2 ms·
It doesn't really work like that. It is ridiculously easy to stamp out ALUs on a die, but the problem is keeping those ALUs fed with useful data. And memory jus
by jcranmer 6y ago
It doesn't really work like that. It is ridiculously easy to stamp out ALUs on a die, but the problem is keeping those ALUs fed with useful data. And memory just can't keep up with modern CPU clock speeds. Modern CPUs use cache hierarchies to push off having to access main memory as late as possible, and even then, they use a lot of transistors to deal with the fact that even on-chip caches are slow compared to the ALUs. GPUs instead use those transistors to keep as many threads running as possible, treating a request to go to main memory as an event to switch to another bank of threads while it waits for the memory request to come back.