Skip to content

Instantly share code, notes, and snippets.

@bjacob
Created July 19, 2026 02:06
Show Gist options
  • Select an option

  • Save bjacob/0928cb4b5810dc5ee528b0e86785b1f8 to your computer and use it in GitHub Desktop.

Select an option

Save bjacob/0928cb4b5810dc5ee528b0e86785b1f8 to your computer and use it in GitHub Desktop.

RocJitsu vs. FFM gfx1250 benchmark

Measured on 2026-07-19 using the same gfx1250 workload binaries and the workspace TheRock ROCm runtime. Each reported row passed its independent correctness oracle. FFM used its default worker configuration; forcing eight workers was separately found to cause severe slowdowns and is not represented in the table.

Test RocJitsu FFM Comparison
1,024-element vector-add 1.32 s 0.13 s FFM about 10x faster; both passed
Jakub attention baseline, three variants 2.44 s 1.59 s FFM about 1.5x faster; all variants passed
Qwen3-0.6B prefill baseline 139.77 s 27.71 s FFM about 5x faster; expected output matched
D128 block Inline Shadow, 18/18 accesses and 8/8 barriers 9.50 s 5.61 s FFM about 1.7x faster; oracle and coverage passed

Interpretation

  • RocJitsu is usable for gfx1250 e2e execution and correctly completed Qwen.
  • On this checkout and these workloads, FFM is faster in every completed paired comparison.
  • RocJitsu remains valuable as an independent emulator cross-check.
  • FFM's default worker configuration should be used. With an explicit eight-worker configuration, uninstrumented Qwen exceeded 180 seconds; with the default configuration it completed in 27.71 seconds.
  • A retained historical Jakub clean Inline result of 7.27 seconds is not directly comparable to the current hook revision. Fresh current-revision attempts with a broader runtime-instrumentation configuration exceeded 120 seconds on both emulators, so they are intentionally excluded from the paired table.

These are software-emulator iteration timings, not hardware performance claims. They are intended to guide backend selection for the gfx1250 ConSan port.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment