2 ms·
Many of us are non-graphics programmers interested in learning to benefit from CUDA-style parallelization. I’ve read they’re mostly incompatible, though. I know
by nickpsecurity 2y ago
Many of us are non-graphics programmers interested in learning to benefit from CUDA-style parallelization. I’ve read they’re mostly incompatible, though. I know there’s vendors supporting OpenCL and Vulkan. I didn’t see many AI projects using them, though. Makes me think cross-platform approaches aren’t great for some reason.
My questions: are there good, native implementations of WebGPU where learning it can replace learning CUDA or ROC for good, GPU utilization? If not, would a book like this teach us enough to easily pick up CUDA etc? Or are they very different to the point we’re better off learning the GPU-specific libraries?
- raphlinus 2y agoWebGPU is not yet performance competitive with CUDA, in part because the cooperative matrix multiplication ("tensor cores" in Nvidia-speak) extension is not done yet. That in turn depends on subgroups, which are pretty far along (but not yet shipping). That said, you can do machine learning in WebGPU, and, if your goal is to ship in a browser, it is viable. I personally think the performance gap can be closed with some engineering effort, and that WebGPU has the potential to become a real contender, but it's too early to say for sure. Certainly CUDA has a major head-start.