Frequently Asked Question

Is CUDA open source?

No. CUDA is NVIDIA's proprietary platform for general-purpose GPU computing, released in 2007, and it is the single most important closed layer in the AI stack. Almost everything above it, including PyTorch, TensorFlow, NumPy, and the Hugging Face libraries, is open source under permissive licences.

The picture on the driver side has improved. NVIDIA's nvidia-open kernel modules are now the recommended path for recent hardware, replacing the proprietary blob that caused years of friction with distribution maintainers. AMD's ROCm stack is fully open, and there are open alternatives to CUDA itself in the form of Vulkan compute, SYCL, and OpenCL, though none has CUDA's ecosystem depth.

The practical position for a Linux user is that the proprietary parts sit at the hardware boundary, which is exactly where proprietary parts have always sat on Linux.