Cerebras AI compute system 200x faster than supercomputer

Cerebras has worked with the Department of Energy’s National Energy Technology Laboratory (NETL) to demonstrate its CS-1 system, which is based on the world’s largest chip.

 The CS-1 is powered by Cerebras’ Wafer Scale Engine (WSE), which was first presented in November 2019.

While chipmakers typically cut a wafer into hundreds of separate chips, Cerebras creates a single enormous chip from a wafer with interconnected 400,000 AI cores, 1.2 trillion transistors, and 18GB of RAM. Each core has its own private memory and is interconnected to other cores in a sophisticated arrangement, which optimises speed.

At 462cm2, the WSE is 56 times larger than the largest GPU.

Cerebras worked with the NETL to see how well the CS-1 could handle extremely computationally expensive tasks such as forecasting the weather, finding the optimal shape for an aircraft wing, predicting temperatures and radiation levels within a nuclear power plant, modelling combustion in a coal plant, or making models of the layers of sedimentary rock in oil- and gas-rich areas.

The CS-1 completed computational fluid dynamics tasks 200 times faster than the NETL’s Joule Supercomputer (ranked 82 in a list of the fastest 500 supercomputers).

According to a paper based on these tests, the CS-1 can provide performance which is unattainable with any number of CPUs and GPUs. Cerebras credits this to the CS-1’s memory performance, the high bandwidth and low latency of its on-wafer communication fabric, and architecture optimised for high-bandwidth computing.

“For these workloads, the wafer scale CS-1 is the fastest machine ever built,” Cerebras CEO Andrew Feldman told VentureBeat. “And it is faster than any other combination or cluster of other processors.”

He added that the CS-1 can complete calculations faster than real time; for instance, it could begin simulating the conditions inside a nuclear reactor core when the reaction begins, and complete the simulation before the real-life conditions have reached that stage.

In practical terms, a CS-1 could be used to carry out complex simulations and machine-learning tasks hundreds of times faster than is possible using supercomputers.

The Cerebras CS-10 costs several million dollars and drains up to 20kW of power.