Scalable Neighbor-Listed Nonbonded Runtime (M5 Max)

Date: 2026-07-13

Status: diagnostic. This report validates the MLX neighbor/nonbonded runtime through 92,001 atoms on a synthetic orthorhombic system. This row does not claim a GPCRmd production run, PME-at-scale support, or an OpenMM/LAMMPS performance ratio; charged PME is validated separately on the JAC workload.

Result

mlx_cell_pairs matched the tiled all-pairs MLX oracle at every requested size: 1,000, 4,000, 16,000, 50,000, and 92,001 atoms. The topology-bearing 1,000-atom case included two exclusions (one explicit and one bond-derived) and one 1-4 pair with LJ/Coulomb scales of 0.5/0.75. The lazy topology’s dense pair cache was never materialized.

atoms	compact pairs	candidates	waste	build (ms)	pair force (ms)	tiled oracle (s)	abs dE	rel dE	max abs dF
1,000	34,278	236,788	85.52%	90.48	32.28	0.025	1.22e-4	1.05e-7	2.35e-6
4,000	108,000	998,000	89.18%	22.94	3.59	0.038	0	0	5.36e-7
16,000	425,760	3,482,556	87.77%	86.87	8.15	0.517	1.95e-3	1.12e-7	7.00e-7
50,000	1,314,708	10,517,242	87.50%	267.24	24.35	12.955	7.81e-3	1.35e-7	9.54e-7
92,001	2,432,693	16,902,196	85.61%	545.01	67.69	112.055	4.69e-2	4.56e-7	8.49e-7

The 92,001-atom explicitly synchronized static decomposition was 89.0% neighbor build and 11.0% compact pair-force evaluation. The tiled all-pairs oracle took 1,655x as long as the compact pair evaluation for the same energy and forces. The unusually high first-row timings include MLX compilation/warm-up work; the larger rows show the steady scaling shape.

Two-step runtime row

A separate 92,001-atom synthetic Langevin run exercised the full MD loop with a 1.2 neighbor skin:

atoms	steps	wall (s)	steps/s	backend	compact pairs	candidates	waste	rebuild (s)	max RSS	finite
92,001	2	1.143	1.749	dynamic-neighbor+mlx_cell_pairs	7,869,217	52,355,335	84.97%	0.411	340 MiB	yes

The run completed with no fallback and relative total-energy drift of -3.55e-7. The MD report’s force_eval_wall_s measures asynchronous enqueue time, so it is not used as the compute-time split above. The synchronized static parity row supplies the build-versus-force wall times.

Provenance

• Host: Apple M5 Max, 18 CPU cores, 128 GB unified memory.
• Runtime: Python 3.13.12, MLX 0.31.2, Device(gpu, 0), Metal available.
• Commit used for measurement: 9c43d83.
• Cell: orthorhombic FCC lattice at reduced density 0.8.
• Nonbonded model: heterogeneous LJ parameters and alternating direct-cutoff charges; cutoff 2.5; no PME.
• Pair path: mlx_cell_pairs, CPU argwhere compaction, unsorted output, no fallback.
• Reference path: tiled MLX all-pairs evaluation. Triclinic compact execution remains fail-closed.

Raw outputs (gitignored):

• results/scalable-neighbor-nonbonded-runtime/parity.json
• results/scalable-neighbor-nonbonded-runtime/synthetic-runtime-92001.json

Reproduce

uv run python -m mlx_atomistic.benchmarks.neighbor_nonbonded_parity \
  --sizes 1000,4000,16000,50000,92001 \
  --out results/scalable-neighbor-nonbonded-runtime/parity.json

uv run python -m mlx_atomistic.benchmarks.md_performance \
  --sizes 92001 --steps 2 --mode dynamic-neighbor \
  --sample-interval 2 --diagnostic-interval 2 --evaluation-interval 2 \
  --json-out results/scalable-neighbor-nonbonded-runtime/synthetic-runtime-92001.json

The GPCRmd cache was absent when this 2026-07-13 measurement was made, so this report contains no real-fixture result. A later source-backed run uses the source protocol and the production NeighborBlocks PME path; its commands and evidence are recorded separately in gpcrmd-729-pme-runtime-m5max.md.

Comparison status and next blocker

This row is diagnostic: the new parity workload uses heterogeneous LJ+Coulomb parameters and topology semantics, while the existing OpenMM/LAMMPS scaling row uses a uniform reduced-unit LJ fluid. A ratio would mix physics and is therefore not reported.

The neighbor/topology axis is proven at the target atom count. A separate 94,232-atom charged JAC report now validates fixed-cell orthorhombic PME with an explicit neutralizing-plasma policy; see scalable-charged-pme-runtime-m5max.md. That does not turn this synthetic row into a GPCRmd result. The membrane fixture has since passed a separate bounded fixed-cell NVT parity/runtime/restart gate. Production NPT, analytic PME virial, triclinic PME, production-length stability, and broad membrane readiness remain deferred.