AMD's Newest AI Chips Just Got a Boost in PyTorch
A quiet fix that helps AMD's newest chips run AI models the way Nvidia's do.
A PyTorch pull request titled "[ROCm][Inductor][UT] Enable the FP8 scaled_mm swizzle v2 tests (#198600)" has landed on the trunk branch. According to the description, swizzled MX scales aren't handled by any inductor template, so test_scaled_mm_v2_swizzle_compile checks that compilation falls back to the ATen _scaled_mm_v2 kernel. That path was being skipped on ROCm because SWIZZLE_32_4_4 has no hipBLASLt scale mode. The fix: gfx950 on ROCm 7.14 packs MX FP8 scales as SWIZZLE_32_8, so the change drops skipIfRocm and selects the layout with rocm_mx_swizzle. Older ROCm and non-gfx950 still skip. The test plan is running python -m pytest test/inductor/test_fp8.py -v -k test_scaled_mm_v2_swizzle_compile. The PR was authored with the assistance of Cursor (Grok 4.6), tagged by jagadish-amd, and approved by jeffdaily.
- PyTorch turned on a previously skipped test for AMD's newest AI chip, gfx950 — a quiet sign AMD hardware is being taken seriously.
- The test covers FP8, a compressed number format that makes AI run roughly twice as fast and cheaper than standard precision.
- Older AMD cards still skip the test, because they physically lack the hardware for this number layout.
- More passing tests means more AI companies willing to buy AMD chips instead of only Nvidia — which is how prices eventually come down.
Why It Matters
More working AMD support means real competition for Nvidia, which could eventually make AI cheaper to run.