GPU Computing on Linux
Every GPU Computing on Linux lesson in the Linux for AI Engineering course, in the order you study them. Each page opens about the first half of its lesson to read, some the whole of it; the full lessons are in the plan.
The NVIDIA CUDA stack: driver and runtime compatibility, forward compatibility and the call path to the GPU
Why nvidia-smi, nvcc and PyTorch report different CUDA versions, the layers between Python and the GPU, and driver compatibility.
AMD ROCm and other accelerators: HIP portability, TPUs, Trainium and CPU inference
Why PyTorch on an AMD GPU fails with invalid device function, gfx chip names, setting up ROCm, HIP and Triton, and TPUs and Trainium.
GPU topology and management: PCIe and NUMA affinity, nvidia-smi, sharing, power caps and throttling
Why CUDA_VISIBLE_DEVICES=0 picks a different GPU than nvidia-smi, GPU topology and NVLink, picking GPUs that sit close, and sharing.
