docs(MRP_MLX): m4pro is hands-on (24GB caps it, not policy)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
John King 2026-08-24 11:51:35 +10:00
parent 8c63e21df7
commit 80ca4574b6

View File

@ -39,7 +39,7 @@ after the rung-3 kernel push; every number here was measured, not estimated.
| m3ultra | M3 Ultra | 256GB | 23.5 | 400 GB/s | primary: TRELLIS.2, queue, heavy gen | | m3ultra | M3 Ultra | 256GB | 23.5 | 400 GB/s | primary: TRELLIS.2, queue, heavy gen |
| m1ultra | M1 Ultra | 128GB | 16.3 | 297 GB/s | second lane: everything (post rung-3) | | m1ultra | M1 Ultra | 128GB | 16.3 | 297 GB/s | second lane: everything (post rung-3) |
| m1max | M1 Max | 32GB | 7.1 | 336 GB/s | bandwidth-friendly light ops | | m1max | M1 Max | 32GB | 7.1 | 336 GB/s | bandwidth-friendly light ops |
| m4pro | M4 Pro | 24GB | 5.2 | 225 GB/s | MB light ops + Ollama (hands off) | | m4pro | M4 Pro | 24GB | 5.2 | 225 GB/s | MB light ops + Ollama (24GB caps it) |
Full 7-machine MLX matrix + probe script: `BENCHMARKS.md` (2026-07-19) + Full 7-machine MLX matrix + probe script: `BENCHMARKS.md` (2026-07-19) +
`scripts/mlxprobe.py`. Rules from it: **fp16 on M1-class** (bf16 20%), `scripts/mlxprobe.py`. Rules from it: **fp16 on M1-class** (bf16 20%),