Inside a conference room at a Silicon Valley data center last week, chip startup Cerebras Systems founder and CEO Andrew Feldman demonstrated how the company’s technology allows people to shift between deploying different versions of an AI natural language model in a matter of moments, a task that usually takes hours or days.
“So we’ve made it 15 keystrokes to move among these largest models that have ever been described on a single machine,” Feldman said.
This, to Feldman and Cerebras, represents a triumph worth noting. Cerebras claims the system that achieved this feat has also accomplished a world first: It can train an entire 20-billion-parameter model on a single nearly foot-wide superchip. Without its technology, the company said scaling an AI model from 1 billion parameters to 20 billion parameters might require users to add more server hardware and reconfigure racks inside of a data center.
Training a natural language AI model on one chip makes it considerably cheaper and delivers a performance boost that is an order of magnitude superior to Nvidia’s flagship graphics processor-based systems, Feldman said. The idea is to give researchers and organizations with tiny budgets — in the tens of thousands of dollars range — access to AI training tools that were previously only available to much larger organizations with lots of money.
“Models have grown really fast in this area. Language processing, and the challenges of delivering compute for these models, is enormous,” Feldman said. “We sort of have made this class of model practical, useful to a whole slice of the economy that couldn’t previously do interesting work.”
The AI models that Feldman is talking about are simply methods of organizing mathematical calculations by breaking them up into steps and then regulating the communication between the steps. The point is to train a model to begin to make accurate predictions, whether that’s the next piece of code that should be written, what constitutes spam and so on.
AI models are typically large to begin with, but those built around language tend to be even larger. For language models, context — as in more text, such as adding an author’s entire body of work to a model that began with a single book — is crucial, but that context can make them far, far more complex to operate. Market-leader Nvidia estimates that AI tasks have spurred a 25-fold increase in the need for processing power every two years.
This exponential increase has led companies like Cerebras and others to chase AI as a potential market. For years, hardware investments were seen as bad bets among venture capitalists who were only willing to fund a few promising ideas. But as it became clear that AI as a class of computation would open the door for fresh ideas beyond the general purpose processors made by the likes of Intel and Nvidia, a new class of startups was born.
Cerebras, which is Latin for “mind,” is one of those startups. Founded in 2015, Feldman and his team, which includes a number of AMD veterans in key technology roles, have raised roughly $735 million — including funding from the CIA venture arm In-Q-Tel, the CEO said — at a $4.1 billion valuation.
Chips on the plate
At the core of Cerebras’ pitch is a chip that is roughly the size of a dinner plate, or an entire foot-wide silicon wafer, called the Wafer Scale Engine.
The idea of a wafer-size chip like the one that powers Cerebras’ systems isn’t a novel concept; similar ideas have been floating around for decades. A failed bid by Trilogy Systems in the early 1980s that raised roughly $750 million in today’s dollars is one notable attempt at a superchip, and IBM and others have studied the idea but never produced a product.
But together with TSMC, Cerebras has settled on a design that could be fabricated into a functioning wafer-size chip. In some ways, Cerebras is almost two startups stuck together: It’s interested in tackling the growing challenge of AI compute, but it has also achieved the technological feat of producing a useful chip the size of a wafer.
A Cerebras CS-2 system running inside a data center.Photo: Max A. Cherney/Protocol
The current generation of what Cerebras calls the WSE-2 can offer considerable performance improvements over stringing together multiple graphics chips to achieve the computational horsepower to train some of the largest AI models, according to Feldman.
“So it's unusual for a startup to have deep fab expertise, [but] we have profound expertise,” Feldman said. “And we had an idea of how they could, within their permitted flexibility in their flow, fit our innovation.”
The advantage of building a chip of that size is that it allows Cerebras to duplicate the performance of dozens of other server chips — roughly 80 graphics processors, for some large AI models — and squishes them onto a single piece of silicon. Doing so makes them considerably faster, because, in part, data can move faster across a single chip than across a network of dozens of chips.
"[Our] machine is built for one type of work,” Feldman said. “If you want to take the kids to soccer practice, no matter how shitty they are to drive, the minivan is the perfect car. But if you've got your minivan and you try and move two-by-fours and 50-pound sacks of concrete, you realize what a terrible machine it is for that job. [Our chip] is a machine for AI.”
This story was updated to correct the amount of money raised by Trilogy Systems.