2 ms·
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
by fennecfoxy 3y ago
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.