The CS-1 computing system from Cerebras recently demonstrated the capacity to perform certain simulations 200 times more quickly than one of the fastest supercomputers in the world.
In order to show the possible uses for the CS-1, Cerebras’ researchers ran a simulation of combustion in a power plant. The complex equations for fluid dynamics take into account over a million variables, including flow velocity, pressure, and density. The likely performance for the system was 200 times faster than the Joule Supercomputer, which is one of the 100 fastest supercomputers in the world.
The Best Modern PC Games (Summer 2020 Update)
“This means the CS-1 for this work is the fastest machine ever built, and it’s faster than any combination of clustering of other processors,” CEO Andrew Feldman told ZDNet. If you start the CS-1’s simulation at the same time as the reaction begins in the power plant’s core, the simulation will finish first. The results of the simulation, from Cerebras researcher Kamil Rocki and others from the National Energy Technology Laboratory, were published in a paper, which has not undergone peer review. Cerebras presented them at the recent SC20 supercomputing conference.
The CS-1 uses the company’s chip that takes up an entire silicon wafer. Chipmakers like Intel and Nvidia take a block of silicon and cut it into hundreds of chips, each smaller than a postage stamp. The Cerebras wafer-scale engine (WSE) is closer to the width of a piece of paper in diameter, and it’s one big chip. While a recent Nvidia supercomputer chip has 54 billion transistors, the WSE has 1.2 trillion. Transistors are the electronic switches that control electron flow, a bit like the tap on a faucet that determines the amount of water that pours out. The computers are already being used in artificial-intelligence training.
Nvidia GeForce RTX 3070
The WSE’s size and design were key in how it performed compared to the Joule supercomputer. In the same test, it was also faster than Intel’s Xeon server chip. Though Xeon cores – individual computing units – have more memory than Cerebras’, Rocki posits that the Cerebras cores are more efficient and aren’t competing for shared RAM memory the way Xeon’s are. The WSE has around 380,000 cores. The next version will have 850,000.
While the Cerebras paper speculated about some of the fluid-dynamics problems the CS-1 could help solve – like landing a helicopter on a windy day or providing equipment failure prediction – the company’s CEO wants to go further. “We need to apply this to problems beyond computational fluid dynamics,” he told ZDNet.
For more on tech, read about how Nvidia RTX GPUs will remain in short supply until at least April, and check out our recent breakdown of how the RTX 3060 stacks up against the PS5 and Xbox Series X.
Jenny McGrath is a science writer for IGN. She never tweets, but here she is @JennyMcGeez.