Cerebras introduced the WSE – now the WSE-1 – in August 2019. It had an area of 46,225 square millimeters and 400,000 cores and 1.2 trillion transistors. The chip is optimized for AI applications.
The WSE-2 occupies the same area as the WSE-1 but contains 850,00 cores and 2.6 trillion transistors, suggesting a manufactruring process shrink from 16nm to 7nm.
The WSE-2 has 40 Gigabytes of on-chip memory, all uniformly distributed alongside the cores, and 20 Petabytes/sec of memory bandwidth. As a result, the WSE-2 can keep the entire neural network model in on-chip memory, all of the time, on the same piece of silicon as the compute cores.
The part and computer system along with compiler are expected to ship in 3Q21.
Related links and articles: