GPU: NVIDIA A100-SXM4-40GB (NERSC Perlmutter, one model per card). Structure: the La-3OH-128H2O nonperiodic droplet, 391 atoms, charge 0, spin multiplicity 1. Timing: four energy-and-force evaluations after the model is loaded. The first call is compile or warmup. The last step is the fourth call. The median is the median of the last three calls. The 150 ps column uses a 0.5 fs timestep and 300000 steps, scaled from the last step only. Measured on 2026-09-25; DPA4 Nano–Pro added 2026-09-30 on A100-SXM4-80GB (other models on A100-SXM4-40GB). 94 models succeeded. uma-s-1 does not load on fairchem 2.23.
Rows are ordered by inference time. The number is the last-step time in milliseconds on an A100-40GB, so a row higher on the chart is the cheaper model. The horizontal axis is the total energy (eV), from −273306.4 to −273301.3. The lowest energy is DPA4C-Mini, -273305.999 eV. Hover a point for the exact energy and the difference from DPA4C-Mini.
Top to bottom is the last-step time (ms). The horizontal axis is total energy (eV). Each segment runs from DPA4C-Mini to that model.
| Family | Environment | Software | Inference settings |
|---|---|---|---|
| DPA | dpa-pt211 + pytorch/2.11.0 | deepmd-kit 3.2.0 | DPA4C: DP_CUDA_INFER=2, DP_TF32_INFER=1, frozen.pt2; DPA4: + DP_TRITON_INFER=2; charge_spin=[0,1] |
| GRACE | grace-tf215 + tensorflow/2.15.0 | tensorpotential 0.6.1 | TPCalculator mode=uniform |
| Orb | orb-speed + pytorch/2.13.0 | orb-models 0.7.0 | precision=float32-high, torch.compile |
| UMA | pytorch/2.13.0 user site | fairchem-core 2.23.0 | inference_settings=turbo, task=omol, TF32 |
| MACE | mace-speed + pytorch/2.13.0 | mace-torch 0.3.16, cuequivariance 0.12.0 | cuEq + reduce-overhead; compile is off for POLAR and omol XL |
The environments are under NERSC ~/software/envs. MACE sets PYTHONNOUSERSITE=1. A compiled MACE median can be lower than the last step. For POLAR the first two steady calls are still warmup, so use the last step. Energies near −273303 eV are the OMOL surface. Energies near −1880 eV are a materials surface. This GRACE weight set has no OMol model.
| Family | Model | Last step (ms) | Median of the last three (ms) | First call (s) | 150 ps at 0.5 fs (h) | Energy (eV) | Surface | Inference settings |
|---|