Add M5 Max benchmark — first H17 ANE on record

[lixiangnlp]

Summary

• Hardware: MacBook Pro · Apple M5 Max (6P + 12E, 128 GB) · macOS 26.4.1
• Key finding: _ANEDeviceInfo.aneSubType returns h17 on M5 Max — distinct from the h16 subtype reported by both M4 and the base M5 (per training/m5result.md). This is the first H17 ANE in community_results.json.
• Peak compute is unchanged from M4: 19.27 TFLOPS FP16 / 35.61 TOPS INT8 W8A8 (within 4% of the README's M4 figures).
• Training gains over base M5 are CPU-driven, not ANE-driven — Stories110M 73.5 ms/step (dynamic) and Qwen3-0.6B 320 ms/step (1.29× the README's M4 412 ms baseline) come from the 12 P-cores + Accelerate, while ane_ms is essentially flat across M4/M5/M5 Max.

Full probe report (inmem_peak, inmem_bench, sram_bench, ane_int8_bench, dynamic + static training) lives in benchmarks/m5max_result.md.

Files changed

• benchmarks/m5max_result.md — new, ~180-line probe + training report aligned with the format used by training/m5result.md
• benchmarks/community_results.json — adds an M5 Max training entry and updates neural_engine_specs with the H17 subtype field

Test plan

• inmem_peak, inmem_bench, sram_bench, ane_int8_bench all build with the existing one-line xcrun clang … invocations on macOS 26.4.1
• training/training_dynamic builds for both MODEL=stories110m and MODEL=qwen3_06b
• training/train_large (static pipeline) builds and runs with random init when no pretrained weight file is present
• python3 -c "import json; json.load(open('benchmarks/community_results.json'))" parses cleanly after the edit

🤖 Generated with Claude Code