LAMMPS OpenCL on Apple M5 Max

Engine: lammps-reference. This report covers the five official top-level LAMMPS benchmark inputs from vendors/lammps/bench/ on the single Apple M5 Max host. It does not describe mlx_atomistic runtime performance.

Result

Case	Description	Status	Acceleration class	Loop time, s	Blocker or caveat
`lj`	Atomic Lennard-Jones fluid	`ok`	`full_gpu_opencl`	0.298473	none
`eam`	Bulk Cu EAM solid	`ok`	`full_gpu_opencl`	0.725426	none
`chain`	FENE bead-spring polymer melt	`diagnostic`	`partial_gpu_opencl`	0.544019	bond and Langevin styles are not mapped to GPU styles
`chute`	Granular chute flow	`diagnostic`	`cpu_only_diagnostic`	0.108473	no relevant mapped GPU/OpenCL styles in this build
`rhodo`	Rhodopsin in solvated lipid bilayer	`blocked`	`partial_gpu_opencl`	N/A	PPPM double-precision build conflicts with the single-precision Apple GPU path

The raw records are under results/m5max-reference/lammps/. Each run copies the official input files into a case-local work directory before invoking LAMMPS, so the shared vendor tree stays unchanged.

Environment

Field	Value
LAMMPS version	20250722
Packaged executable	`.venv/lib/python3.13/site-packages/lammps/lmp`
GPU package	`PKG_GPU=ON`
GPU API	`opencl`
GPU precision	`single`
Device	Apple M5 Max
Run command root	`UV_CACHE_DIR=/tmp/mlx-atomistic-uv-cache uv run python scripts/benchmark_m5max_reference.py`

Do not rely on .venv/bin/lmp as the final reproducer. The harness bypasses the console script and records the active packaged executable.

Acceleration Mapping

Case	Relevant styles	GPU/OpenCL mapping
`lj`	`pair lj/cut`, `fix nve`	both map to `lj/cut/gpu` and `nve/gpu`
`eam`	`pair eam`, `fix nve`	both map to `eam/gpu` and `nve/gpu`
`chain`	`bond fene`, `pair lj/cut`, `fix nve`, `fix langevin`	pair and NVE map to GPU; bond and Langevin do not
`chute`	`pair gran/hooke/history`, `fix gravity`, `fix freeze`, `fix nve/sphere`	no mapped GPU styles in this local build
`rhodo`	harmonic bonds, CHARMM angles/dihedrals, `lj/charmm/coul/long`, `pppm`, `shake`, `npt`	pair, PPPM, and NPT have GPU mappings; bonded terms and SHAKE do not

The classification is about this local LAMMPS OpenCL build, not about every possible LAMMPS build. A different build or precision setting can change the mapping and the rhodo outcome.

Blocked Case

rhodo reaches PPPM initialization and then fails:

ERROR: PPPM was compiled for double precision floating point but GPU device supports single precision only.

That is why the suite manifest status is blocked even though the manifest validates structurally. The failure is preserved in results/m5max-reference/lammps/rhodo.json, with the LAMMPS log under results/m5max-reference/lammps/rhodo/work/log.lammps.

Reproducer

UV_CACHE_DIR=/tmp/mlx-atomistic-uv-cache uv run python scripts/benchmark_m5max_reference.py lammps --classify-only --json
UV_CACHE_DIR=/tmp/mlx-atomistic-uv-cache uv run python scripts/benchmark_m5max_reference.py lammps --json
UV_CACHE_DIR=/tmp/mlx-atomistic-uv-cache uv run python scripts/benchmark_m5max_reference.py validate --manifest results/m5max-reference/manifest.json --json

Use a host terminal or approved outside-sandbox command for final GPU/OpenCL measurements. The sandbox can hide device access and should only be used for dry classification or schema tests.