Maybe https://en.wikipedia.org/wiki/Minifloat is a way to cheat flop metric?
That’s already what TPUs do, basically
I think that higher precision isn’t always needed (or used efficiently).
Maybe https://en.wikipedia.org/wiki/Minifloat is a way to cheat flop metric?
That’s already what TPUs do, basically
I think that higher precision isn’t always needed (or used efficiently).