Comment by chickenzzzzu

Vulkan is at 95% of CUDA performance already. The remaining 5% is CUDA's small dispatch logic.

The reason why people continue to use CUDA and Pytorch and so on is because they are literally too stupid and too lazy to do it any other way