The Evolution of AI Chips and Their Business Impact - with Dr. Gregory Diamos

Automatic TRANSCRIPT

So. Brag. I want to kick things off by talking about this theme of a I chips kind of term being thrown around now around software serve built to purpose for ai he gives a bit of a background on what they are and maybe even where they're starting to move forward today. Sure. So if we look historically on ships, there's a few trends that are really important One really powerful tragedy in computing has just been personal computing and especially driven by the C language. So in in the past, actually, you'll probably aware everybody these days has personal computers interact with computers. They're probably aware that computers have gotten substantially faster. Since they were introduced. If we look back fifty years, the speed of computers for the types of applications that we know in love the. Power personal computers and power the Internet have gotten tremendously faster. So can actually measure this. there's a trend around. Around nineteen eighty driven by DISA-, Group of people who are early founders and early manufacturers computers. To try to quantify the speed of computers, one of the things that they developed at that time was called the stack benchmark, and this is a way of determining really objective. How fast is your computer? Suspect started around. One, thousand, nine, hundred, five, a little bit early though not a really danced around nineteen, ninety five. and. This was a metric that. Anybody was producing a computer You could runs back on it and it would give you an idea of how fast it was. So, if we look at stock performance going around nineteen, hundred, five, about two, thousand, ten, it increased basically exponentially over that thirty year run I'm so we saw this enormous over a thousand times improvement in the speed of those computers. From around nine, hundred, eighty, five, two, thousand, ten. So for for various reasons, actually after that it's it's still been increasing, but it's slowed down a little bit. So we have been seeing the same. Performance increases out of our personal computers out of our phones out of the computers that are powering are our data centers. As we have in the past. So the the modern view, those those machines like the current name that we usually call them in industry is is called a CPU. CPU sometimes dance for Central Processing Unit really I think about it as a as a computer that's designed to run a specific type of application like an in particular it's designed to run kinds of applications that we build an in languages like C or or Java, or. Running on operating systems like windows or Android or us. So those computers in those chips are powering quite a lot of the workloads, quite a lot of the applications or the the computing technologies that have been transformative in personal computing in in smartphones in in in data centers. And they're connected the world through the Internet. So, ai chips are actually something different. And There's actually a convergence of maybe two forces here. That are driving this and one is the the forest I just mentioned that. You. Know as we look at the existing designs. For various reasons sometimes people one of the reasons people often referred to as the death of Moore's law whether or not. It's actually dying is is actually a highly contentious topic, but the death of Moore's law is actually one force. That's. Slowing down progress in in Cebu speeded efficiency Almond CPU development. There's some other forces to one fourth that sits a little bit technical. It's related to it's usually called Denard. Scaling. This is just the idea that it's it's getting harder and harder to make computers consume less energy. Some of the technologies that have been driving reductions energy are kind running out of steam. So for various reasons, actually progress in CBS has been slowing down not stopping but but slowing down since about two thousand ten, actually a little before that. Is, really showing up an industry you know after about two thousand ten. In around that time, a lot of people including myself were working on this. It seemed like a really hard problem. You Know How do we actually? Continue the enormous improvements we see in computing. Face with all of these challenges, there's actually a number of people in the industry who call. They're all of these walls that you'd run into and you know. So there is a joke in the industry that we basically run into a brick wall at this point. There's no way around this wall or it's it's really hard for your the leading experts in the world to figure out what to do next. You mean like n Hey, Moore's law. Hey as denard factor that all of these are just kind of like physical barriers that you couldn't surmount Yes Yeah it's hard to say like actually You know who? Maybe there will be a breakthrough in the future but in a lot of people spend a lot of really smart people spend a Lotta time young. Yeah. Yeah. And we haven't been able to find a way around young tough stuff. Yeah. So Luckily, at the same time, we figured out how to get a it to work. And that was so fortunate because the type of a sometimes called machine learning I'm just GONNA call it a I. The type of a technology that started to work really well, sometimes people called his deep learning one of the driving technologies is. The back propagation Algorithm in neural networks. And this technology is wonderful for computers. You know if you want to build a computer like a CPU that's good at running windows or running. The android operating system on your phone. Is Hard, you know the CPU. Is the best way that we know how to do that. But if you want to run something else. So if you want to run an AI algorithm like neural network if you want to train with back propagation. You can do something different you can do something completely different. And that opens up so many more opportunities to continue to improve performance and so you know since two thousand ten. As some companies have started to invest in these technologies I. Think. Actually. GP Use especially GP's from. Our an early era of investment. we've seen enormous gains like we've seen 10x gains are actually in terms of real applications, not not all real application running real applications actually seen a hundred times gains. In performance

Coming up next