Cerebras’ wafer-size chip is 10,000 times faster than a GPU | VentureBeat

https://venturebeat.com/2020/11/17/cerebras-wafer-size-chip-is-10000-times-faster-than-a-gpu/amp/

Cerebras is finally hitting the racks in some locations and doing benchmarks to see how well it performs. And as compared to custom made/bespoke computer clusters with a mix of CPU/GPU, it’s performing very well. One benchmark has shown it is a significant improvement over the #82 ranked supercomputer (200x faster on certain benchmarks) and performs those calculations at a reduced power consumption 20KW versus 450KW (for #82 ranked super computer). So as a first time out, it seems like some of the promises and hand-waving made by the company are being proved out. The question still remains as Venture Beat points out, is how applicable Cerebras CS-1 is to a range of compute problems. The value of any of these multi-million dollar specialty compute machines is only as good as the range of problems you can solve with them. Thinking back to the 1970s, Control Data Corporation spent a fortune optimizing for vector math with their first vector-based super computer. But that was all for naught when scalar type calculations were mixed in. Seymour Cray saw this and addressed it with the Cray-1 by allowing a performant mix of scalar, with much better performance vector calculations. So having a good enough performance on scalar, was so far ahead of the CDC vector-optimized that people gravitated to the Cray-1 instead. We’ll see how closely the Cerebras CS-1 hews to this pattern over the next year or so as people figure out what it is good at and whether or not it’s good value for the money.