Re:In mobile, 62.7% of computation energy cost is spent on d...
by localhost on Wednesday November 27, 2019 @11:54AM
(#21649522)
Cerebras's "wafer scale engine" [1] takes some of these ideas and applies them narrowly to deep learning training. 400,000 cores, 18GB of on-chip memory, 9.6 petabytes of memory bandwidth in a 1.2 trillion transistor package in a gigantic 46,225mm^2 package.
[1] https://www.cerebras.net/cerebras-wafer-scale-engine-why-we-...