Skip to content

Add Apple M2 local ANE benchmark results#51

Open
dailykim149656-source wants to merge 1 commit into
maderix:mainfrom
dailykim149656-source:contribute-m2-benchmark-results
Open

Add Apple M2 local ANE benchmark results#51
dailykim149656-source wants to merge 1 commit into
maderix:mainfrom
dailykim149656-source:contribute-m2-benchmark-results

Conversation

@dailykim149656-source

Copy link
Copy Markdown

Summary

Adds local Apple M2 benchmark results from a Mac mini (Mac14,3, 8 GB) on macOS 26.5.

Included:

  • Raw summarized M2 results in benchmarks/local_m2_results.md
  • M2 entry in benchmarks/community_results.json
  • M2 peak-throughput and key-finding notes in benchmarks/ANE_BENCHMARK_REPORT.md

Measurement notes

  • Tested commit: d91c9845c0784dec7753048954fc6d0e8411fe29 (origin/main)
  • Measurements were rerun from a clean worktree at /private/tmp/ANE-m2-clean
  • ANE benchmarks were run sequentially to avoid accelerator contention
  • Repeated benchmark tables use the median of three runs
  • Dynamic training is reported separately from the static training table
  • Qwen3-0.6B was not run on this 8 GB M2 machine due expected memory pressure

Verification

  • python3 -m json.tool benchmarks/community_results.json
  • git -c core.whitespace=blank-at-eol,blank-at-eof,space-before-tab,cr-at-eol diff --check -- benchmarks/local_m2_results.md benchmarks/community_results.json benchmarks/ANE_BENCHMARK_REPORT.md
  • Clean-worktree benchmark logs retained locally under /private/tmp/ane_m2_2026-06-05_clean/

Record M2 measurements from a clean origin/main worktree so cross-generation benchmark coverage includes the base M2 Mac mini alongside existing Pro/Max/Ultra data.

Constraint: Results are documentation-only and measured on Apple M2 Mac mini 8 GB, macOS 26.5, commit d91c984.

Rejected: Mixing benchmark tool fixes into this contribution | output quirks in inmem_peak and ane_int8_bench are noted but left for a separate tooling PR.

Confidence: medium

Scope-risk: narrow

Directive: Keep dynamic training separate from the static pipeline table unless a comparable static M2 run is submitted.

Tested: python3 -m json.tool benchmarks/community_results.json; git -c core.whitespace=blank-at-eol,blank-at-eof,space-before-tab,cr-at-eol diff --check -- benchmarks/local_m2_results.md benchmarks/community_results.json benchmarks/ANE_BENCHMARK_REPORT.md; clean-worktree benchmark runs logged under /private/tmp/ane_m2_2026-06-05_clean/.

Not-tested: Static train_large pipeline on M2; Qwen3-0.6B dynamic training on 8 GB M2.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant