header\",\"data\":{\"text\":\"Why full models need a different path\",\"level\":2},\"tunes\":{}},{\"id\":\"p-models-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"Hand-writing a few matrix operations in HLSL is practical for small neural functions. It becomes much less practical when the workload is a complete modern model with many layers, dependencies and intermediate tensors.\"},\"tunes\":{}},{\"id\":\"p-models-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"Microsoft's DirectX Compute Graph Compiler is designed for this larger class of workload.\"},\"tunes\":{}},{\"id\":\"p-models-3\",\"type\":\"paragraph\",\"data\":{\"text\":\"Instead of rewriting the model as custom shader code, the compiler can accept a computation graph, analyze the whole graph, plan memory, fuse operations and lower the result into GPU work that integrates with DirectX 12.\"},\"tunes\":{}},{\"id\":\"h-boundary\",\"type\":\"header\",\"data\":{\"text\":\"The Shader-or-Graph Boundary\",\"level\":2},\"tunes\":{}},{\"id\":\"boundary-table\",\"type\":\"comparison\",\"data\":{\"title\":\"When the ML workload belongs in HLSL vs a model compiler\",\"layout\":\"table\",\"columns\":[{\"id\":\"shader\",\"label\":\"Shader-level path\"},{\"id\":\"graph\",\"label\":\"Model-level path\"}],\"rows\":[{\"id\":\"size\",\"label\":\"Model size\",\"values\":[\"\",\"\"]},{\"id\":\"location\",\"label\":\"Execution style\",\"values\":[\"\",\"\"]},{\"id\":\"authoring\",\"label\":\"Authoring\",\"values\":[\"\",\"\"]},{\"id\":\"optimization\",\"label\":\"Optimization scope\",\"values\":[\"\",\"\"]},{\"id\":\"use\",\"label\":\"Typical use\",\"values\":[\"\",\"\"]}]},\"tunes\":{}},{\"id\":\"h-crossvendor\",\"type\":\"header\",\"data\":{\"text\":\"Why cross-vendor support matters more than another AI feature\",\"level\":2},\"tunes\":{}},{\"id\":\"p-cross-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"AMD, Intel, NVIDIA and Qualcomm expose different GPU architectures and different forms of dedicated matrix acceleration.\"},\"tunes\":{}},{\"id\":\"p-cross-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"A DirectX abstraction gives Microsoft and the driver vendors a place to translate common HLSL or graph-level ML into the correct hardware path.\"},\"tunes\":{}},{\"id\":\"p-cross-3\",\"type\":\"paragraph\",\"data\":{\"text\":\"That does not make every GPU equally fast, but it can reduce the need for a game engine to implement a completely different neural-rendering API for every vendor.\"},\"tunes\":{}},{\"id\":\"portability-note\",\"type\":\"callout\",\"data\":{\"variant\":\"note\",\"title\":\"Portability is the real platform story\",\"body\":\"The interesting part is not that DirectX can “run AI.” GPUs already could. The platform change is that \u003Cstrong>ML becomes expressible through a common graphics API and toolchain\u003C\u002Fstrong>.\"},\"tunes\":{}},{\"id\":\"h-support\",\"type\":\"header\",\"data\":{\"text\":\"What hardware supports the current preview?\",\"level\":2},\"tunes\":{}},{\"id\":\"p-support-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"The answer depends on the specific Linear Algebra operation and preview driver.\"},\"tunes\":{}},{\"id\":\"p-support-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"Microsoft's Agility SDK 1.721 preview support table lists LinAlg VectorAccumulate for AMD Radeon RX 9000 Series hardware, Intel Xe2-or-newer hardware through an upcoming driver, and NVIDIA RTX hardware through the supported preview path.\"},\"tunes\":{}},{\"id\":\"p-support-3\",\"type\":\"paragraph\",\"data\":{\"text\":\"This is preview-era support, not a universal retail guarantee. Hardware and driver support can change before finalization.\"},\"tunes\":{}},{\"id\":\"h-infra\",\"type\":\"header\",\"data\":{\"text\":\"Neural rendering is becoming infrastructure\",\"level\":2},\"tunes\":{}},{\"id\":\"p-infra-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"DLSS, FSR and other neural graphics technologies are often discussed as branded features visible in a game's settings menu.\"},\"tunes\":{}},{\"id\":\"p-infra-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"DirectX Linear Algebra points to a deeper change. The neural operation can become an internal implementation detail inside the renderer rather than a single optional post-processing feature.\"},\"tunes\":{}},{\"id\":\"p-infra-3\",\"type\":\"paragraph\",\"data\":{\"text\":\"A developer could use ML for textures, materials, lighting, reconstruction or other local functions without exposing each one as a consumer-facing AI toggle.\"},\"tunes\":{}},{\"id\":\"ref-dlss5\",\"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 3D-Guided Neural Rendering Actually Changes\",\"excerpt\":\"How neural rendering is moving beyond upscaling and generated frames into material and lighting reconstruction inside the graphics pipeline.\",\"ctaLabel\":\"Read the DLSS 5 neural-rendering guide\"},\"tunes\":{}},{\"id\":\"h-stack\",\"type\":\"header\",\"data\":{\"text\":\"The Neural Rendering Stack is becoming layered\",\"level\":2},\"tunes\":{}},{\"id\":\"stack-flow\",\"type\":\"processFlow\",\"data\":{\"title\":\"A possible future DirectX game pipeline\",\"orientation\":\"auto\",\"steps\":[{\"label\":\"Traditional engine work\",\"description\":\"Simulation, geometry, visibility and base rendering remain standard engine responsibilities.\"},{\"label\":\"Inline neural shaders\",\"description\":\"Small learned functions reconstruct textures, materials or lighting inside HLSL.\"},{\"label\":\"Larger ML graphs\",\"description\":\"Full reconstruction or inference models execute through a graph-level DirectX path.\"},{\"label\":\"Vendor hardware mapping\",\"description\":\"Drivers map common DirectX operations onto the GPU's specialized AI and matrix units.\"},{\"label\":\"PIX visibility\",\"description\":\"Graphics and ML work can be profiled together instead of living in separate opaque runtimes.\"}]},\"tunes\":{}},{\"id\":\"h-pix\",\"type\":\"header\",\"data\":{\"text\":\"Why unified profiling matters\",\"level\":2},\"tunes\":{}},{\"id\":\"p-pix-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"A neural workload that improves image quality can still damage a game if it unexpectedly consumes frame time, memory bandwidth or VRAM.\"},\"tunes\":{}},{\"id\":\"p-pix-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"Microsoft explicitly includes unified PIX visibility as part of the Compute Graph Compiler direction. That matters because developers need to see graphics and ML work in the same frame capture.\"},\"tunes\":{}},{\"id\":\"p-pix-3\",\"type\":\"paragraph\",\"data\":{\"text\":\"If neural rendering becomes infrastructure, it must be measurable like any other rendering stage.\"},\"tunes\":{}},{\"id\":\"h-not-everything\",\"type\":\"header\",\"data\":{\"text\":\"This does not mean every shader will become a neural network\",\"level\":2},\"tunes\":{}},{\"id\":\"p-not-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"Traditional shader math remains efficient, deterministic and easy to reason about for many workloads.\"},\"tunes\":{}},{\"id\":\"p-not-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"A neural function makes sense when it can approximate or reconstruct something expensive more efficiently, compress data, or produce quality that would otherwise require too much conventional compute or memory.\"},\"tunes\":{}},{\"id\":\"p-not-3\",\"type\":\"paragraph\",\"data\":{\"text\":\"The right architecture will remain hybrid.\"},\"tunes\":{}},{\"id\":\"h-test\",\"type\":\"header\",\"data\":{\"text\":\"The Neural Shader Value Test\",\"level\":2},\"tunes\":{}},{\"id\":\"value-test\",\"type\":\"processFlow\",\"data\":{\"title\":\"When does putting ML into the rendering pipeline actually make sense?\",\"orientation\":\"auto\",\"steps\":[{\"label\":\"1. Identify the expensive traditional operation\",\"description\":\"What compute, bandwidth, memory or storage cost are you trying to replace?\"},{\"label\":\"2. Define the neural substitute\",\"description\":\"What can a small model reconstruct, predict or compress?\"},{\"label\":\"3. Measure inference cost\",\"description\":\"The ML model itself consumes GPU time, memory and bandwidth.\"},{\"label\":\"4. Measure quality stability\",\"description\":\"Look for temporal artifacts, reconstruction errors and failure cases.\"},{\"label\":\"5. Measure net savings\",\"description\":\"The technique is useful only if the avoided traditional cost is worth the added ML cost.\"},{\"label\":\"6. Test across vendors\",\"description\":\"A DirectX abstraction helps portability, but real hardware performance still differs.\"}]},\"tunes\":{}},{\"id\":\"h-gamers\",\"type\":\"header\",\"data\":{\"text\":\"What this means for gamers\",\"level\":2},\"tunes\":{}},{\"id\":\"p-gamers-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"Gamers may never see a “DX Linear Algebra” switch in a graphics menu.\"},\"tunes\":{}},{\"id\":\"p-gamers-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"The likely impact is indirect: smaller texture footprints, better lighting reconstruction, more efficient neural graphics, or new visual techniques that become practical because the engine can access matrix acceleration through a standard path.\"},\"tunes\":{}},{\"id\":\"p-gamers-3\",\"type\":\"paragraph\",\"data\":{\"text\":\"The feature is more important as infrastructure than as a brand.\"},\"tunes\":{}},{\"id\":\"h-change\",\"type\":\"header\",\"data\":{\"text\":\"What would change this answer?\",\"level\":2},\"tunes\":{}},{\"id\":\"p-change-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"The biggest uncertainty is the final retail shape of these APIs. DX Linear Algebra remains in preview, and the Compute Graph Compiler has not yet reached broad retail availability.\"},\"tunes\":{}},{\"id\":\"p-change-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"Final hardware support, API details, compiler behavior and engine adoption may change before these systems become normal shipping-game infrastructure.\"},\"tunes\":{}},{\"id\":\"p-change-3\",\"type\":\"paragraph\",\"data\":{\"text\":\"If major engines adopt the abstractions directly, neural shaders could become much more common without individual game teams implementing every technique from scratch.\"},\"tunes\":{}},{\"id\":\"h-limit\",\"type\":\"header\",\"data\":{\"text\":\"Limitations\",\"level\":2},\"tunes\":{}},{\"id\":\"p-limit-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"This article explains Microsoft's documented DirectX architecture and preview APIs. It does not claim that DX Linear Algebra currently improves performance in every game or that the Compute Graph Compiler is a finished retail product.\"},\"tunes\":{}},{\"id\":\"p-limit-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"Microsoft's examples describe capability and intended use. Real benefits depend on the model, engine integration, GPU architecture, drivers and workload.\"},\"tunes\":{}},{\"id\":\"h-conclusion\",\"type\":\"header\",\"data\":{\"text\":\"Conclusion\",\"level\":2},\"tunes\":{}},{\"id\":\"p-conc-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"The important DirectX change is not another checkbox called AI.\"},\"tunes\":{}},{\"id\":\"p-conc-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"It is that machine-learning math is moving into the graphics programming model itself. Small neural functions can sit inside shaders, larger models can move toward graph-level compilation, and the same DirectX toolchain can expose those workloads across multiple GPU vendors.\"},\"tunes\":{}},{\"id\":\"p-conc-3\",\"type\":\"paragraph\",\"data\":{\"text\":\"That is how neural rendering stops being one branded feature and starts becoming part of the rendering infrastructure.\"},\"tunes\":{}},{\"id\":\"h-faq\",\"type\":\"header\",\"data\":{\"text\":\"FAQ\",\"level\":2},\"tunes\":{}},{\"id\":\"faq\",\"type\":\"faq\",\"data\":{\"title\":\"DirectX Linear Algebra and neural shaders\",\"items\":[{\"id\":\"faq1\",\"question\":\"What is DirectX Linear Algebra?\",\"answer\":\"It is a DirectX\u002FHLSL feature set for hardware-accelerated vector and matrix operations used by machine learning, neural rendering and image-processing workloads.\"},{\"id\":\"faq2\",\"question\":\"What is a neural shader?\",\"answer\":\"A useful plain-English description is a shader that includes a learned neural function or ML inference step as part of its graphics work.\"},{\"id\":\"faq3\",\"question\":\"Is DX Linear Algebra already a normal retail DirectX feature?\",\"answer\":\"As of September 2026 it remains in the Shader Model 6.10 \u002F Agility SDK preview path.\"},{\"id\":\"faq4\",\"question\":\"What is the DirectX Compute Graph Compiler?\",\"answer\":\"It is Microsoft's announced model-level ML compiler API intended to take larger computation graphs and lower them into optimized GPU workloads integrated with D3D12.\"},{\"id\":\"faq5\",\"question\":\"Why not run every ML model directly in HLSL?\",\"answer\":\"Small functions fit shader-level authoring well, while large models benefit from whole-graph optimization, memory planning and operator fusion.\"},{\"id\":\"faq6\",\"question\":\"Will this replace DLSS or FSR?\",\"answer\":\"Not directly. DirectX provides lower-level infrastructure that vendors and developers can use for neural graphics. Branded technologies can still implement their own models and integration strategies.\"}]},\"tunes\":{}},{\"id\":\"h-glossary\",\"type\":\"header\",\"data\":{\"text\":\"Glossary\",\"level\":2},\"tunes\":{}},{\"id\":\"glossary\",\"type\":\"glossary\",\"data\":{\"title\":\"Key DirectX ML terms\",\"entries\":[{\"term\":\"DX Linear Algebra\",\"definition\":\"DirectX\u002FHLSL APIs for accelerated vector and matrix operations intended for neural rendering, ML and image-processing workloads.\",\"anchor\":\"dx-linear-algebra\"},{\"term\":\"Cooperative Vector\",\"definition\":\"An earlier DirectX approach for accelerated vector-matrix operations inside shaders that helped establish the shader-level neural-rendering path.\",\"anchor\":\"cooperative-vector\"},{\"term\":\"Shader Model 6.10\",\"definition\":\"The preview shader-model generation containing the current DirectX Linear Algebra matrix APIs.\",\"anchor\":\"shader-model-6-10\"},{\"term\":\"DirectX Compute Graph Compiler\",\"definition\":\"Microsoft's announced compiler API for optimizing and executing larger ML computation graphs as native DirectX GPU workloads.\",\"anchor\":\"compute-graph-compiler\"},{\"term\":\"Neural shader\",\"definition\":\"A practical term for a shader that performs learned inference or neural computation as part of graphics processing.\",\"anchor\":\"neural-shader\"},{\"term\":\"Shader-or-Graph Boundary\",\"definition\":\"A Figure Rocks model for deciding whether an ML workload belongs as small inline shader math or as a larger model-level computation graph.\",\"anchor\":\"shader-or-graph-boundary\"},{\"term\":\"Neural Shader Value Test\",\"definition\":\"A Figure Rocks workflow for deciding whether the traditional compute or memory cost avoided by a neural technique is worth its inference cost and quality trade-offs.\",\"anchor\":\"neural-shader-value-test\"}]},\"tunes\":{}},{\"id\":\"h-sources\",\"type\":\"header\",\"data\":{\"text\":\"Primary sources\",\"level\":2},\"tunes\":{}},{\"id\":\"src-ms-ml-era\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fdevblogs.microsoft.com\u002Fdirectx\u002Fevolving-directx-for-the-ml-era-on-windows\u002F\",\"meta\":{\"image\":{\"url\":\"\"},\"title\":\"Microsoft DirectX — Evolving DirectX for the ML Era on Windows\",\"description\":\"Official GDC 2026 architecture overview covering shader-level ML, DX Linear Algebra and the DirectX Compute Graph Compiler.\"}},\"tunes\":{}},{\"id\":\"src-ms-linalg\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fdevblogs.microsoft.com\u002Fdirectx\u002Fd3d12-linalg-preview\u002F\",\"meta\":{\"image\":{\"url\":\"\"},\"title\":\"Microsoft DirectX — D3D12 LinAlg Matrix Preview\",\"description\":\"Official April 2026 preview explaining the unified Linear Algebra APIs, matrix operations and neural-rendering motivation.\"}},\"tunes\":{}},{\"id\":\"src-ms-agility721\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fdevblogs.microsoft.com\u002Fdirectx\u002Fannouncing-agilitysdk-721-preview-and-more-shader-model-6-10-features\u002F\",\"meta\":{\"image\":{\"url\":\"\"},\"title\":\"Microsoft DirectX — Agility SDK 1.721 Preview\",\"description\":\"Official May 2026 release documenting Shader Model 6.10 Linear Algebra updates and preview hardware support.\"}},\"tunes\":{}},{\"id\":\"src-ms-gdc\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fdeveloper.microsoft.com\u002Fen-us\u002Fgames\u002Farticles\u002F2026\u002F03\u002Fgdc-2026-evolving-directx-for-ml-era-on-windows\u002F\",\"meta\":{\"image\":{\"url\":\"\"},\"title\":\"Microsoft Game Dev — GDC 2026: Evolving DirectX for the ML Era\",\"description\":\"Official Microsoft Game Dev summary explaining shader-level and model-level ML and the role of the Compute Graph Compiler.\"}},\"tunes\":{}},{\"id\":\"src-ms-coop\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fdevblogs.microsoft.com\u002Fdirectx\u002Fcooperative-vector\u002F\",\"meta\":{\"image\":{\"url\":\"\"},\"title\":\"Microsoft DirectX — D3D12 Cooperative Vector\",\"description\":\"Official background on hardware-accelerated vector\u002Fmatrix operations and neural rendering directly from shader threads.\"}},\"tunes\":{}}],\"version\":\"2.31.6\"}",{"time":1224,"blocks":1225,"version":1732},1790405124627,[1226,1230,1235,1240,1244,1248,1252,1256,1260,1264,1279,1284,1288,1292,1296,1300,1319,1323,1327,1331,1335,1339,1343,1365,1369,1373,1377,1381,1388,1392,1396,1400,1404,1408,1433,1437,1441,1445,1449,1454,1458,1462,1466,1470,1474,1478,1482,1486,1493,1497,1517,1521,1525,1529,1533,1537,1541,1545,1549,1553,1576,1580,1584,1588,1592,1596,1600,1604,1608,1612,1616,1620,1624,1628,1632,1636,1640,1663,1667,1693,1697,1704,1711,1718,1725],{"id":541,"data":1227,"type":544,"tunes":1229},{"text":1228},"DirectX is no longer only a graphics API that sends traditional shaders to the GPU. Microsoft is adding machine-learning primitives directly to HLSL and a second path for running larger ML graphs inside the DirectX ecosystem. That changes where neural rendering can live inside future PC games.",{},{"id":547,"data":1231,"type":552,"tunes":1234},{"body":1232,"title":1233,"variant":551},"\u003Cstrong>DirectX is becoming an ML-capable graphics platform.\u003C\u002Fstrong> Small neural workloads can run directly inside shaders through DX Linear Algebra, while larger models are being targeted by the DirectX Compute Graph Compiler. The goal is to let game engines use GPU AI hardware without building a separate vendor-specific path for every technique.","Direct answer",{},{"id":555,"data":1236,"type":552,"tunes":1239},{"body":1237,"title":1238,"variant":559},"As of September 2026, \u003Cstrong>DX Linear Algebra is still a preview technology\u003C\u002Fstrong> in the Shader Model 6.10 \u002F Agility SDK preview path. Microsoft announced the DirectX Compute Graph Compiler for private preview rather than as a broadly shipping retail feature. This article explains the architecture and direction, not a claim that every current game can use it today.","Current status",{},{"id":562,"data":1241,"type":566,"tunes":1243},{"title":1242,"maxLevel":565,"minLevel":47},"Contents",{},{"id":569,"data":1245,"type":572,"tunes":1247},{"text":1246,"level":47},"Why neural rendering needs new DirectX primitives",{},{"id":575,"data":1249,"type":544,"tunes":1251},{"text":1250},"Modern GPUs already contain specialized hardware for matrix operations used by machine learning. The problem for a game developer is not only whether that hardware exists, but how to access it efficiently from a real-time graphics pipeline.",{},{"id":580,"data":1253,"type":544,"tunes":1255},{"text":1254},"Microsoft first explored this with Cooperative Vector support. In 2026, that work evolved into a broader DirectX Linear Algebra design that supports both vector-matrix and matrix-matrix operations.",{},{"id":585,"data":1257,"type":544,"tunes":1259},{"text":1258},"That matters because different neural-rendering jobs have different shapes. A tiny model evaluating material behavior per pixel is not the same workload as a large super-resolution or denoising graph.",{},{"id":590,"data":1261,"type":572,"tunes":1263},{"text":1262,"level":47},"The Two-Level DirectX ML Model",{},{"id":595,"data":1265,"type":612,"tunes":1278},{"steps":1266,"title":1277,"orientation":611},[1267,1270,1272,1275],{"label":1268,"description":1269},"1. Shader-level ML","Small neural or linear-algebra workloads execute directly from HLSL alongside traditional shader code.",{"label":602,"description":1271},"The shader can request hardware-accelerated vector and matrix operations instead of manually implementing every ML primitive.",{"label":1273,"description":1274},"3. Model-level ML","Larger neural networks are represented as complete computation graphs rather than hand-written shader fragments.",{"label":608,"description":1276},"Microsoft's planned compiler path analyzes and lowers those graphs into optimized GPU workloads integrated with D3D12 queues and command lists.","Two ways ML can enter the DirectX graphics pipeline",{},{"id":615,"data":1280,"type":552,"tunes":1283},{"body":1281,"title":1282,"variant":619},"\u003Cstrong>DX Linear Algebra:\u003C\u002Fstrong> put small ML math inside the shader.\u003Cbr>\u003Cstrong>Compute Graph Compiler:\u003C\u002Fstrong> bring a larger ML model into the engine as a graph.","The simple difference",{},{"id":622,"data":1285,"type":572,"tunes":1287},{"text":1286,"level":47},"What DX Linear Algebra actually gives a shader",{},{"id":627,"data":1289,"type":544,"tunes":1291},{"text":1290},"Traditional HLSL is built around graphics and compute operations. Neural workloads rely heavily on linear algebra: vectors, matrices, multiplication, accumulation and data layouts optimized for those operations.",{},{"id":632,"data":1293,"type":544,"tunes":1295},{"text":1294},"The Shader Model 6.10 preview adds first-class matrix APIs so developers can express those workloads more directly and let the driver map them to specialized hardware.",{},{"id":637,"data":1297,"type":544,"tunes":1299},{"text":1298},"Microsoft's April 2026 preview explicitly describes this as a unified path for neural rendering, ML and image-processing workloads rather than a graphics-only feature.",{},{"id":642,"data":1301,"type":666,"tunes":1318},{"rows":1302,"title":1312,"layout":658,"columns":1313},[1303,1306,1309],{"id":646,"label":1304,"values":1305},"Typical work",[13,13],{"id":650,"label":1307,"values":1308},"Hardware path",[13,13],{"id":654,"label":1310,"values":1311},"Developer expression",[13,13],"Traditional shader math vs ML-oriented shader math",[1314,1316],{"id":661,"label":1315},"Traditional shader focus",{"id":664,"label":1317},"ML-oriented focus",{},{"id":669,"data":1320,"type":572,"tunes":1322},{"text":1321,"level":47},"Why Cooperative Vector was only the beginning",{},{"id":674,"data":1324,"type":544,"tunes":1326},{"text":1325},"Cooperative Vector allowed shader threads to request vector-matrix work that drivers could map to specialized hardware. That was useful for highly parallel per-pixel workloads.",{},{"id":679,"data":1328,"type":544,"tunes":1330},{"text":1329},"Microsoft later concluded that many important ML workloads need more than vector-matrix operations. Super resolution, denoising, temporal reconstruction, larger image models and general inference can require matrix-matrix operations and shared work across many threads.",{},{"id":684,"data":1332,"type":544,"tunes":1334},{"text":1333},"DX Linear Algebra therefore broadens the model instead of treating Cooperative Vector as the final abstraction.",{},{"id":689,"data":1336,"type":572,"tunes":1338},{"text":1337,"level":47},"What can actually run inside a shader?",{},{"id":694,"data":1340,"type":544,"tunes":1342},{"text":1341},"The most interesting shader-level ML workloads are small enough to execute close to the graphics data they operate on.",{},{"id":699,"data":1344,"type":658,"tunes":1364},{"content":1345,"stretched":720,"withHeadings":15},[1346,1349,1352,1355,1358,1361],[1347,1348],"Workload","Why shader-level ML fits",[1350,1351],"Neural texture compression","A small network can reconstruct texture information near the point where the shader needs it",[1353,1354],"Neural material evaluation","A learned function can replace or augment expensive hand-authored material math",[1356,1357],"Neural radiance caching","Per-pixel or local inference can estimate lighting information from learned scene behavior",[1359,1360],"Small denoising\u002Freconstruction kernels","ML operations can sit directly beside the rendering stage they improve",[1362,1363],"Image-processing inference","Matrix operations can be embedded into GPU processing without a separate external runtime",{},{"id":723,"data":1366,"type":572,"tunes":1368},{"text":1367,"level":47},"Why this matters for texture memory",{},{"id":728,"data":1370,"type":544,"tunes":1372},{"text":1371},"One of Microsoft's recurring examples is neural texture compression.",{},{"id":733,"data":1374,"type":544,"tunes":1376},{"text":1375},"Instead of storing every texture channel at conventional fidelity, a game can store a more compact representation and use a small neural network to reconstruct detail during rendering.",{},{"id":738,"data":1378,"type":544,"tunes":1380},{"text":1379},"That trades some GPU inference work for lower storage or memory pressure. The exact benefit depends on the technique and hardware, but the architectural change is important: some visual detail can become computation instead of stored data.",{},{"id":743,"data":1382,"type":749,"tunes":1387},{"url":1383,"title":1384,"excerpt":1385,"ctaLabel":1386},"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 guide to VRAM capacity, residency budgets, working sets and why memory pressure is more complicated than one usage number.","Read the VRAM guide",{},{"id":752,"data":1389,"type":572,"tunes":1391},{"text":1390,"level":47},"Why full models need a different path",{},{"id":757,"data":1393,"type":544,"tunes":1395},{"text":1394},"Hand-writing a few matrix operations in HLSL is practical for small neural functions. It becomes much less practical when the workload is a complete modern model with many layers, dependencies and intermediate tensors.",{},{"id":762,"data":1397,"type":544,"tunes":1399},{"text":1398},"Microsoft's DirectX Compute Graph Compiler is designed for this larger class of workload.",{},{"id":767,"data":1401,"type":544,"tunes":1403},{"text":1402},"Instead of rewriting the model as custom shader code, the compiler can accept a computation graph, analyze the whole graph, plan memory, fuse operations and lower the result into GPU work that integrates with DirectX 12.",{},{"id":772,"data":1405,"type":572,"tunes":1407},{"text":1406,"level":47},"The Shader-or-Graph Boundary",{},{"id":777,"data":1409,"type":666,"tunes":1432},{"rows":1410,"title":1426,"layout":658,"columns":1427},[1411,1414,1417,1420,1423],{"id":781,"label":1412,"values":1413},"Model size",[13,13],{"id":785,"label":1415,"values":1416},"Execution style",[13,13],{"id":789,"label":1418,"values":1419},"Authoring",[13,13],{"id":793,"label":1421,"values":1422},"Optimization scope",[13,13],{"id":797,"label":1424,"values":1425},"Typical use",[13,13],"When the ML workload belongs in HLSL vs a model compiler",[1428,1430],{"id":803,"label":1429},"Shader-level path",{"id":806,"label":1431},"Model-level path",{},{"id":810,"data":1434,"type":572,"tunes":1436},{"text":1435,"level":47},"Why cross-vendor support matters more than another AI feature",{},{"id":815,"data":1438,"type":544,"tunes":1440},{"text":1439},"AMD, Intel, NVIDIA and Qualcomm expose different GPU architectures and different forms of dedicated matrix acceleration.",{},{"id":820,"data":1442,"type":544,"tunes":1444},{"text":1443},"A DirectX abstraction gives Microsoft and the driver vendors a place to translate common HLSL or graph-level ML into the correct hardware path.",{},{"id":825,"data":1446,"type":544,"tunes":1448},{"text":1447},"That does not make every GPU equally fast, but it can reduce the need for a game engine to implement a completely different neural-rendering API for every vendor.",{},{"id":830,"data":1450,"type":552,"tunes":1453},{"body":1451,"title":1452,"variant":559},"The interesting part is not that DirectX can “run AI.” GPUs already could. The platform change is that \u003Cstrong>ML becomes expressible through a common graphics API and toolchain\u003C\u002Fstrong>.","Portability is the real platform story",{},{"id":836,"data":1455,"type":572,"tunes":1457},{"text":1456,"level":47},"What hardware supports the current preview?",{},{"id":841,"data":1459,"type":544,"tunes":1461},{"text":1460},"The answer depends on the specific Linear Algebra operation and preview driver.",{},{"id":846,"data":1463,"type":544,"tunes":1465},{"text":1464},"Microsoft's Agility SDK 1.721 preview support table lists LinAlg VectorAccumulate for AMD Radeon RX 9000 Series hardware, Intel Xe2-or-newer hardware through an upcoming driver, and NVIDIA RTX hardware through the supported preview path.",{},{"id":851,"data":1467,"type":544,"tunes":1469},{"text":1468},"This is preview-era support, not a universal retail guarantee. Hardware and driver support can change before finalization.",{},{"id":856,"data":1471,"type":572,"tunes":1473},{"text":1472,"level":47},"Neural rendering is becoming infrastructure",{},{"id":861,"data":1475,"type":544,"tunes":1477},{"text":1476},"DLSS, FSR and other neural graphics technologies are often discussed as branded features visible in a game's settings menu.",{},{"id":866,"data":1479,"type":544,"tunes":1481},{"text":1480},"DirectX Linear Algebra points to a deeper change. The neural operation can become an internal implementation detail inside the renderer rather than a single optional post-processing feature.",{},{"id":871,"data":1483,"type":544,"tunes":1485},{"text":1484},"A developer could use ML for textures, materials, lighting, reconstruction or other local functions without exposing each one as a consumer-facing AI toggle.",{},{"id":876,"data":1487,"type":749,"tunes":1492},{"url":1488,"title":1489,"excerpt":1490,"ctaLabel":1491},"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 3D-Guided Neural Rendering Actually Changes","How neural rendering is moving beyond upscaling and generated frames into material and lighting reconstruction inside the graphics pipeline.","Read the DLSS 5 neural-rendering guide",{},{"id":884,"data":1494,"type":572,"tunes":1496},{"text":1495,"level":47},"The Neural Rendering Stack is becoming layered",{},{"id":889,"data":1498,"type":612,"tunes":1516},{"steps":1499,"title":1515,"orientation":611},[1500,1503,1506,1509,1512],{"label":1501,"description":1502},"Traditional engine work","Simulation, geometry, visibility and base rendering remain standard engine responsibilities.",{"label":1504,"description":1505},"Inline neural shaders","Small learned functions reconstruct textures, materials or lighting inside HLSL.",{"label":1507,"description":1508},"Larger ML graphs","Full reconstruction or inference models execute through a graph-level DirectX path.",{"label":1510,"description":1511},"Vendor hardware mapping","Drivers map common DirectX operations onto the GPU's specialized AI and matrix units.",{"label":1513,"description":1514},"PIX visibility","Graphics and ML work can be profiled together instead of living in separate opaque runtimes.","A possible future DirectX game pipeline",{},{"id":910,"data":1518,"type":572,"tunes":1520},{"text":1519,"level":47},"Why unified profiling matters",{},{"id":915,"data":1522,"type":544,"tunes":1524},{"text":1523},"A neural workload that improves image quality can still damage a game if it unexpectedly consumes frame time, memory bandwidth or VRAM.",{},{"id":920,"data":1526,"type":544,"tunes":1528},{"text":1527},"Microsoft explicitly includes unified PIX visibility as part of the Compute Graph Compiler direction. That matters because developers need to see graphics and ML work in the same frame capture.",{},{"id":925,"data":1530,"type":544,"tunes":1532},{"text":1531},"If neural rendering becomes infrastructure, it must be measurable like any other rendering stage.",{},{"id":930,"data":1534,"type":572,"tunes":1536},{"text":1535,"level":47},"This does not mean every shader will become a neural network",{},{"id":935,"data":1538,"type":544,"tunes":1540},{"text":1539},"Traditional shader math remains efficient, deterministic and easy to reason about for many workloads.",{},{"id":940,"data":1542,"type":544,"tunes":1544},{"text":1543},"A neural function makes sense when it can approximate or reconstruct something expensive more efficiently, compress data, or produce quality that would otherwise require too much conventional compute or memory.",{},{"id":945,"data":1546,"type":544,"tunes":1548},{"text":1547},"The right architecture will remain hybrid.",{},{"id":950,"data":1550,"type":572,"tunes":1552},{"text":1551,"level":47},"The Neural Shader Value Test",{},{"id":955,"data":1554,"type":612,"tunes":1575},{"steps":1555,"title":1574,"orientation":611},[1556,1559,1562,1565,1568,1571],{"label":1557,"description":1558},"1. Identify the expensive traditional operation","What compute, bandwidth, memory or storage cost are you trying to replace?",{"label":1560,"description":1561},"2. Define the neural substitute","What can a small model reconstruct, predict or compress?",{"label":1563,"description":1564},"3. Measure inference cost","The ML model itself consumes GPU time, memory and bandwidth.",{"label":1566,"description":1567},"4. Measure quality stability","Look for temporal artifacts, reconstruction errors and failure cases.",{"label":1569,"description":1570},"5. Measure net savings","The technique is useful only if the avoided traditional cost is worth the added ML cost.",{"label":1572,"description":1573},"6. Test across vendors","A DirectX abstraction helps portability, but real hardware performance still differs.","When does putting ML into the rendering pipeline actually make sense?",{},{"id":979,"data":1577,"type":572,"tunes":1579},{"text":1578,"level":47},"What this means for gamers",{},{"id":984,"data":1581,"type":544,"tunes":1583},{"text":1582},"Gamers may never see a “DX Linear Algebra” switch in a graphics menu.",{},{"id":989,"data":1585,"type":544,"tunes":1587},{"text":1586},"The likely impact is indirect: smaller texture footprints, better lighting reconstruction, more efficient neural graphics, or new visual techniques that become practical because the engine can access matrix acceleration through a standard path.",{},{"id":994,"data":1589,"type":544,"tunes":1591},{"text":1590},"The feature is more important as infrastructure than as a brand.",{},{"id":999,"data":1593,"type":572,"tunes":1595},{"text":1594,"level":47},"What would change this answer?",{},{"id":1004,"data":1597,"type":544,"tunes":1599},{"text":1598},"The biggest uncertainty is the final retail shape of these APIs. DX Linear Algebra remains in preview, and the Compute Graph Compiler has not yet reached broad retail availability.",{},{"id":1009,"data":1601,"type":544,"tunes":1603},{"text":1602},"Final hardware support, API details, compiler behavior and engine adoption may change before these systems become normal shipping-game infrastructure.",{},{"id":1014,"data":1605,"type":544,"tunes":1607},{"text":1606},"If major engines adopt the abstractions directly, neural shaders could become much more common without individual game teams implementing every technique from scratch.",{},{"id":1019,"data":1609,"type":572,"tunes":1611},{"text":1610,"level":47},"Limitations",{},{"id":1024,"data":1613,"type":544,"tunes":1615},{"text":1614},"This article explains Microsoft's documented DirectX architecture and preview APIs. It does not claim that DX Linear Algebra currently improves performance in every game or that the Compute Graph Compiler is a finished retail product.",{},{"id":1029,"data":1617,"type":544,"tunes":1619},{"text":1618},"Microsoft's examples describe capability and intended use. Real benefits depend on the model, engine integration, GPU architecture, drivers and workload.",{},{"id":1034,"data":1621,"type":572,"tunes":1623},{"text":1622,"level":47},"Conclusion",{},{"id":1039,"data":1625,"type":544,"tunes":1627},{"text":1626},"The important DirectX change is not another checkbox called AI.",{},{"id":1044,"data":1629,"type":544,"tunes":1631},{"text":1630},"It is that machine-learning math is moving into the graphics programming model itself. Small neural functions can sit inside shaders, larger models can move toward graph-level compilation, and the same DirectX toolchain can expose those workloads across multiple GPU vendors.",{},{"id":1049,"data":1633,"type":544,"tunes":1635},{"text":1634},"That is how neural rendering stops being one branded feature and starts becoming part of the rendering infrastructure.",{},{"id":1054,"data":1637,"type":572,"tunes":1639},{"text":1638,"level":47},"FAQ",{},{"id":1059,"data":1641,"type":1059,"tunes":1662},{"items":1642,"title":1661},[1643,1646,1649,1652,1655,1658],{"id":1063,"answer":1644,"question":1645},"It is a DirectX\u002FHLSL feature set for hardware-accelerated vector and matrix operations used by machine learning, neural rendering and image-processing workloads.","What is DirectX Linear Algebra?",{"id":1067,"answer":1647,"question":1648},"A useful plain-English description is a shader that includes a learned neural function or ML inference step as part of its graphics work.","What is a neural shader?",{"id":1071,"answer":1650,"question":1651},"As of September 2026 it remains in the Shader Model 6.10 \u002F Agility SDK preview path.","Is DX Linear Algebra already a normal retail DirectX feature?",{"id":1075,"answer":1653,"question":1654},"It is Microsoft's announced model-level ML compiler API intended to take larger computation graphs and lower them into optimized GPU workloads integrated with D3D12.","What is the DirectX Compute Graph Compiler?",{"id":1079,"answer":1656,"question":1657},"Small functions fit shader-level authoring well, while large models benefit from whole-graph optimization, memory planning and operator fusion.","Why not run every ML model directly in HLSL?",{"id":1083,"answer":1659,"question":1660},"Not directly. DirectX provides lower-level infrastructure that vendors and developers can use for neural graphics. Branded technologies can still implement their own models and integration strategies.","Will this replace DLSS or FSR?","DirectX Linear Algebra and neural shaders",{},{"id":1089,"data":1664,"type":572,"tunes":1666},{"text":1665,"level":47},"Glossary",{},{"id":1094,"data":1668,"type":1094,"tunes":1692},{"title":1669,"entries":1670},"Key DirectX ML terms",[1671,1674,1677,1680,1683,1686,1689],{"term":1672,"anchor":1100,"definition":1673},"DX Linear Algebra","DirectX\u002FHLSL APIs for accelerated vector and matrix operations intended for neural rendering, ML and image-processing workloads.",{"term":1675,"anchor":1104,"definition":1676},"Cooperative Vector","An earlier DirectX approach for accelerated vector-matrix operations inside shaders that helped establish the shader-level neural-rendering path.",{"term":1678,"anchor":1108,"definition":1679},"Shader Model 6.10","The preview shader-model generation containing the current DirectX Linear Algebra matrix APIs.",{"term":1681,"anchor":1112,"definition":1682},"DirectX Compute Graph Compiler","Microsoft's announced compiler API for optimizing and executing larger ML computation graphs as native DirectX GPU workloads.",{"term":1684,"anchor":1116,"definition":1685},"Neural shader","A practical term for a shader that performs learned inference or neural computation as part of graphics processing.",{"term":1687,"anchor":1120,"definition":1688},"Shader-or-Graph Boundary","A Figure Rocks model for deciding whether an ML workload belongs as small inline shader math or as a larger model-level computation graph.",{"term":1690,"anchor":1123,"definition":1691},"Neural Shader Value Test","A Figure Rocks workflow for deciding whether the traditional compute or memory cost avoided by a neural technique is worth its inference cost and quality trade-offs.",{},{"id":1127,"data":1694,"type":572,"tunes":1696},{"text":1695,"level":47},"Primary sources",{},{"id":1132,"data":1698,"type":1139,"tunes":1703},{"link":1134,"meta":1699},{"image":1700,"title":1701,"description":1702},{"url":13},"Microsoft DirectX — Evolving DirectX for the ML Era on Windows","Official GDC 2026 architecture overview covering shader-level ML, DX Linear Algebra and the DirectX Compute Graph Compiler.",{},{"id":1142,"data":1705,"type":1139,"tunes":1710},{"link":1144,"meta":1706},{"image":1707,"title":1708,"description":1709},{"url":13},"Microsoft DirectX — D3D12 LinAlg Matrix Preview","Official April 2026 preview explaining the unified Linear Algebra APIs, matrix operations and neural-rendering motivation.",{},{"id":1151,"data":1712,"type":1139,"tunes":1717},{"link":1153,"meta":1713},{"image":1714,"title":1715,"description":1716},{"url":13},"Microsoft DirectX — Agility SDK 1.721 Preview","Official May 2026 release documenting Shader Model 6.10 Linear Algebra updates and preview hardware support.",{},{"id":1160,"data":1719,"type":1139,"tunes":1724},{"link":1162,"meta":1720},{"image":1721,"title":1722,"description":1723},{"url":13},"Microsoft Game Dev — GDC 2026: Evolving DirectX for the ML Era","Official Microsoft Game Dev summary explaining shader-level and model-level ML and the role of the Compute Graph Compiler.",{},{"id":1169,"data":1726,"type":1139,"tunes":1731},{"link":1171,"meta":1727},{"image":1728,"title":1729,"description":1730},{"url":13},"Microsoft DirectX — D3D12 Cooperative Vector","Official background on hardware-accelerated vector\u002Fmatrix operations and neural rendering directly from shader threads.",{},"2.31.6","DirectX is moving beyond traditional graphics shaders. Microsoft is adding hardware-accelerated linear algebra directly to HLSL and a separate path for larger ML models, laying the foundation for neural textures, learned materials, neural lighting and other AI-driven rendering techniques.",{"lang":7,"title":534,"content":536,"contentJson":1735,"excerpt":1178},{"time":538,"blocks":1736,"version":1177},[1737,1740,1743,1746,1749,1752,1755,1758,1761,1764,1772,1775,1778,1781,1784,1787,1800,1803,1806,1809,1812,1815,1818,1828,1831,1834,1837,1840,1843,1846,1849,1852,1855,1858,1875,1878,1881,1884,1887,1890,1893,1896,1899,1902,1905,1908,1911,1914,1917,1920,1929,1932,1935,1938,1941,1944,1947,1950,1953,1956,1966,1969,1972,1975,1978,1981,1984,1987,1990,1993,1996,1999,2002,2005,2008,2011,2014,2024,2027,2038,2041,2046,2051,2056,2061],{"id":541,"data":1738,"type":544,"tunes":1739},{"text":543},{},{"id":547,"data":1741,"type":552,"tunes":1742},{"body":549,"title":550,"variant":551},{},{"id":555,"data":1744,"type":552,"tunes":1745},{"body":557,"title":558,"variant":559},{},{"id":562,"data":1747,"type":566,"tunes":1748},{"title":564,"maxLevel":565,"minLevel":47},{},{"id":569,"data":1750,"type":572,"tunes":1751},{"text":571,"level":47},{},{"id":575,"data":1753,"type":544,"tunes":1754},{"text":577},{},{"id":580,"data":1756,"type":544,"tunes":1757},{"text":582},{},{"id":585,"data":1759,"type":544,"tunes":1760},{"text":587},{},{"id":590,"data":1762,"type":572,"tunes":1763},{"text":592,"level":47},{},{"id":595,"data":1765,"type":612,"tunes":1771},{"steps":1766,"title":610,"orientation":611},[1767,1768,1769,1770],{"label":599,"description":600},{"label":602,"description":603},{"label":605,"description":606},{"label":608,"description":609},{},{"id":615,"data":1773,"type":552,"tunes":1774},{"body":617,"title":618,"variant":619},{},{"id":622,"data":1776,"type":572,"tunes":1777},{"text":624,"level":47},{},{"id":627,"data":1779,"type":544,"tunes":1780},{"text":629},{},{"id":632,"data":1782,"type":544,"tunes":1783},{"text":634},{},{"id":637,"data":1785,"type":544,"tunes":1786},{"text":639},{},{"id":642,"data":1788,"type":666,"tunes":1799},{"rows":1789,"title":657,"layout":658,"columns":1796},[1790,1792,1794],{"id":646,"label":647,"values":1791},[13,13],{"id":650,"label":651,"values":1793},[13,13],{"id":654,"label":655,"values":1795},[13,13],[1797,1798],{"id":661,"label":662},{"id":664,"label":665},{},{"id":669,"data":1801,"type":572,"tunes":1802},{"text":671,"level":47},{},{"id":674,"data":1804,"type":544,"tunes":1805},{"text":676},{},{"id":679,"data":1807,"type":544,"tunes":1808},{"text":681},{},{"id":684,"data":1810,"type":544,"tunes":1811},{"text":686},{},{"id":689,"data":1813,"type":572,"tunes":1814},{"text":691,"level":47},{},{"id":694,"data":1816,"type":544,"tunes":1817},{"text":696},{},{"id":699,"data":1819,"type":658,"tunes":1827},{"content":1820,"stretched":720,"withHeadings":15},[1821,1822,1823,1824,1825,1826],[703,704],[706,707],[709,710],[712,713],[715,716],[718,719],{},{"id":723,"data":1829,"type":572,"tunes":1830},{"text":725,"level":47},{},{"id":728,"data":1832,"type":544,"tunes":1833},{"text":730},{},{"id":733,"data":1835,"type":544,"tunes":1836},{"text":735},{},{"id":738,"data":1838,"type":544,"tunes":1839},{"text":740},{},{"id":743,"data":1841,"type":749,"tunes":1842},{"url":745,"title":746,"excerpt":747,"ctaLabel":748},{},{"id":752,"data":1844,"type":572,"tunes":1845},{"text":754,"level":47},{},{"id":757,"data":1847,"type":544,"tunes":1848},{"text":759},{},{"id":762,"data":1850,"type":544,"tunes":1851},{"text":764},{},{"id":767,"data":1853,"type":544,"tunes":1854},{"text":769},{},{"id":772,"data":1856,"type":572,"tunes":1857},{"text":774,"level":47},{},{"id":777,"data":1859,"type":666,"tunes":1874},{"rows":1860,"title":800,"layout":658,"columns":1871},[1861,1863,1865,1867,1869],{"id":781,"label":782,"values":1862},[13,13],{"id":785,"label":786,"values":1864},[13,13],{"id":789,"label":790,"values":1866},[13,13],{"id":793,"label":794,"values":1868},[13,13],{"id":797,"label":798,"values":1870},[13,13],[1872,1873],{"id":803,"label":804},{"id":806,"label":807},{},{"id":810,"data":1876,"type":572,"tunes":1877},{"text":812,"level":47},{},{"id":815,"data":1879,"type":544,"tunes":1880},{"text":817},{},{"id":820,"data":1882,"type":544,"tunes":1883},{"text":822},{},{"id":825,"data":1885,"type":544,"tunes":1886},{"text":827},{},{"id":830,"data":1888,"type":552,"tunes":1889},{"body":832,"title":833,"variant":559},{},{"id":836,"data":1891,"type":572,"tunes":1892},{"text":838,"level":47},{},{"id":841,"data":1894,"type":544,"tunes":1895},{"text":843},{},{"id":846,"data":1897,"type":544,"tunes":1898},{"text":848},{},{"id":851,"data":1900,"type":544,"tunes":1901},{"text":853},{},{"id":856,"data":1903,"type":572,"tunes":1904},{"text":858,"level":47},{},{"id":861,"data":1906,"type":544,"tunes":1907},{"text":863},{},{"id":866,"data":1909,"type":544,"tunes":1910},{"text":868},{},{"id":871,"data":1912,"type":544,"tunes":1913},{"text":873},{},{"id":876,"data":1915,"type":749,"tunes":1916},{"url":878,"title":879,"excerpt":880,"ctaLabel":881},{},{"id":884,"data":1918,"type":572,"tunes":1919},{"text":886,"level":47},{},{"id":889,"data":1921,"type":612,"tunes":1928},{"steps":1922,"title":907,"orientation":611},[1923,1924,1925,1926,1927],{"label":893,"description":894},{"label":896,"description":897},{"label":899,"description":900},{"label":902,"description":903},{"label":905,"description":906},{},{"id":910,"data":1930,"type":572,"tunes":1931},{"text":912,"level":47},{},{"id":915,"data":1933,"type":544,"tunes":1934},{"text":917},{},{"id":920,"data":1936,"type":544,"tunes":1937},{"text":922},{},{"id":925,"data":1939,"type":544,"tunes":1940},{"text":927},{},{"id":930,"data":1942,"type":572,"tunes":1943},{"text":932,"level":47},{},{"id":935,"data":1945,"type":544,"tunes":1946},{"text":937},{},{"id":940,"data":1948,"type":544,"tunes":1949},{"text":942},{},{"id":945,"data":1951,"type":544,"tunes":1952},{"text":947},{},{"id":950,"data":1954,"type":572,"tunes":1955},{"text":952,"level":47},{},{"id":955,"data":1957,"type":612,"tunes":1965},{"steps":1958,"title":976,"orientation":611},[1959,1960,1961,1962,1963,1964],{"label":959,"description":960},{"label":962,"description":963},{"label":965,"description":966},{"label":968,"description":969},{"label":971,"description":972},{"label":974,"description":975},{},{"id":979,"data":1967,"type":572,"tunes":1968},{"text":981,"level":47},{},{"id":984,"data":1970,"type":544,"tunes":1971},{"text":986},{},{"id":989,"data":1973,"type":544,"tunes":1974},{"text":991},{},{"id":994,"data":1976,"type":544,"tunes":1977},{"text":996},{},{"id":999,"data":1979,"type":572,"tunes":1980},{"text":1001,"level":47},{},{"id":1004,"data":1982,"type":544,"tunes":1983},{"text":1006},{},{"id":1009,"data":1985,"type":544,"tunes":1986},{"text":1011},{},{"id":1014,"data":1988,"type":544,"tunes":1989},{"text":1016},{},{"id":1019,"data":1991,"type":572,"tunes":1992},{"text":1021,"level":47},{},{"id":1024,"data":1994,"type":544,"tunes":1995},{"text":1026},{},{"id":1029,"data":1997,"type":544,"tunes":1998},{"text":1031},{},{"id":1034,"data":2000,"type":572,"tunes":2001},{"text":1036,"level":47},{},{"id":1039,"data":2003,"type":544,"tunes":2004},{"text":1041},{},{"id":1044,"data":2006,"type":544,"tunes":2007},{"text":1046},{},{"id":1049,"data":2009,"type":544,"tunes":2010},{"text":1051},{},{"id":1054,"data":2012,"type":572,"tunes":2013},{"text":1056,"level":47},{},{"id":1059,"data":2015,"type":1059,"tunes":2023},{"items":2016,"title":1086},[2017,2018,2019,2020,2021,2022],{"id":1063,"answer":1064,"question":1065},{"id":1067,"answer":1068,"question":1069},{"id":1071,"answer":1072,"question":1073},{"id":1075,"answer":1076,"question":1077},{"id":1079,"answer":1080,"question":1081},{"id":1083,"answer":1084,"question":1085},{},{"id":1089,"data":2025,"type":572,"tunes":2026},{"text":1091,"level":47},{},{"id":1094,"data":2028,"type":1094,"tunes":2037},{"title":1096,"entries":2029},[2030,2031,2032,2033,2034,2035,2036],{"term":1099,"anchor":1100,"definition":1101},{"term":1103,"anchor":1104,"definition":1105},{"term":1107,"anchor":1108,"definition":1109},{"term":1111,"anchor":1112,"definition":1113},{"term":1115,"anchor":1116,"definition":1117},{"term":1119,"anchor":1120,"definition":1121},{"term":952,"anchor":1123,"definition":1124},{},{"id":1127,"data":2039,"type":572,"tunes":2040},{"text":1129,"level":47},{},{"id":1132,"data":2042,"type":1139,"tunes":2045},{"link":1134,"meta":2043},{"image":2044,"title":1137,"description":1138},{"url":13},{},{"id":1142,"data":2047,"type":1139,"tunes":2050},{"link":1144,"meta":2048},{"image":2049,"title":1147,"description":1148},{"url":13},{},{"id":1151,"data":2052,"type":1139,"tunes":2055},{"link":1153,"meta":2053},{"image":2054,"title":1156,"description":1157},{"url":13},{},{"id":1160,"data":2057,"type":1139,"tunes":2060},{"link":1162,"meta":2058},{"image":2059,"title":1165,"description":1166},{"url":13},{},{"id":1169,"data":2062,"type":1139,"tunes":2065},{"link":1171,"meta":2063},{"image":2064,"title":1174,"description":1175},{"url":13},{},"Post erfolgreich abgerufen",{"items":2068,"source":2144,"manualIds":2145,"manualMatchedIds":2146},[2069,2075,2081,2087,2094,2101,2107,2113,2120,2126,2132,2138],{"id":2070,"slug":2071,"title":2072,"excerpt":2073,"featuredImage":14,"publishedAt":2074},"418","blathers","Blathers","Blathers amiibo属于任天堂amiibo平台大规模推广期间发布的《动物森友会》手办系列。每款手办底座内部都嵌有NFC芯片，兼具小型收藏雕塑功能。当放置在兼容的读取器上时，游戏主机会读取手办中存储的角色ID。实际应用中，这使特定游戏能直接调用对应角色。Blathers amiibo主要在支持的《动物森友会》游戏中用于解锁博物馆馆长形象或相关小型角色功能。","2026-03-06T13:06:00.000Z",{"id":2076,"slug":2077,"title":2078,"excerpt":2079,"featuredImage":14,"publishedAt":2080},"415","reese","Reese","Reese amiibo属于任天堂amiibo系列中的《动物之森》角色线，代表游戏内城镇经济体系中的商店经营者之一。与该系列其他手办相同，其价值主要不在于塑料模型本身，而在于底座内置的NFC芯片。当通过兼容的任天堂设备扫描时，该手办能触发游戏内的小型互动、解锁角色登场，或根据具体游戏版本开启额外对话与道具。","2026-03-06T09:00:00.000Z",{"id":2082,"slug":2083,"title":2084,"excerpt":2085,"featuredImage":14,"publishedAt":2086},"423","kapp-n","Kapp’n","Kapp'n amiibo属于《动物之森》amiibo手办系列，首发于该系列联动手办的第一波产品。与同系列其他手办相同，其内置微型NFC芯片，可将实体手办与多款任天堂游戏进行联动。扫描该手办可激活角色关联内容。其实用价值主要体现在：在支持游戏中召唤Kapp'n角色，并解锁与其在系列作品中角色定位相关的小型主题内容。","2026-03-06T15:42:00.000Z",{"id":2088,"slug":2089,"title":2090,"excerpt":2091,"featuredImage":2092,"publishedAt":2093},"445","amd-fsr-redstone-is-not-just-fsr-4-upscaling-frame-generation-ray-regeneration-and-radiance-caching-explained","AMD FSR Redstone 不仅仅是 FSR 4：超分辨率、帧生成、光线再生与辐射缓存详解","AMD 的 FSR 命名之所以改变，是因为 FSR 不再是一项单一功能。Redstone 现在是一套神经渲染套件，包含用于超分辨率、帧生成、光线追踪重建和学习型全局光照的独立技术。","\u002Fuploads\u002F2026\u002F09\u002Famd-fsr-redstone-is-not-just-fsr-4-upscaling-frame-generation-ray-regeneration-and-radiance-caching-explained-1790377725715-arwmri.webp","2026-09-25T17:07:00.000Z",{"id":2095,"slug":2096,"title":2097,"excerpt":2098,"featuredImage":2099,"publishedAt":2100},"451","windows-auto-sr-is-not-dlss-how-npu-upscaling-works-without-game-integration","Windows Auto SR 不是 DLSS：NPU 如何在无需游戏集成的情况下实现超分辨率","Windows Auto SR 可以在没有 DLSS、FSR 或 XeSS 集成的情况下对支持的游戏进行超分辨率处理。Windows 不是在游戏内于 GPU 上运行重建模型，而是使用 NPU 从较低分辨率的渲染中重建出更高分辨率的图像。","\u002Fuploads\u002F2026\u002F09\u002Fwindows-auto-sr-is-not-dlss-how-npu-upscaling-works-without-game-integration-1790406942266-77ihme.webp","2026-09-26T03:14:00.000Z",{"id":2102,"slug":2103,"title":2104,"excerpt":2105,"featuredImage":14,"publishedAt":2106},"419","celeste","Celeste","Celeste amiibo属于《动物森友会》系列amiibo手办的首发阵容，专为动物森友会宇宙设计。与此系列的其他手办一样，它作为一个小型NFC载体，连接至任天堂的amiibo生态系统。扫描后，该手办可将角色Celeste关联至兼容游戏中。这款amiibo的价值主要在于能够解锁角色登场及小型游戏互动，这些内容通常仅在特定条件下才会出现。","2026-03-06T14:42:00.000Z",{"id":2108,"slug":2109,"title":2110,"excerpt":2111,"featuredImage":14,"publishedAt":2112},"395","mabel","Mabel","The Mabel amiibo belongs to the Animal Crossing amiibo figure line. It represents the hedgehog tailor connected to the clothing shop that appears across the series. The figure does not introduce a new character. It transfers an established shop role into a scannable format for compatible Nintendo systems.","2026-02-23T01:58:00.000Z",{"id":2114,"slug":2115,"title":2116,"excerpt":2117,"featuredImage":2118,"publishedAt":2119},"444","pubg-ally-shows-why-ai-teammates-need-two-brains-fast-reflexes-and-slow-reasoning","PUBG Ally展示了为什么AI队友需要两个大脑：快速反射和慢速推理","语言模型可以理解战术和玩家意图，但不应该直接控制每一个动作和战斗反应。PUBG Ally展示了一种更实用的架构：用快速的行为树控制反射动作，并结合一个小型语言模型进行规划、协调和自然对话。","\u002Fuploads\u002F2026\u002F09\u002Fpubg-ally-shows-why-ai-teammates-need-two-brains-fast-reflexes-and-slow-reasoning-1790376777825-bi2zzb.webp","2026-09-25T16:51:00.000Z",{"id":2121,"slug":2122,"title":2123,"excerpt":2124,"featuredImage":14,"publishedAt":2125},"416","cyrus","Cyrus","西施惠amiibo属于任天堂将《动物森友会》系列拓展至实体NFC手办时期推出的产品线。它充当着塑料手办与支持该功能的任天堂游戏之间的桥梁。当扫描时，存储在NFC芯片中的角色便能在游戏内解锁使用。这款手办的实用价值在于能够开启与西施惠相关的互动内容和隐藏要素，这些内容通常难以通过常规方式获取。","2026-03-06T08:32:00.000Z",{"id":2127,"slug":2128,"title":2129,"excerpt":2130,"featuredImage":14,"publishedAt":2131},"420","resetti","Resetti","Resetti amiibo属于任天堂早期扩展基于NFC技术角色手办系列中的《动物森友会》amiibo手办线。与此系列其他产品类似，该手办既是角色的实体化呈现，又内置微型NFC芯片，可与兼容的任天堂设备进行通信。扫描该手办后，可将游戏角色Reset先生关联至支持的游戏，解锁与其在《动物森友会》世界观中角色定位相关的小型互动或角色登场内容。","2026-03-06T13:59:00.000Z",{"id":2133,"slug":2134,"title":2135,"excerpt":2136,"featuredImage":14,"publishedAt":2137},"241","overdrive-tuning-the-clean-way-to-reduce-blur-without-ghosting","超频调校：无重影降噪的纯净之道","过驱动可以提升清晰度或产生难看的晕影。采用这个简单方法，选择干净的中等设置，以减少模糊而不产生重影伪影。","2026-02-21T03:30:00.000Z",{"id":2139,"slug":2140,"title":2141,"excerpt":2142,"featuredImage":14,"publishedAt":2143},"127","motion-clarity-why-blur-happens-and-the-fixes-that-actually-work","运动清晰度：为何会产生模糊及其有效解决方案","动态模糊不仅仅是一个设置——它是时机与持续性的结合。掌握实用顺序：先选对模式，再调对刷新率，最后进行关键的少数清晰度优化。","2026-02-20T11:00:00.000Z","fallback",[],[]]