COMPUTER BRAINS are tiny rectangles, becoming tinier with each new generation. Or so it used to be. These days Andrew Feldman, the boss of Cerebras, a startup, pulls a block of Plexiglas out of his backpack. Baked into it is a microprocessor the size of letter paper. “It’s the world’s biggest,” he says proudly, rattling off its technical specs: 400,000 cores (sub-brains), 18 gigabytes of memory and 1.2trn transistors. That is, respectively, about 78, 3,000 and 57m times more than the largest existing processor from Nvidia, a chipmaker.
Cerebras is leading a shift in semiconductors that was on display at Hot Chips, an industry gathering at Stanford University, where startups like Mr Feldman’s and giants such as Nvidia and Intel showed off their new silicon wares on August 19th.
Cramming ever more transistors on standard chips—twice as many every 18 months, according to Moore’s Law, which has turned from an empirical observation to an industry benchmark—used to be the way to go. But with transistors now the size of dozens of atoms, improvements have become less predictable. And with the spread of artificial intelligence (AI), demand for computing power has grown by more than 300,000 times for certain applications between 2012 and 2018, according to some estimates, much faster than the 16 times or so ordained by Moore’s Law.
As a result, chipmakers are now dialling up performance by, among other things, increasing the size of processors that inhale data to train AI services, from facial recognition to drug discovery. Cerebras has pushed this approach to the limit: its chip is the biggest that can be cut from the largest available wafers, the round sheets of silicon onto which transistors are etched.
To get there, the firm had to overcome more than one technical hurdle. One is defects: every wafer has some, so Mr Feldman’s team had to find a way to bypass faulty cores. Another is cooling: water pumped through tiny pipes carries away the great heat that cores generate. Cerebras has also built a specialised computer for its new chip which it claims will deliver 150 times more number-crunching power than the best server based on graphics-processing units, today’s AI workhorses.
Those attending Hot Chips were passably impressed when Cerebras presented its new processor. But the biggest hurdle for Cerebras may be economic, not technical, says Linley Gwennap of Microprocessor Report, an industry newsletter.
One question is whether other firms that have a huge demand for computing power, including banks and oil giants, will buy such AI supercomputers, instead of having their data crunched in a cloud. And Mr Feldman has yet to convince big providers of cloud computing, such as Amazon Web Services, Microsoft Azure and Google Cloud, that Cerebras’s superior performance relative to machines packed with Nvidia chips outweighs the extra costs, for instance in higher power consumption.
Though the name Cerebras is meant to echo cerebrum, the largest part of the human brain, it also bears a resemblance to Cerberus, the giant three-headed dog guarding the entrance to Hades. The industry’s more fearsome beasts, whether chipmakers or computer manufacturers, may need little convincing to gobble it up while it is still a puppy, as they have done with other AI-chip pioneers before it.■
This post was originally posted at https://www.economist.com/business/2019/08/22/cerebras-unveils-the-worlds-chunkiest-ai-chip.