GPU在計算機架構的新黃金時代還會繼續閃耀嗎？

{"type":"doc","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"italic"},{"type":"size","attrs":{"size":10}},{"type":"strong"}],"text":"本文最初發佈於Medium網站，經原作者授權由InfoQ中文站翻譯並分享。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"John Hennessy和David Patterson在2018年6月4日以2017年圖靈獎（相當於計算機科學諾貝爾獎）的獲得者身份發表了他們的圖靈講座《"},{"type":"link","attrs":{"href":"https:\/\/cacm.acm.org\/magazines\/2019\/2\/234352-a-new-golden-age-for-computer-architecture\/fulltext","title":"","type":null},"content":[{"type":"text","text":"計算機架構新"}]},{"type":"link","attrs":{"href":"https:\/\/cacm.acm.org\/magazines\/2019\/2\/234352-a-new-golden-age-for-computer-architecture\/fulltext","title":"","type":null},"content":[{"type":"text","text":"的"}]},{"type":"link","attrs":{"href":"https:\/\/cacm.acm.org\/magazines\/2019\/2\/234352-a-new-golden-age-for-computer-architecture\/fulltext","title":"","type":null},"content":[{"type":"text","text":"黃金時代"}]},{"type":"text","text":"》。講座的"},{"type":"link","attrs":{"href":"https:\/\/cacm.acm.org\/magazines\/2019\/2\/234352-a-new-golden-age-for-computer-architecture\/fulltext#body-2","title":"","type":null},"content":[{"type":"text","text":"三個關鍵見解"}]},{"type":"text","text":"分別是："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"numberedlist","attrs":{"start":1,"normalizeStart":1},"content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":1,"align":null,"origin":null},"content":[{"type":"text","text":"軟件進步可以激發架構創新。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":2,"align":null,"origin":null},"content":[{"type":"text","text":"硬件\/軟件接口的進化爲架構創新創造了機會。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":3,"align":null,"origin":null},"content":[{"type":"text","text":"市場最終會解決架構爭論。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"我想再補充第四點，補全這個循環："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"blockquote","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在競爭中勝出的架構促進了後續的軟件進化。"}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"自從Hennessy\/Patterson的演講以來，市場可以說已經在AI領域中實現了見解#3，將圖形處理單元（GPU）推舉爲推動AI革命的架構勝出者。在本文中，我將探討AI革命是如何激發架構創新和重新發明GPU的。我希望本文能回答我自己的一個重要問題："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"blockquote","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"GPU能否在計算機架構新的黃金時代繼續閃耀？"}]}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"領域特定架構"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Henessy和Patterson提出了領域特定架構（DSA）的概念，旨在爲計算機架構帶來創新，努力邁向新的黃金時代。顧名思義，GPU是3D圖形領域的DSA。它的目標在3D虛擬世界中渲染照片般逼真的圖像；然而，幾乎所有人工智能研究人員都在使用GPU來探索超越3D圖形領域的想法，並在人工智能的“軟件”，也就是神經網絡架構方面取得了一系列突破。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"GPU在3D世界仍然是不可或缺的，同時它已成爲人工智能世界的“CPU”，因爲它促進了AI的軟件創新。除了3D用途之外，GPU架構師一直在努力將GPU的計算資源用於非3D用例。我們將這種設計理念稱爲通用GPU（GPGPU）。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"如今，我們看到行業中湧現了大批AI DSA而非GPGPU，前者試圖憑藉更好的性能來取代GPU。甚至GPU本身也掙扎在它的雙重屬性，AI DSA和3D DSA之間。原因是AI DSA需要加速張量運算，這在AI中是很常見的運算，但在3D世界中是沒有的。同時，爲3D用途準備的固定功能硬件對AI來說一般是不需要的。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"於是，架構爭論的主題應該集中在以下幾個方面："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"numberedlist","attrs":{"start":1,"normalizeStart":1},"content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":1,"align":null,"origin":null},"content":[{"type":"text","text":"GPU能否保住人工智能世界“CPU”的寶座？"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":2,"align":null,"origin":null},"content":[{"type":"text","text":"GPU是否會分成兩種DSA，一種用於AI，另一種用於3D？"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"我的預測如下："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"numberedlist","attrs":{"start":1,"normalizeStart":1},"content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":1,"align":null,"origin":null},"content":[{"type":"text","text":"GPU硬件\/軟件接口將維持GPU作爲AI世界“CPU”的地位。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":2,"align":null,"origin":null},"content":[{"type":"text","text":"基於AI的渲染會讓張量加速成爲GPU的一大支柱。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":3,"align":null,"origin":null},"content":[{"type":"text","text":"虛擬世界和現實世界互相映射的數字孿生理念將主導市場，最終解決架構爭論。"}]}]}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"GPU硬件\/軟件接口"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"我們可以將GPU在3D領域中的主導地位和在AI世界中取得的巨大成功歸功於它的硬件\/軟件接口，這種接口是GPU和3D圖形軟件架構師努力推行的。這種接口是解決以下悖論的關鍵。雖然GPU社區在繼續提升GPU的通用性，但業界的其他人已轉向更專業的硬件，以應對摩爾定律終結的困境。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/resource\/image\/50\/77\/50e88c28cc900bf1a381e7156cdb7677.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","marks":[{"type":"size","attrs":{"size":10}}],"text":"GPU流水線"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"雙層可編程性"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"從概念上講，GPU是一個處理很多階段的較長線性流水線。不同類型的工作項目在流經這個流水線時被一一處理。在早期，每個處理階段都是一個功能固定的塊。程序員對GPU能做的唯一控制就是調整每個塊的參數。如今，GPU硬件\/軟件接口讓程序員可以自由地處理每個工作項目，無論它們是頂點還是像素。開發者無需在每個頂點或像素循環中處理循環頭，因爲GPU架構師在一個固定函數中實現了它。這種架構選擇讓程序員負責關注循環體，或稱“着色器”，它通常以工作項目的類型命名，例如用於處理頂點的“頂點着色器”和用於處理像素的“像素着色器”。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"現代遊戲是如何通過這種線性流水線生成令人驚歎的畫面的呢？除了通過流水線在一個pass中控制不同類型的着色器之外，程序員還可以通過流水線的多個pass逐步生成多張中間圖像，最終生成屏幕上看到的圖像。程序員快速創建了一個計算圖，描述了中間圖像之間的關係。圖中的每個節點代表通過GPU流水線的一個pass。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"通用計算資源的集中池"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"一箇中心化的通用計算資源池被不同的處理階段之間共享，用來完成繁重的工作。這種方案的最初動機是爲了負載平衡。在不同的使用場景中，各個處理階段可能具有截然不同的負載。被稱爲着色器核心的計算資源逐漸變得更具通用性，以實現靈活性和產品差異化。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"某一天，GPU架構師嘗試將中心化着色器池作爲GPGPU提供給了非3D應用程序。這種設計方案讓GPU在AI任務方面取得了突破，甚至將AI任務作爲了自己的兼職工作。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"均衡的專業能力"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"GPU架構師時不時會在不改變硬件\/軟件接口的情況下，通過添加協處理單元來“加速”或“對領域定製”着色器池。紋理單元就是這樣一個協處理單元，紋理貼圖中的紋素通過它在到達着色器池的途中被提取和過濾。特殊函數單元（SFU）是負責執行超越數學函數的另一種協處理單元，處理對數、平方根倒數等函數。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"雖然這聽起來很像CPU中的超標量設計，但它們有一個顯著區別：GPU架構師根據着色器程序使用協處理單元的“平均”頻率來分配協處理單元的吞吐量。例如，我們可以爲紋理單元提供着色器池吞吐量的八分之一，因爲我們假設紋理操作出現在基準測試或遊戲中的時間平均爲八分之一。當協處理單元處於繁忙狀態時，GPU會切換任務以讓自己的資源被充分利用。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"用於3D用途的張量加速"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在前文中，我指出了GPU在3D用途中難以利用張量加速。我們看看如果我們改變GPU渲染典型遊戲幀的方式，這種狀況能否改變。GPU首先爲每個像素生成爲像素着色所需的所有信息，並存儲在"},{"type":"text","marks":[{"type":"strong"}],"text":"G-buffer"},{"type":"text","text":"中。從G-buffer中，我們會計算如何點亮一個像素，然後是幾個處理步驟，包括："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"numberedlist","attrs":{"start":1,"normalizeStart":1},"content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":1,"align":null,"origin":null},"content":[{"type":"text","text":"去除鋸齒邊緣（抗鋸齒，AA）"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":2,"align":null,"origin":null},"content":[{"type":"text","text":"將低分辨率圖像放大到更高精度的圖像（超分辨率，SR）"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":3,"align":null,"origin":null},"content":[{"type":"text","text":"向整個幀添加一些特定視覺效果，例如環境光遮蔽、運動模糊、布隆過濾或景深。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"我們稱這種渲染方案爲"},{"type":"link","attrs":{"href":"https:\/\/en.wikipedia.org\/wiki\/Deferred_shading","title":"","type":null},"content":[{"type":"text","text":"延遲着色"}]},{"type":"text","text":"，因爲對像素的着色是“延遲”的，直到每個像素都獲得所需的信息後纔開始。我們將照明之後的處理步驟稱爲後處理。今天，後處理消耗了大約90%的渲染時間，這意味着GPU的屏幕時間主要用在2D而非3D上！"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"NVIDIA已經展示了用來做AA和SR，基於AI的DLSS 2.0，這項技術聲稱可以生成比沒有DLSS 2.0的原生渲染圖像更好看的畫面。此外，NVIDIA還爲光線追蹤提供了基於AI的蒙特卡羅去噪算法，這樣我們就可以使用很少的光線來實現原本需要更多光線才能做到的畫面質量。另外，人工智能技術爲其他許多後處理類型提供了一類新的解決方案，例如用於環境遮蔽的NNAO和用於景深的DeepLens。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"blockquote","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"如果基於AI的後處理成爲主流，張量加速將成爲GPU在3D用途上的支柱。GPU分化爲3D DSA和AI DSA的可能性也會下降。"}]}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"3D\/AI融合"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"爲了解決架構爭論，我們要解決最後一個難題：我們最後是否應該移除3D渲染中的固定功能硬件，尤其是在用於AI用途時這樣做？請注意，通過GPGPU，GPU可以將3D渲染作爲純“軟件”來實現，而無需使用任何固定功能的硬件。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"嚴格意義上講，給定場景參數，3D渲染模擬的是光子如何從光源穿過空間，與3D虛擬世界中的對象交互。GPU的傳統3D渲染過程是這個過程的一個非常粗略的近似。因此，微軟將光線追蹤宣傳爲“未來的完整3D效果”時表示，“[基於傳統光柵化的]3D圖形是一個謊言”。然而，一位3D渲染純粹主義者可能仍然不會理會光線追蹤技術，因爲在光線追蹤過程中，我們是將光線從像素向後追蹤到3D虛擬世界來實現3D渲染的，這也是不真實的。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"這兩種方法都是基於模擬的3D渲染的近似方案。在兩種方案下，我們都會將3D虛擬世界的建模，或者說內容創建與渲染分離開來。在第一種方案下，對3D虛擬世界建模需要工程師和藝術家進行大量艱苦而富有創造性的工作，來描述每個對象及其與燈光交互方式的物理屬性。在第二種方案下，通過渲染做到完全真實是不可能的，因爲我們需要大幅簡化3D渲染以在資源預算內達成多個性能目標。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"相比使用最著名的科學知識和數學理論爲給定問題尋找解決方案，人工智能方法是從數據中“學習”計算模型，或者說神經網絡。我們通過反覆試驗迭代地調整網絡參數。我們通過先前的參數估計向前運行網絡並測量失配或稱“損失”。然後我們根據梯度調整參數以減少損失，有效地向梯度的相反方向引導損失圖景。這種機制稱爲反向傳播，要求沿前向路徑的所有計算都是可微的，以參與梯度計算。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"神經渲染是一個新興的人工智能研究領域，它使用上述方法來研究3D渲染。下面是我用來跟蹤神經渲染進展的思維導圖："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/resource\/image\/04\/7c\/045d95371f9525c4356c7b86914ff97c.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"這個3D虛擬世界模型隱式表示爲神經網絡參數（參見NeRF、GRAF、GIRAFFE），我們將真實世界圖像與我們從虛擬世界渲染的圖像對比來推斷出這些參數。然後我們反向傳播對比的梯度來調整神經網絡參數。或者，我們可以從數據中學習顯式3D網格（參見DeepMarching Cube，GAN2Shape）。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"實際上，對3D虛擬世界建模與學習神經網絡參數是一回事。這個過程要求我們在前向路徑中包含一個3D渲染流水線，並在多個緊密循環中集成3D虛擬世界的建模和渲染。通過對真實世界圖像迭代多個渲染和測試，我們獲得了可用於渲染虛擬世界新視圖的所需模型和場景參數。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在這個框架內，我們可以選擇不調整每個參數的整體，例如，保持物體的形狀完整但估計其位置（參見iNeRF）。這樣，我們可以高效地嘗試識別和定位有問題的對象，而不是對其建模。建模和識別任務之間不再存在區別。相反，問題在於我們想要“學習”或“估計”哪些場景參數。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"結論"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"因此，在人工智能解決問題的範式下，3D渲染的目標不僅是生成3D虛擬世界的逼真圖像，而且還是根據現實世界來構建虛擬世界。此外，新的框架通過以下方式重新定義了3D和AI："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"numberedlist","attrs":{"start":1,"normalizeStart":1},"content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":1,"align":null,"origin":null},"content":[{"type":"text","text":"3D渲染成爲AI訓練循環中必不可少的操作"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":2,"align":null,"origin":null},"content":[{"type":"text","text":"訓練，或“梯度下降”，過去只出現在雲中的神經網絡訓練過程，現在則是推理的一部分。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":3,"align":null,"origin":null},"content":[{"type":"text","text":"照片級圖像與保持真實世界和虛擬世界之間的對應關係是同樣重要的。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"數字孿生理念要求將龐大且不斷變化的現實世界帶到其尚未開發的孿生虛擬世界中，並不斷維持兩者之間的對應關係。通過神經渲染獲得的虛擬對象需要與通過經典方法構建的虛擬對象共存。因此，我相信神經渲染和傳統渲染將在GPU上融合，充分利用其成熟和高性能的3D流水線。數字孿生的需求將由未來的GPU負責實現。我們還需要在GPU端做很多工作來實現“可微”，以參與AI訓練循環的梯度計算，。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"假設GPU因響應3D世界中的AI進展而獲得原生可微和張量加速能力，我預計GPU的雙重人格將化爲一體。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"blockquote","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"然後，GPU會維持其首選架構的地位，繼續促進AI軟件的進一步發展，並最終在計算機架構新的黃金時代繼續閃耀。"}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"原文鏈接："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"link","attrs":{"href":"https:\/\/towardsdatascience.com\/will-the-gpu-star-in-a-new-golden-age-of-computer-architecture-3fa3e044e313","title":"","type":null},"content":[{"type":"text","text":"https:\/\/towardsdatascience.com\/will-the-gpu-star-in-a-new-golden-age-of-computer-architecture-3fa3e044e313"}]}]}]}