AMD Engineer Leverages AI To Help Make A Pure-Python AMD GPU User-Space Driver
An AMD VP used Claude AI to create a pure-Python user-space driver for AMD GPUs for debugging and testing ROCm code.Summary
Anush Elangovan, AMD's VP of AI Software, utilized Claude Code (an AI agent) to develop a pure-Python user-space driver for AMD GPUs. This driver bypasses the standard ROCm/HIP user-space stack, communicating directly with /dev/kfd and /dev/dri/renderD* via ctypes ioctls, primarily to exercise other ROCm code and aid in debugging the user-space stack, especially for SDMA and compute/communications overlap.
The driver, inspired by Tinygrad's implementation, supports features like KFD ioctl bindings, GPU family registry (RDNA2/3/4, CDNA2/3), SDMA copy engine, PM4 compute packet building, and timeline semaphores. Elangovan noted that AI agents act as a "great equalizer" in software development speed. The project has rapidly evolved to include multi-GPU support and compute-bound kernel capabilities, with initial commits showing 130 passing tests on MI300X/gfx942 hardware.
(Source：Phoronix)