OpenMM OpenCL — Amber20 Cellulose & STMV on Apple M5 Max

Engine: openmm-reference. Large-system measurement of OpenMM’s OpenCL backend on the M5 Max for the two heavy AMBER 20 benchmark systems.

Result

Test	Atoms	M5 Max OpenCL (ns/day)	A100 (ns/day)²	H100 1× (ns/day)²	H100 4× (ns/day)²	B200 (ns/day)²	M5 Max / A100
Cellulose PME	408,609	55.89	131.85	216.98	342.21	276.41	42%
STMV PME	1,067,095	18.14	46.05	70.16	147.09	116.91	39%

² openmm.org/benchmarks, OpenMM 8.4.

The full M5 Max scaling curve, all benchmarks so far

Combining this report with openmm-opencl-dhfr.md and openmm-opencl-apoa1.md:

System	Atoms	PME?	M5 Max ns/day	M5 Max / A100
DHFR Implicit	~2.5k	❌	1762.1	91%
DHFR Explicit-RF	23.6k	❌	1018.4	70%
DHFR Explicit-PME	23.6k	✅	752.5	58%
ApoA1 PME	92.2k	✅	231.1	48%
Cellulose PME	408.6k	✅	55.9	42%
STMV PME	1,067k	✅	18.1	39%

What this tells us

The M5 Max / A100 ratio on PME-heavy systems decelerates as system size grows, approaching an asymptote near ~35–40%:

• 23.6k → 92.2k (≈4×): 58% → 48%, a 10-point drop.
• 92.2k → 408.6k (≈4×): 48% → 42%, a 6-point drop.
• 408.6k → 1.07M (≈2.6×): 42% → 39%, a 3-point drop.

The ratio is not crashing toward zero. M5 Max is not memory-pressure-limited at STMV (1M atoms): unified memory holds the system fine, and the relative deficit vs A100 stabilizes rather than blowing up. This is the most informative thing this run produced — Apple’s unified memory does carry its weight at large scale; the gap stays bound by the per-step PME-FFT deficit, not by data movement.

For reference, even a single A100 only reaches 46 ns/day on STMV; the OpenMM table shows you need 3–4× H100 to make STMV throughput meaningful (130 to 147 ns/day). M5 Max at 18 ns/day is in the regime where single-GPU NVIDIA is also a struggle.

Provenance

• Engine: OpenMM 8.5.1.dev-f7fa0c2 (vendored at vendors/openmm/, run via the upstream stock benchmark script)
• Platform: OpenCL
• OpenCL platform name: Apple
• Device: Apple M5 Max (DeviceIndex 0)
• Host: AppCubics-MacBook-Pro.local, Darwin arm64
• Date: 2026-05-15
• Raw output: results/openmm-opencl-amber20-m5max.json (gitignored)

External inputs

These tests require the AMBER 20 Benchmark Suite (not shipped with OpenMM):

Field	Value
Source	https://ambermd.org/Amber20_Benchmark_Suite.tar.gz
Tarball size	~75 MB
Extracted size	~411 MB
Local path	results/inputs/Amber20_Benchmark_Suite/ (gitignored)
Fetched	2026-05-15

results/inputs/README.md records what is in that directory and why. The benchmark script handles the download automatically on first run; subsequent runs reuse the local copy. See the reproducer below.

Config

Parameter	Cellulose	STMV
Force field	AMBER (suite default)	AMBER (suite default)
Integrator	Langevin (NVT)	Langevin (NVT)
Timestep	4 fs	4 fs
Constraints	HBonds	HBonds
Hydrogen mass	1.5 amu	1.5 amu
PME cutoff	0.9 nm	0.9 nm
Precision	single	single
Target wall time	30 s	30 s

Identical to the configuration used at openmm.org/benchmarks.

Reproducer

cd results/inputs
UV_CACHE_DIR=/tmp/mlx-atomistic-uv-cache uv run --project ../.. \
  python ../../vendors/openmm/examples/benchmarks/benchmark.py \
    --platform OpenCL \
    --test amber20-cellulose,amber20-stmv \
    --seconds 30 \
    --precision single \
    --outfile ../openmm-opencl-amber20-m5max.json

The script downloads Amber20_Benchmark_Suite.tar.gz into the current working directory on first run and extracts it next to itself. By running from results/inputs/, the download stays inside the gitignored results/ tree rather than polluting the vendored OpenMM source under vendors/.

Open questions for follow-up benchmarks

• --disable-pme-stream on STMV: would directly test whether the PME-stream overlap path is the deficit, or whether the FFT itself is.
• Cellulose with mixed precision: large-system mixed-precision benefits on Apple GPU are unmeasured publicly.
• mlx-atomistic at the same scales — once feasible, an MLX run on a comparable system would land alongside these and tell us where the MLX runtime fits relative to OpenMM’s OpenCL ceiling on this hardware.