MLIP benchmarks
Atomli 0.1.4 · 1 host CPU thread
Nequix MP-1 PFT
Caffeine
24 atoms · energy and forces
| Hardware | Time | Peak RSS |
|---|---|---|
| Apple M4 Max · CPU | 1.7× faster 7.78 msmlip.rs 12.8 msNequix / JAX | 78% less memory 172 MiBmlip.rs 793 MiBNequix / JAX |
| Apple M3 Ultra · CPU | 1.3× faster 9.17 msmlip.rs 12.0 msNequix / JAX | 81% less memory 173 MiBmlip.rs 929 MiBNequix / JAX |
| AMD EPYC 7702P · CPU | 4× faster 21.5 msmlip.rs 85.7 msNequix / JAX | 85% less memory 177 MiBmlip.rs 1.13 GiBNequix / JAX |
| NVIDIA A100 | 2.7× faster 3.28 msmlip.rs 9.01 msNequix / JAX | 77% less memory 398 MiBmlip.rs 1.69 GiBNequix / JAX |
| Apple M4 Max · Metal | 3.96 msmlip.rs | 239 MiBmlip.rs |
| Apple M3 Ultra · Metal | 6.10 msmlip.rs | 244 MiBmlip.rs |
TiO₂ anatase
6 atoms · energy, forces, and stress
| Hardware | Time | Peak RSS |
|---|---|---|
| Apple M4 Max · CPU | 1.8× faster 6.51 msmlip.rs 11.8 msNequix / JAX | 78% less memory 172 MiBmlip.rs 795 MiBNequix / JAX |
| Apple M3 Ultra · CPU | 1.4× faster 7.92 msmlip.rs 11.1 msNequix / JAX | 81% less memory 171 MiBmlip.rs 918 MiBNequix / JAX |
| AMD EPYC 7702P · CPU | 4.6× faster 17.4 msmlip.rs 79.9 msNequix / JAX | 85% less memory 177 MiBmlip.rs 1.15 GiBNequix / JAX |
| NVIDIA A100 | 4.3× faster 3.62 msmlip.rs 15.6 msNequix / JAX | 76% less memory 395 MiBmlip.rs 1.62 GiBNequix / JAX |
| Apple M4 Max · Metal | 4.43 msmlip.rs | 241 MiBmlip.rs |
| Apple M3 Ultra · Metal | 7.33 msmlip.rs | 241 MiBmlip.rs |
Equiformer V3
Caffeine
24 atoms · energy and forces
| Hardware | Time | Peak RSS |
|---|---|---|
| Apple M4 Max · CPU | 2.1× faster 54.5 msmlip.rs 115 msPyTorch | 72% less memory 307 MiBmlip.rs 1.08 GiBPyTorch |
| Apple M3 Ultra · CPU | 64.6 msmlip.rs | 328 MiBmlip.rs |
| AMD EPYC 7702P · CPU | 589 msmlip.rs | 307 MiBmlip.rs |
Methods
Each pair uses identical weights and geometry. Equiformer uses OMat24 direct weights. Times include neighbor graphs and result transfer to the CPU. Timing starts after model loading and one warm-up. Time is the median of three evaluations.
Nequix uses its JAX calculator. Equiformer uses upstream PyTorch layers with a molecular graph adapter that evaluates energy and forces, excluding the unused stress head. The Equiformer PyTorch comparison is available on M4 Max only.