Connect with us

Components

Giant Cerebras Processor Hundreds of Times Faster than NVIDIA’s GPU-Based Supercomputer

Published

on

The huge Cerebras WSE (Water Scale Engine) processor with 21.5 x 21.5 cm die and 1.2 trillion transistors has shown incredible performance in a variety of tasks.

CS-1 systems with this chip occupy only 1/3 of an ordinary server rack, the blocks are only 66 cm high. But their power is colossal. And it is achieved thanks to only one processor. In computational tasks, Cerebras WSE occupies the entire silicon wafer and is 10,000 times faster than even the most powerful graphics chips.

Giant Cerebras Processor Hundreds of Times Faster than NVIDIA's GPU-Based Supercomputer

The US Department of Energy’s National Energy Technology Laboratory (NETL) compared the CS-1 to an x86-based supercomputer with nVidia GPUs called Joule. It is number 82 in the top 500 most powerful systems on the planet.

Giant Cerebras Processor Hundreds of Times Faster than NVIDIA's GPU-Based Supercomputer

Joule is equipped with 84,000 computing cores, occupies dozens of racks, consumes 450 kW and costs tens of millions of dollars. The CS-1 takes only a third of the rack, is estimated at several million dollars and consumes only 20 kW. And in a task that loaded 16 384 Joule cores, CS-1 showed an advantage of about 200 times!

However, it should be borne in mind that Cerebras processor is not suitable for all calculations. It is ideal for machine learning and AI operations. In addition, it cannot work with “raw” data. He needs an additional powerful external server for preparing and processing information.

But in hydrodynamic problems, Cerebras is simply not equal. He is able to solve them even faster than if the simulated processes took place in real time. For example, he calculates a nuclear reaction even before it stops in a real reactor.

Source: VentureBeat

Popular on Gadget Tendency