AMD Graphics Cards Get Faster AI Performance After Code Fix
This fix could make AI run smoother on AMD GPUs, saving time and frustration.
This PR fixes a ROCm reduction-heuristic bug: gfx1100, gfx1151, and gfx1200 were incorrectly treated as SM100+ NVIDIA hardware because ROCm reports the gfx major in the same field. That caused the no-split threshold to return 32768 instead of 8192, and the first R0_BLOCK value to return 1024 instead of 2048.
The change limits the no-split thresholds to normalized DeviceProperties.type == "cuda" and overrides only the architecture cap in ROCmReductionHeuristic. Workload-specific caps remain unchanged.
Regression tests cover HIP and CUDA major versions and pass without ROCm hardware, and real-device probes confirmed the corrected values. The work is scoped to RDNA, and ROCm CI should cover MI architectures.
- A bug made AMD graphics cards use inefficient settings for AI tasks.
- The fix ensures AMD cards get the correct settings, improving speed.
- This means faster AI performance on AMD-powered computers without new hardware.
Why It Matters
Faster AI on AMD graphics cards means quicker video edits, smoother games, and better AI tools for everyday users.