The AI ​​accelerator Cerebras Wafer-Scale Engine, introduced last year, breaks a lot of records: No processor even comes close to its size of 215 x 200 mm (462 cm 2 ) or contains a similar number of cores (400. 000) and transistors (1,2 Trillion). It is around 50 times as big as other CPUs and GPUs. Now the chip has shown that it can also beat records in arithmetic.

According to a paper presented at the Supercomputing Conference 2020, a Cerebras CS-1 computer with this processor has the Joule 2.0 supercomputer – at least Place 82 on the current TOP 500 List from November – beaten by the factor 200. The task consisted of a complex system of linear equations, such as those required for flow simulations or weather forecasts. As Cerebras explains in a blog post, the lead of the CS-1 is primarily based on its high memory performance and the fast connections between the cores.

Giant processor vs. . Cluster Joule 2.0 contains over 4000 20 – Kerner of the type Xeon Gold 6148 with a total of 86. 400 cores. He is at the US National Energy Technology Laboratory of the US Department of Energy (DOE). Its computing power is around the factor 86 below the currently fastest computer in the world Fugaku

In the meantime, Cerebras has presented the successor Cerebras Wafer Scale Engine 2. He owns 850.000 Cores and 2.6 trillion transistors. As with the predecessor, the cores are connected to each other via a 3D mesh.

(odi)