Files
derek.holloway 7502bf64a4
CodeQL / Analyze (actions) (push) Canceled after 0s
CodeQL / Analyze (python) (push) Canceled after 0s
pre-commit / pre-commit (push) Canceled after 0s
Multi-Arch CI / setup (push) Canceled after 0s
Publish build_manylinux_x86_64 images / publish_build_manylinux_x86_64 (push) Canceled after 0s
Setup (Multi-Arch) / setup (push) Canceled after 0s
Multi-Arch CI / Linux::${{ fromJSON(needs.setup.outputs.linux_build_config || '{}').build_variant_label || 'skip' }} (push) Canceled after 0s
Publish TheRock Docker image / build-and-push-image (push) Canceled after 0s
Publish no_rocm_image_ubi10 image / publish_no_rocm_image_ubi10 (push) Canceled after 0s
Multi-Arch CI / Windows::${{ fromJSON(needs.setup.outputs.windows_build_config || '{}').build_variant_label || 'skip' }} (push) Canceled after 0s
Publish TheRock Docker image / build-and-push-image (push) Canceled after 0s
Multi-Arch CI / Manifest Diff (push) Canceled after 0s
TheRock Manifest Diff Report / Generate Manifest Diff Report (push) Canceled after 0s
Multi-Arch CI / CI Summary (push) Canceled after 0s
Publish no_rocm_image_ubuntu24_04 images / publish_no_rocm_image_ubuntu24_04 (push) Canceled after 0s
Publish no_rocm_image_ubuntu24_04_ocl_rt images / publish_no_rocm_image_ubuntu24_04_ocl_rt (push) Canceled after 0s
Publish TheRock Docker image / build-and-push-image (push) Canceled after 0s
Publish TheRock Docker image / build-and-push-image (push) Canceled after 0s
Publish no_rocm_image_ubuntu24_04 images / publish_no_rocm_image_ubuntu24_04_rocgdb (push) Canceled after 0s
Publish no_rocm_image_ubuntu24_04_rocgdb images / publish_no_rocm_image_ubuntu24_04_rocgdb (push) Canceled after 0s
Publish no_rocm_image_ubuntu24_04 images / publish_no_rocm_image_ubuntu24_04_ocl_rt (push) Canceled after 0s
Publish TheRock Docker image / build-and-push-image (push) Canceled after 0s
Unit Tests / Unit Tests :: ubuntu-24.04 (push) Canceled after 0s
Unit Tests / Unit Tests :: windows-2022 (push) Canceled after 0s
Publish no_rocm_image_ubuntu24_04 images / publish_no_rocm_image_ubuntu24_04_media (push) Canceled after 0s
Unit Tests / Unit Tests Summary (push) Canceled after 0s
Publish no_rocm_image_ubuntu24_04 images / publish_no_rocm_image_ubuntu24_04_openmpi (push) Canceled after 0s
Flatten subprojects into a single base monorepo
2026-09-02 16:34:47 -07:00

1.9 KiB
Raw Permalink Blame History

MI300X Benchmark Baseline

Generated from benchmarks/baseline/mi300x_baseline.json.

Hardware

  • MI300X (gfx942), 304 CUs, 192 GB HBM3
  • ROCm 6.4, kernel 5.15, Ubuntu 22.04

Suites

  1. TritonBench-ROCm — 40+ Triton-lowered kernels
  2. KernelBench-ROCm level-1 — 100+ elementary kernels

Methodology

For each kernel:

  1. rocprofv3 --sys-trace → rocprof.db
  2. perfxpert analyze -i rocprof.db --offline --emit-patch .perfxpert.patch
  3. git apply .perfxpert.patch
  4. Benchmark kernel in optimized state
  5. Report baseline_ns, optimized_ns, speedup, whether perfxpert emitted a recommendation

Per-kernel measurements use median of 30 runs after 3 warm-up iterations.

Gate

  • Weighted geomean across kernels where perfxpert recommended a change must be ≥ 1.20×
  • Weight: baseline_ns (longer-running kernels count more)

Current published numbers

See mi300x_baseline.json for machine-readable per-kernel data.

TritonBench-ROCm

  • matmul_f16_256x256: 1.000 → 0.700 ms (1.429× speedup, perfxpert applied)
  • softmax_f32_1024: 0.500 → 0.300 ms (1.667× speedup, perfxpert applied)
  • layernorm_f16_4096: 0.800 → 0.800 ms (1.000× speedup, no perfxpert recommendation)

KernelBench-ROCm level-1

  • gemm_rocblas: 1.000 → 0.750 ms (1.333× speedup, perfxpert applied)
  • softmax_fused: 0.500 → 0.500 ms (1.000× speedup, no perfxpert recommendation)

Combined

  • Total kernels: 5
  • Kernels with perfxpert recommendation: 3
  • Weighted geomean (applied only): 1.433×

Tolerances for nightly

  • Geomean delta > -2% vs baseline → investigate
  • Any single kernel regression > -10% → fail nightly
  • New kernels added → geomean computed on intersection set only

How to refresh the baseline

  1. Open PR editing benchmarks/baseline/mi300x_baseline.json
  2. CI runs suites on self-hosted MI300X runner
  3. If geomean ≥ 1.20× and no kernel regresses > 10%, PR is reviewable
  4. Two core reviewers approve → merge