4 ms·

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.

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.

floatboth 5y ago
> help the GPU scheduler assign tasks to specialized blocks Source? I can't find anything for '"cublas" scheduler'…
- rubatuga 5y ago
  The GEMM instructions are sent to Tensor Cores by the scheduler. The cutlass library is the open source version of cublas: https://github.com/NVIDIA/cutlass https://github.com/NVIDIA/cutlass
  - floatboth 5y ago
    Seems like you can leverage those via VK_NV_cooperative_matrix…