As I dive deeper into the world of machine learning with PyTorch, I’ve been wondering about the role of graphics processing units (GPUs) in accelerating ML computations. Specifically, I’ve been using compute shaders with graphics APIs like Unreal and Vulkan for work, and I’m curious about their potential in ML.
It seems like CUDA is the primary GPU backend for most ML applications, but isn’t CUDA exclusive to NVIDIA? Can we use compute shaders for ML directly via Vulkan or DX12? I’ve been exploring options like DirectML and Onyx, and it appears that compute shaders might offer a more cross-platform solution, supporting both AMD and NVIDIA GPUs.
But is the ML world largely dominated by NVIDIA and CUDA? I’d love to hear from those with more experience in this space. Are compute shaders a viable alternative, or is CUDA the way to go for ML development? Share your thoughts!