老黃

在最近的next platform採訪中有這麼一段：

1 Omniverse vs Video games： 前者比後者難的多。因為後者的細節影象處理光追效果是

pre-baked

， rendered off line，

視覺戲法

（ compositing light layer - ie lighting equation is additive， by adding dynamic light to the setting will make the environment look dynamic， 再比如一個動態場景不需要顯示所有

環境因子

的實時變化） 大於真實計算；而前者是all based on physical world， real time，每一幀畫面都要實打實算出來的。這意味著數量級的算力需求；

Dassault， PTC - industry 3D simulation/digital twin

Unity， Adobe， Autodesk - gaming engine， design

USD - universal scene description -

a framework for interchange of 3D computer graphics data

， created by Pixar。 The framework focuses on collaboration， non-destructive editing， and enabling multiple views and opinions about graphics data。 。。。 usd， which can be either ASCII or binary-encoded。

目前英偉達最大的雄心壯志是earth 2， 全景地球70年氣候變遷的digital twin，這是個巨大工程，希望不會把英偉達的財力拖垮。

2 英偉達的價值大頭在於軟體，公司3/4的人力放在軟體上。軟體給客戶帶來巨大好處：只需硬體投入一筆錢，能享受從此以後多年的軟體升級帶來的performance和生產力翻N倍，以及工具箱的擴充套件擴深 （這不就是90年代的Windows平臺麼）。競爭對手很難做到這一點。為什麼？

CUDA庫是免費的，只有賣出足夠多的GPGPU（通用平臺），涵蓋各行各業的use cases， 才能養得起高質量的CUDA軟體團隊。

Hardware vs Software model： 使用者希望前者是一次性的，而後者是持續不斷的更新。這決定硬體難賺錢。推開了想，軟硬合一的境界，就是讓使用者的老舊

晶片

透過軟體更新能發揮10X效應，使用者就會願意為軟體付錢。

3 Enterprise使用者（大概指的是Adobe PTC Dassault）會要求晶片配合自己的釋出節奏，和特殊定製需求（適配自己的軟體）。這一點感受最深應該是AMD。但是通用的好處是自己能控制節奏

4 我一直有個問題，以英特爾之強大，為什麼20年還做不出來GPU？ GPU為什麼這麼難？老黃在回答AMD最新顯示卡

Aldebaran Instinct MI200

部分解答了這個問題。

4。1 超算是最簡單的，砸錢堆peak PF64 flop和記憶體。客戶不在乎錢，不在乎生態

4。2 難的是通用的。使用者啥都在乎，而且要求千奇百怪，千變萬化。而且還要跟更通用CPU競爭。而後者有

摩爾定律

支援。以前聽他講英偉達不得不打敗摩爾定律還不明白，這次聽懂了。

4。3 順序： 問題 - 解決方法（application） - algorithm - system/architecture 這才剛加速了一個問題，還有千萬個問題怎麼辦？“And that’s the reason why， in the history of time， there’s never been another computing architecture aside from CPUs until now， with accelerated computing。 And the reason for that is all the reasons that I said。 It’s just insanely hard to refactor the entire stack on the one hand， not to mention convincing the person who owns the application to do it in the first place。 They will just wait for Moore’s Law。 It’s so much easier just to buy a few more nodes。” 也就是說，他不在意AMD GPU但是在意英特爾CPU。要不是牙膏是在擠得太慢，10奈米結點出了致命delay也不會有這麼天賜的機會。To do better than that， you have to go full stack， you have to go domain by domain， and you are going to have to develop a lot of software。 You are going to be working on a lot of solvers， hacking away at it like we are。 And then of course， after almost 25 years， the architecture becomes trusted everywhere。 And so this is where we feel quite privileged。 But nonetheless， the ultimate competitor is doing nothing and waiting for Moore’s Law。 We are a $10 billion datacenter business， which is maybe five percent of datacenters。 That’s another way of saying that 95 percent of datacenters are all CPUs。 And that’s the competition。

In this new world of computing， most of the hard problems that you want to solve are not are not 50 percent away or 2X away。 It’s a million times away。 For the very first time in history， the confluence of three or four things， you know， have come together that makes it possible for us to go after the 1，000，000X。 The climate

science community

will tell you that at a reasonable scale to succeed， we are probably somewhere between 100 million times or a billion times of computing away from solving the problems。 Because every time you increase the resolution by half， the amount of computation explodes – because it’s volumetric。 And the amount of physics that comes into the domain of simulation explodes。 We are talking about 1，000，000，000X computing problems。

But the thing that’s really quite amazing is that GPUs led to the democratization of deep learning， which led to physics-informed neural networks。 Can you imagine a neural network that obeys the laws of physics？ It learns physics from principle equations and by observing nature， but whenever it predicts， the loss function is governed by the

laws of physics

， so it doesn’t leave the boundaries up。 It obeys the laws of conservation of energy …So here’s the amazing thing。 If I can get a neural network to obey the laws of physics， I know one thing about neural networks that I can do very well， which is I can scale it up incredibly。 This algorithm， we know how to scale it。 So GPUs made it possible for us to do physics-informed neural networks， which allows us to then scale it back up with GPUs。 GPU acceleration gave us 20X to 50X on physics。 Now， all of a sudden， this

neural network

， because it’s multi physics， all of the partial differential equations are all learned away。 All you have left now is neural networks， and we have seen examples 10，000X to 100，000X faster – and I could keep it synchronized with my observed data that I’m getting every day from the real world。 And then on top of that， because I can parallelize it， I can now distribute it across 10，000 or 20，000 GPUs and get somewhere between 100，000，000X to 1，000，000，000X。 That’s why I’m going to go build Omniverse。 The time has come where we could go create these incredible digital simulations of the world。 We are going to give our ourselves a leap and this will change computer science。 I am completely convinced this is going to change scientific computing altogether。 This is really about how computing has fundamentally changed because the computer science changed the algorithms and now the algorithms are coming came back to change the computer science。