Probably in memory? Most GPUs use different ISAs, and they're mostly proprietary, so they usually have to be compiled (at least partially) at runtime.
Then again, never used CUDA so who knows
AMD provides their specification. I guess up to HD4000 series.
Re:C for high level programmers (slides)
by TheAceOfHearts on Monday June 01, 2015 @11:41AM
(#9639479)
The OS is big and scary. When you're working directly with the metal, it's SUPER simple, there's no magical abstractions. But when you're using libraries and doing system calls you don't really know what's going on. Sure, you can dig into em sometimes, but it's more than a bit daunting.
On the other hand, I downloaded the CUDA SDK. I couldn't even figure out where the GPU compiled code even resided. I suppose I was just supposed to take it as it "just works" (and it did), but it all left me highly uncomfortable.