dfc2bf4a51
hdim-56 fallback is textbook here (1.00x x3); padding = 2.0x. Fits the corrected fused-kernel-quality theory as a prediction, not a fit: 0.50x fused/unfused -> 2.0x padding win, exactly between M3 Air (0.40x->2.6x) and M1 Max (0.87x->1.3x). At 10c matched width the ladder is 138.9/48.8/44.0ms at hdim56 but 54.3/24.4/8.9ms at hdim64 — M5's gains live almost entirely in the fused kernel, making the head-dim fix most valuable on the newest silicon.
241 lines
13 KiB
Markdown
241 lines
13 KiB
Markdown
---
|
||
title: "M-series fleet ablation: 7 machines, M1 → M5 (16GB mini → 256GB Ultra)"
|
||
date: 2026-07-16
|
||
device: "M1 Max 32GB (24c) · M1 Ultra 128GB (64c) · M3 Air 16GB (10c, fanless) · M3 Ultra 256GB (80c) · M4 mini 16GB (10c) · M4 Pro mini 24GB · M5 MacBook Pro 16GB (10c)"
|
||
script: scripts/bench_optimizations.py (unmodified) + engine-level tiled×compile extension
|
||
---
|
||
|
||
# M-series Fleet Ablation Benchmarks
|
||
|
||
Same methodology as the 2026-03-09 wave2 doc (`--sweep ablation --resolution 512 1024 2048`,
|
||
checkpoint weights, 3 warmup + 10 bench runs), executed across **seven machines** spanning the
|
||
Apple Silicon range — **M1 Max 32GB (24c) · M1 Ultra 128GB (64c) · M3 Air 16GB (10c, fanless) ·
|
||
M3 Ultra 256GB (80c) · M4 mini 16GB (10c) · M4 Pro mini 24GB · M5 MacBook Pro 16GB (10c)** —
|
||
i.e. every current GPU generation, Studio→fanless laptop, an M1 width ladder (24c→64c) and a
|
||
**matched-width M3→M4→M5 generation ladder (all 10-core)**. Contributed from a multi-Mac render-farm setup; happy to run follow-ups on any of them.
|
||
|
||
## Headline findings
|
||
|
||
1. **`sdpa` is a ~4.5× regression on M1-class GPUs at 2048 — and free on M3.** Single-toggle
|
||
isolation at 2048 (median of 6, all-off baseline):
|
||
|
||
| toggle | M1 Ultra | M3 Ultra |
|
||
|---|---|---|
|
||
| all off | 4139 ms | 1743 ms |
|
||
| **bf16 only** | **4120 ms (neutral!)** | 1766 ms |
|
||
| **sdpa only** | **18592 ms (4.5× slower)** | 1763 ms |
|
||
| bf16+sdpa | 6388 ms | 1767 ms |
|
||
|
||
The suspected M1 danger (bf16 emulation — M1 lacks hardware bf16) is **innocent** on this
|
||
workload.
|
||
|
||
**Root cause (follow-up, same day):** it is *not* the SDPA kernel. Instrumenting a real
|
||
2048 forward shows the model's attention shapes are `head_dim=56` (global blocks
|
||
`(1,8,16384,56)` ×15, `(1,16,4096,56)` ×3; windowed `(4096,·,·,56)`). At head_dim 56,
|
||
`mx.fast.scaled_dot_product_attention` **silently falls back to the unfused path** —
|
||
fused == unfused within 1 % on every shape on both M1 and M3 Ultra. So `use_sdpa=True`
|
||
currently buys zero kernel benefit on any machine, while its branch pays extra 5-D
|
||
transpose/reshape choreography — which is what costs ~4.5× on M1 (layout-sensitive),
|
||
~free on M3.
|
||
|
||
**The actionable win: pad head_dim 56 → 64.** The fast path then engages and beats
|
||
unfused decisively — *despite ~14 % more FLOPs* (fp32, exact model shapes):
|
||
|
||
| shape (hdim 64) | M3 fused/unfused | M1 fused/unfused |
|
||
|---|---|---|
|
||
| (1, 8, 16384, ·) | 28.6 / 63.1 ms → **2.2×** | 53.8 / 144.5 ms → **2.7×** |
|
||
| (1, 16, 4096, ·) | 3.8 / 7.2 ms → 1.9× | 7.5 / 11.1 ms → 1.5× |
|
||
| (4096, 2, 64, ·) | 1.3 / 2.5 ms → 1.9× | 1.5 / 2.6 ms → 1.7× |
|
||
|
||
**On M5 it's worth ~5× — the largest win we measured.** M5 MacBook Pro (10-core GPU,
|
||
16 GB), hdim 56 vs 64, fp32:
|
||
|
||
| shape | hdim 56 (today) | hdim 64 (padded) | **speedup from padding** |
|
||
|---|---|---|---|
|
||
| (1, 16, 4096) | 44.00 ms *(fallback: 1.05× vs unfused)* | **8.91 ms** *(fast path: 0.21×)* | **4.9×** |
|
||
| (1, 16, 2048) | 11.44 ms *(1.06×)* | **2.23 ms** *(0.22×)* | **5.1×** |
|
||
| (4096, 2, 64) | 16.72 ms *(0.97×)* | **5.29 ms** *(0.25×)* | **3.2×** |
|
||
|
||
**The fallback is universal; the size of the win is not.** Measured on six machines
|
||
(padding 56→64 at `(1,16,4096)` fp32). Every machine shows the fallback at hdim 56
|
||
(fused/unfused 0.97–1.06×, i.e. the fast kernel never engages) — but the payoff varies
|
||
~4× and **does not track GPU width monotonically**:
|
||
|
||
| machine | GPU cores | hdim 56 | hdim 64 | **padding speedup** | fused/unfused @4096 |
|
||
|---|---|---|---|---|---|
|
||
| M5 MacBook Pro | 10c | 44.00 ms | 8.91 ms | **4.9×** | 0.14× |
|
||
| M1 Ultra | 64c | 150.5 ms | 53.8 ms | **2.7×** | 0.59× |
|
||
| M3 Air | 10c | 138.9 ms | 54.3 ms | **2.6×** | 0.40× |
|
||
| M3 Ultra | 80c | 63.2 ms | 28.6 ms | **2.2×** | 0.53× |
|
||
| M4 mini | 10c | 48.79 ms | 24.40 ms | **2.0×** | 0.50× |
|
||
| **M1 Max** | **24c** | 20.16 ms | 15.79 ms | **1.3×** | **0.87×** |
|
||
|
||
Note how tightly the padding win tracks the fused/unfused column (0.14×→4.9×,
|
||
0.40×→2.6×, 0.50×→2.0×, 0.87×→1.3×) and how poorly it tracks core count. The
|
||
M4 mini was measured *after* the correction below and fits it — the theory made a
|
||
prediction and held.
|
||
|
||
We initially read this as "narrower GPU ⇒ bigger win" (M5 10c gains most, M3 Ultra 80c
|
||
least). **The M1 Max refutes that**: it is mid-width (24c) yet gains the *least* of all.
|
||
The actual predictor is each chip's **fused-kernel quality relative to its own raw matmul
|
||
throughput** — the M1 Max's fused path is only ~0.87× its unfused path at every size
|
||
(i.e. MLX's SDPA kernel barely beats a plain matmul there), whereas the M5's is 0.14×.
|
||
Where the fused kernel is strong, missing it is expensive; where it's weak, missing it
|
||
barely matters.
|
||
|
||
Practical upshot is unchanged and *strictly positive*: padding to 64 helps on **every
|
||
Apple GPU generation tested** — by 1.3× (M1 Max) to 4.9× (M5) — and hurts nowhere,
|
||
despite ~14 % more FLOPs.
|
||
|
||
Suggested changes: (a) pad qkv projections to head_dim 64 (at minimum for the global
|
||
blocks) so sdpa's fast kernel actually engages — M1 benefits *more* than M3; (b) until
|
||
then, default `use_sdpa=False` (it is currently overhead-only). For MLX upstream: a
|
||
warning (or doc note) when sdpa silently falls back on unsupported head dims would have
|
||
made this obvious much sooner.
|
||
|
||
2. **`stage_gc` is harmful on Ultra-class machines at every resolution** — 0.53× at 512,
|
||
0.64× at 1024, 0.66× at 2048 on M3 Ultra, for a 1–5 % peak-memory saving. Wave2 measured
|
||
a mild 0.79–0.99× on its reference hardware; on big-memory machines it's pure overhead.
|
||
Suggestion: document as a low-memory-only flag.
|
||
|
||
3. **Tiled + compile is the best 2048 config on both Ultras** — upstream `engine.py` forced
|
||
`compile=False` in tiled mode; tiles are fixed-shape, so fused compilation applies
|
||
(patched in this fork, output bit-identical, verified against ground-truth alpha):
|
||
|
||
| engine config (2048 input) | M3 Ultra | M1 Ultra | peak | alpha MAE* |
|
||
|---|---|---|---|---|
|
||
| full-frame 2048 | 2788 ms | 5373 ms | 27.9 GB | 0.00906 |
|
||
| tiled 512/64 | 2760 ms | 4576 ms | 2.2 GB | **0.00821** |
|
||
| tiled 512/64 + compile | 2478 ms | 4222 ms | 2.3 GB | **0.00821** |
|
||
| **tiled 768/64 + compile** | **1949 ms** | **3275 ms** | 2.4 GB | 0.00842 |
|
||
| tiled 1024/64 + compile | 3475 ms | 5864 ms | 3.7 GB | 0.00915 |
|
||
|
||
\*alpha MAE vs exact ground truth: synthetic 2048² green-screen plates (soft-alpha
|
||
subject + motion-blur stripes + defocus disk composited over chroma green), hint =
|
||
8× downscaled truth. Tiled beats full-frame on *accuracy* as well as memory — the model
|
||
runs at native tile scale over full-res input. Confirms wave2's tiled-768 pick and adds
|
||
~8–13 % from compiling the tile graph.
|
||
|
||
4. **24 GB Macs must tile at 2048.** On the M4 Pro mini (24 GB), every full-frame 2048
|
||
config lands at 20–24 s/run — the ~26 GB working set swaps; toggle choice becomes noise.
|
||
At ≤1024 the M4 is healthy (218 ms @512, 1149 ms @1024). Tiled 2048 runs in ~2.3 GB.
|
||
|
||
## Cross-machine baselines (all-off)
|
||
|
||
| res | M3 Ultra | M1 Ultra | M4 Pro 24GB | wave2 reference |
|
||
|---|---|---|---|---|
|
||
| 512 | 53.3 ms | 80.8 ms | 218.1 ms | 119.6 ms |
|
||
| 1024 | 247.0 ms | 350.6 ms | 1149.2 ms | 610.7 ms |
|
||
| 2048 | 1750.2 ms | 3527.3 ms | 23147 ms (swap) | 4984.7 ms |
|
||
|
||
### Raw SDPA kernel scaling (pure MLX, no CorridorKey) — fused vs unfused, hdim 64, fp32
|
||
|
||
Useful context for *why* the head-dim fix matters differently per machine. Fused-vs-unfused
|
||
ratio (lower = fused is winning by more):
|
||
|
||
| seq | M3 Ultra (80c) | M1 Ultra (64c) | M5 (10c laptop) |
|
||
|---|---|---|---|
|
||
| 1024 | 0.46× | 0.54× | 0.28× |
|
||
| 2048 | 0.53× | 0.71× | **0.21×** |
|
||
| 4096 | 0.53× | 0.59× | **0.14×** |
|
||
|
||
Absolute fused times at 4096/fp32: M3 3.86 ms · M1 7.19 ms · **M5 7.22 ms** — an M5
|
||
laptop matches an M1 Ultra on *fused* attention while being ~4× slower unfused (51.5 ms
|
||
vs 12.2 ms). **How much the fast path is worth is chip-specific, not width-specific**
|
||
(M1 Max 24c: 0.87×; M5 10c: 0.14×) — see the six-machine table above. (M5 numbers taken
|
||
on a machine in active desktop use; treat as indicative, not lab-clean.)
|
||
|
||
### Generational delta at *matched* GPU width — M3 → M4 → M5, all 10-core
|
||
|
||
The Ultras confound generation with width. These three don't: all are 10-core GPUs, one
|
||
generation apart each (fused SDPA, hdim 64, fp32):
|
||
|
||
| seq | M3 Air (10c) | M4 mini (10c) | M5 (10c) | **M5 vs M3** |
|
||
|---|---|---|---|---|
|
||
| 1024 | 2.44 ms | 1.86 ms | 0.80 ms | **3.1×** |
|
||
| 2048 | 17.51 ms | 6.35 ms | 2.32 ms | **7.5×** |
|
||
| 4096 | 55.77 ms | 24.47 ms | 7.22 ms | **7.7×** |
|
||
|
||
At hdim **56** (what the model actually runs today) the same ladder is 138.9 → 48.8 →
|
||
44.0 ms: the M4 nearly triples the M3, but the **M5's advantage is concentrated almost
|
||
entirely in the fused kernel** — its unpadded step is barely better than the M4's, while
|
||
its padded step is ~3× better. i.e. M5's headline MLX gains show up *only if you hit the
|
||
fast path*, which makes the head-dim fix especially valuable on the newest silicon.
|
||
|
||
~7.7× at identical core count — i.e. the M5's gain is architectural, not width. (Apple's own
|
||
published MLX figures claim ~3.8× M4→M5 on FLUX image-gen; this attention workload shows more.)
|
||
The head_dim-56 fallback reproduces on the M3 Air too — padding to 64 gives **2.6×**
|
||
(138.91 → 54.26 ms at (1,16,4096)), so the finding now holds on **M1, M3 Ultra, M3 Air, M4 Pro
|
||
and M5** — every generation Apple currently ships.
|
||
|
||
### Fanless sustained load: the M3 Air does *not* throttle on this workload (negative result)
|
||
|
||
Every number above is a burst measurement, so we checked whether a fanless chassis invalidates
|
||
them. Sustained SDPA `(1,16,2048,64)` fp32 on the M3 Air (MacBook Air, no fan), per-30s medians
|
||
over 5 minutes:
|
||
|
||
| elapsed | 30s | 60s | 150s | 240s | 300s |
|
||
|---|---|---|---|---|---|
|
||
| median | 13.85 ms | 14.91 ms | 14.55 ms | 14.43 ms | **14.04 ms** |
|
||
| vs first | 1.00× | 1.08× | 1.05× | 1.04× | **1.01×** |
|
||
|
||
**No meaningful throttle** — ~4 % wobble, ending where it started. We expected a decay curve and
|
||
did not find one; the burst numbers in this report are therefore not flattered by short runs.
|
||
Caveat: attention at this shape may not be power-dense enough to reach the thermal ceiling —
|
||
a full multi-minute diffusion pipeline could still behave differently.
|
||
|
||
## Full ablation tables
|
||
|
||
### M3 Ultra 256GB
|
||
|
||
| Config | 512 | 1024 | 2048 | peak @2048 |
|
||
|---|---|---|---|---|
|
||
| baseline | 53.3 | 247.0 | 1750.2 | 26689 MB |
|
||
| slim+sdpa+bf16+fused_decode+gpu_preprocess | 53.5 | 248.2 | 1790.1 | 27245 MB |
|
||
| slim+stage_gc+bf16+fused_decode+gpu_preprocess | 95.4 | 382.7 | 2644.3 | 26689 MB |
|
||
| slim+stage_gc+sdpa+fused_decode+gpu_preprocess | 96.3 | 385.9 | 2676.9 | 26661 MB |
|
||
| slim+stage_gc+sdpa+bf16+gpu_preprocess | 99.4 | 372.1 | 2618.6 | 26661 MB |
|
||
| slim+stage_gc+sdpa+bf16+fused_decode | 99.0 | 386.5 | 2688.0 | 26661 MB |
|
||
| stage_gc+sdpa+bf16+fused_decode+gpu_preprocess | 100.1 | 391.4 | 2628.2 | 26661 MB |
|
||
| slim+stage_gc+sdpa+bf16+fused_decode+gpu_preprocess | 101.0 | 390.9 | 2617.9 | 26661 MB |
|
||
|
||
### M1 Ultra 128GB
|
||
|
||
| Config | 512 | 1024 | 2048 | peak @2048 |
|
||
|---|---|---|---|---|
|
||
| baseline | 80.8 | 350.6 | 3527.3 | 26689 MB |
|
||
| slim+sdpa+bf16+fused_decode+gpu_preprocess | 79.1 | 353.3 | **33137.8** | 27245 MB |
|
||
| slim+stage_gc+bf16+fused_decode+gpu_preprocess | 189.5 | 705.5 | 4919.0 | 26689 MB |
|
||
| slim+stage_gc+sdpa+fused_decode+gpu_preprocess | 195.6 | 711.1 | 4939.2 | 26661 MB |
|
||
| slim+stage_gc+sdpa+bf16+gpu_preprocess | 198.2 | 681.6 | 4853.8 | 26661 MB |
|
||
| slim+stage_gc+sdpa+bf16+fused_decode | 197.7 | 714.1 | 4822.9 | 26661 MB |
|
||
| stage_gc+sdpa+bf16+fused_decode+gpu_preprocess | 194.2 | 733.2 | 6274.1 | 26661 MB |
|
||
| slim+stage_gc+sdpa+bf16+fused_decode+gpu_preprocess | 193.8 | 737.1 | 5128.2 | 26661 MB |
|
||
|
||
Note the 33.1 s outlier: the only 2048 config *without* `stage_gc` but *with* `sdpa` —
|
||
sdpa's slow path dominating once nothing throttles it (see headline 1 for the isolation).
|
||
The sdpa cliff appears only at 2048; 512/1024 are unaffected.
|
||
|
||
### M4 Pro Mac mini 24GB
|
||
|
||
| Config | 512 | 1024 | 2048 (swap-bound) |
|
||
|---|---|---|---|
|
||
| baseline | 218.1 | 1149.2 | 23147.1 |
|
||
| slim+sdpa+bf16+fused_decode+gpu_preprocess | 221.4 | 1139.0 | 20994.5 |
|
||
| slim+stage_gc+… (all stage_gc combos) | 266–273 | 1268–1292 | 20233–24253 |
|
||
|
||
At 2048 the ~26 GB working set exceeds 24 GB unified memory; all configs swap and
|
||
differences are not meaningful. ≤1024 full-frame or tiled-anything is the usable envelope.
|
||
|
||
## Recommended settings by hardware
|
||
|
||
| Hardware | 2048 recommendation |
|
||
|---|---|
|
||
| M3-class (Max/Ultra) | full-frame or tiled 768/64 + compile; every toggle optional; avoid `stage_gc` |
|
||
| M1/M2-class | **avoid `sdpa`**; tiled 768/64 + compile; avoid `stage_gc` |
|
||
| ≤24–32 GB any gen | **tiled required** (2.3 GB vs 26 GB); tiled 768/64 + compile |
|
||
|
||
*Environment: MLX (venv per `uv sync --extra mlx`), macOS 26.5, checkpoint v1.0.0. Quality
|
||
harness (ground-truth plates + scoring) available on request — it's ~150 lines and
|
||
reproduces the alpha-MAE column.*
|