Pretty sure it's all about Vulkan compute nowadays

>>108668287
Why aren't we using a standard GPU compute API?
why are there so many different fucking GPU APIs in the first place?

>>108668336
vulkan is the unified standard, but it's too late since everyone is already locked into cuda workflows

i'm not reading your ai slop
i have plenty of my own slop to read

>>108668287
I don't even have to read it to know the answer is no.

>>108668319
AItards can't into vulkan compute, even though it's faster than cuda for most tasks and lets you tune scheduling further.

But vulkan is basically impossible to debug. you send your shader through the ICD and it gets compiled away on the GPU without any information on what instruction gets generated. It's still pretty much impossible to program GPUs directly.

Cuda however has a nice debugger and profiler.

>>108668336
>what is openmp

>>108669337
openmp works on gpu, but it's still way behind openacc in terms of specs. openmp doesn't distinguish between the different types of gpu memory. sadly only a small fraction of openacc has been implemented on gcc and clang.

also it's a lie to believe that a good multithreaded algorithm will run well on gpu. gpu are not simple co-processors and translating even simple algorithm to run on a gpu isn't an easy task.