3 ms·
I've never understood why people say you "can't" do this or that on GPU. A GPU is made of SMs, and each SM is just a CPU with very wide SIMD pipes and very good
by quotemstr 2y ago
I've never understood why people say you "can't" do this or that on GPU. A GPU is made of SMs, and each SM is just a CPU with very wide SIMD pipes and very good hyperthreading. You can take one thread of a warp in a SM and do exactly the same things a CPU would do. Would you get 1/32 potential performance? Sure. But so what? Years ago, we did plenty of useful work with less than 1/32 of a modest CPU, and we can again.
One of the more annoying parts of the Nvidia experience is PTX. I. I know perfectly well that your CPU/SM/whatever has a program counter. Let me manipulate it directly!
- bee_rider 2y agoI think people are just saying it isn’t very cost effective to use a whole GPU as 1/32 of a modest (modern?) CPU.
- quotemstr 2y agoCars are slow and inefficient in reverse gear but a car that couldn't drive in reverse would be broken.
- bee_rider 2y agoYou put “can’t” in quotes, so I guess you are quoting somebody, but I don’t see where the quote is from, so I’m not sure what they actually meant. But I suspect they are using “can’t” informally. Like: You can’t run a drag race in reverse. Ok, you technically could, but it would be a silly thing to do.
- imtringued 2y agoWhy would that matter? Adding extra CPU cores to the GPU sounds like a much dumber idea to me. You'd be wasting silicon on something that is meant to be used infrequently. Is it really that difficult to imagine 1% of your workload doing some sort of management code, except since it is running on your GPU, you can now freely intersperse those 1% every 50 microseconds without too much overhead?