That doesn't get us full CUDA support, but we can run OpenCL apps at least (with x11/nvidia-driver, x11/linux-nvidia-libs and the nv-sglrun script from emulators/libc6-shim). And NVENC if you bother to compile it.
Why would you want to limit yourself to the operating system that cannot train, say, a pytorch model on your nvidia GPU?
We had some success running inference on CPU, which works for as long as the model isn't that large. But that doesn't scale for larger models, where you also want to batch requests and run them all at once on GPU.
Comments
From https://forums.freebsd.org/threads/cuda.84926/#post-564624
Why would you want to limit yourself to the operating system that cannot train, say, a pytorch model on your nvidia GPU?
We had some success running inference on CPU, which works for as long as the model isn't that large. But that doesn't scale for larger models, where you also want to batch requests and run them all at once on GPU.
Maybe because you have different requirements and preferences (than running CUDA applications)?
Maybe people don't want to support a vendor locked environment that requires overpriced hardware that loses support quickly.