Andrew Feldman, CEO of Cerebras Systems, describes the wafer-scale processor (WSE) built to handle explosive AI compute demand. The chip has 1.2 trillion transistors and 400,000 cores, cutting training time from months to minutes and inference latency from milliseconds to microseconds. Early customers include Argonne and Lawrence Livermore national labs.
Chapters
Greetings everybody, I'm Andrew Feldman. I'm founder and CEO of Cerebra Systems. Thank you so much for joining me today. Uh, in late 2015, this is what the AI landscape looked like when my co-founders and I were together thinking about what we should Since then, it's changed a little bit. This is the AI landscape right now, and with any luck— and with any luck, that is, that is us. Um, just for those of you who love to surf, this is the, the break at Nazaré in Portugal. AI over the past 4 or 5 years has become the most important workload of our generation. It's proliferating across industries and applications. It's growing at prodigious rates, sort of a doubling every 3.5 months, vastly exceeding Moore's Law. I found this poster at a bus stop in Palo Alto recommending your third graders, uh, uh, begin after-school machine learning classes, and it reminds you that adults are, are welcome. Uh, we have grown a little bit since we started the company in, uh, by 2015. Today we are 250 people. We have offices in San Diego, Silicon Valley, and Toronto. We've raised more than $400 million backed by the industry's leading investors. We now have customers around the world And, uh, for those of you who are exceptional engineers and looking for work, we have a yawning demand for world-class engineering talent. Our mission is to change compute, and to, to do that by building a deep learning system that delivers more performance than up to 1,000 graphics processing units. While we're doing that, we deliver compute at a tiny fraction of the power and the space per unit compute And we do that while supporting all major deep learning frameworks without changes to software. The first step in that path was to solve an open problem in the industry, a problem that had not been solved for 70 years, the entire history of compute, which was to build a very large chip. This is my co-founder Sean holding our wafer-scale engine. This is a 46,000 square millimeter chip. It's 1.2 trillion transistors and 400,000 AI-optimized cores. This is a comparison between, uh, our wafer-scale engine and the largest graphics processing unit available at market. Uh, it has 54 billion transistors, is 826 millimeters square, For those of you who grew up outside of the metric system, that's about the size of a postage stamp. Uh, the Cerebras wafer-scale engine is 1.2 trillion transistors and more than 46,000 square millimeters of silicon, or roughly the size of a large dinner plate. On every important vector that underpins performance, the wafer-scale engine dwarfs the, uh, the competitive landscape. We are 56 times larger, we have 54 times more cores and 450 times more on-chip high-performance memory, nearly 6,000 times more memory bandwidth and more than 20,000 times more fabric bandwidth. It is our view that you can't build a Ferrari engine like the WaferScale Engine and put it in a Volkswagen. And expect to get race car performance. So we built the entire system, and this is the Cerebra CS1. It is 15 rack units tall, so that's about 26 inches tall, fits in a standard rack, and is the, the solution unit that we sell. No matter how fast your, your race car, no matter how fast the, the performance of your, of your system, if it's not easy use, quick to deploy, you're going to struggle. And so it was extremely important to us at, at Cerebris that you be able to deploy your existing AI models using your standard ML frameworks and existing workflows without modification. So the user writes an ML model in a standard open source framework using their existing tools and flows, and that's it. How does that happen? Well, the Cerebras Graph Compiler handles the mapping of the neural network starting in TensorFlow to the computational elements inside the CS-1 on the WaferScale engine. It goes through a set of steps, but, but simply, it takes the the neural network. It understands the calculations that are prescribed in that model. It allocates compute resources on the, on the wafer, laying out the entire neural network onto the wafer. It then does all layer work simultaneously in a pipeline. So, uh, to put simply, the value we bring to our customers is that we reduce training time from months to minutes, from weeks to seconds. That is crucial. That allows you and your, your data scientists to test thousands of new hypotheses, different ideas model variance in the same period. We also, once you've settled on a model, reduce the latency and inference not by a little bit but by a lot. Here latency shrinks from milliseconds to microseconds. One of the areas we are, are most excited about is our ability to explore networks that are not possible on graphics processing units. These are very large deep networks. These are extraordinarily sparse networks. These are wide shallow networks. All of these present fundamental problems to the computation in a graphics processing unit. All of these are, are made simple in the Cerebrus CS-1. And finally, we bring to customers a system that allows you to pour vastly more data per unit time for your training sets. And of course, more data improves results. Here's a little bit of our press about some of our wins and some of the things customers are saying about our, our product. Um, initial deployments were in some of the most challenging computational environments in the world, and these are deployed inside of supercomputers. And, uh, one of the, one of the places is at Argonne National Labs. And their, their leader said, "We are doing in a few months what would normally have taken a drug development process years to do." And that's with our system. Yesterday we announced another win in addition to the wins here, in which our system was integrated into the supercomputer at Lawrence Livermore Labs. But these are just a few of the, uh, sort of most challenging applications in the largest compute domains. We have customers around the world in a variety of different industries, from pharmaceuticals to heavy manufacturing to oil and gas. To advertising, uh, soup to nuts. So I'd like to— that's a little bit about us. I'd like to thank you for listening to our talk today. I will exit presentation mode, and let's see if I have a minute or two for, for any questions that you might have via the chat line. Left just a couple minutes, uh, for chat questions, any thoughts or questions that I might be able to answer. Greetings, Joseph. Thank you, uh, for asking a question. These are not cheap systems. They're designed to replace, uh, somewhere between hundreds and thousands of GPUs depending on your work. So this is a several million dollar machine. Other, other thoughts or questions? All right, folks, if there aren't any more questions, I'm going to give you 1 minute and 6 seconds back into your day and say that I very much appreciate you taking some time today to listen learn a little bit about, about Cerebras systems. If there's anything I can do to answer questions now or later, you can reach me directly. I'm just andrew@cerebras.net. I'm happy to answer any questions you might have at that time. Energy consumption is— Joseph, thank you for asking. Energy consumption is a fraction of what it would take to power that amount of compute in graphics processing units, but it's still a meaningful block because it's a big powerful machine. A single system takes order 20 kilowatts, so two 30-amp whips should be just fine. They're telling me that I now have one
