Re:AMD Reports Q1 2021 Earnings
by floatboth on Wednesday April 28, 2021 @08:34AM
(#26968611)
It's really disappointing that the GPGPU space isn't adopting Vulkan all that fast. These "compute specific" things such as OpenCL/CUDA/ROCm should not be necessary and should die. Just use Vulkan for fuck's sake.
[ Parent
| on HN ]
Libraries like cublas and cudnn help the GPU scheduler assign tasks to specialized blocks on the graphics card. Vulkan would slow this innovation down severely. Unless they make Vulkan extensions? I’m not too sure how that would work.
Personally I haven’t used CUDA, but if anybody could comment on its ease of use compared to Vulkan that would be great.
> help the GPU scheduler assign tasks to specialized blocks
Source? I can't find anything for '"cublas" scheduler'…
The GEMM instructions are sent to Tensor Cores by the scheduler. The cutlass library is the open source version of cublas: