Comment by chickenzzzzu

It is possible to run ML workloads on for example AMD devices via Vulkan. With newer extensions like cooperative matrix, and maybe also in the future some scheduling magic exposed by the driver through a new extension, the remaining single digit percent gap CUDA has will evaporate.