Cerebras's Condor Galaxy AI supercomputer takes flight carrying 36 exaFLOPS

Someone should unit-ify floating point precision. It makes it tough cause we see "36 EFLOPS" not knowing till we dig deeper that it's FP16.

Like 36 EFLOPS2, or say 10 TFLOPS4 (for FP32) or kFLOPS8 (for FP64). I suppose as long as it's byte aligned and IEEE flops. Hmmm. Or just include "36 EFLOPS FP16" in the title I guess, ha ha.

On the plus side, if killer AIs need datacenters this big, it will make it easier to kill them back (if anybody survives their initial rampage, that is.)

I think the article is wrong, it should have 2.56TB of SRAM with 64 nodes.

But can it run PyTorch?

It can indeed