Cerebras launches waferscale-based supercomputer

Author: EIS Release Date: Nov 21, 2022

Cerebras, the waferscale supercomputer startup, has made available for research a machine capable of an exaflop or a least a quintillion ops.

 

The machine, calked Andromeda,  is based on its waferscale processor containing 2.6 trillion transistors called Wafer-Scale Engine 2 (WSE-2) (pictured)

 

Andromeda is built by linking up 16 Cerebras CS-2 systems, the company’s latest AI computer built around WSE-2.

 

WSE-2 has 850,000 cores each of which is independently programmable and optimised for the tensor-based, sparse linear algebra operations that underpin neural network training and inference for deep learning.

 

In Andromeda, 16 WSE-2s in the CS-2 subsystem deliver 13.5 million of these cores. As the sub-systems are added to a system, they deliver near-perfect linear scaling – the performance increases directly with the number of subsystems added.

 

Key characteristics are:

 

1 ExaFLOP Al Compute, 120 PetaFLOPs Dense Compute, 16-bit half precision

• 16 CS-2s; 13.5 Million Al Cores (Frontier, #1 on the TOP500 has 8.7 Million cores)

96.8 Terabits of Internal Bandwidth

Fed by 18,176 AMD EPYC M Gen 3 processor cores (284 EPYC Gen 3 processors)