Table 3.
Comparison of molecular dynamics simulation speeds for various NN models, DPA-2-Drug, MACE-OFF23 (small, medium, and large configurations), and AIMNet2, across molecules with atom counts ranging from 50 to 1,490. The semi-empirical method GFN2-xTB, tested using different numbers of CPU cores, is included for comparison. All simulations were conducted using the atomic simulation environment [28]. Computation times per step (in milliseconds) represent averages from 3 independent trajectories, each comprising 5,000 time steps.
| # Atoms | Time per MD step (ms) | ||||||
|---|---|---|---|---|---|---|---|
| DPA-2-Drug a | MACE-OFF23 a | AIMNet2 a | GFN2-xTB b | ||||
| Small (S) | Medium (M) | Large (L) | 4 cores | 8 cores | |||
| 50 | 51.9 | 29.1 | 33.9 | 42.7 | 18.9 | 54.4 | 50.1 |
| 60 | 49.9 | 28.7 | 35.1 | 46.4 | 19.7 | 80.6 | 71.1 |
| 100 | 50.0 | 29.0 | 34.8 | 58.9 | 21.5 | 187.4 | 158.9 |
| 148 | 51.2 | 30.6 | 38.9 | 83.9 | 19.4 | 403.2 | 324.7 |
| 375 | 53.4 | 79.1 | 105.3 | 233.8 | 23.9 | - | - |
| 745 | 62.8 | 107.4 | 163.0 | 409.0 | 24.9 | - | - |
| 1,490 | 104.6 | 251.8 | 367.7 | 926.9 | 59.9 | - | - |
a
Simulations were run on an NVIDIA A800 GPU.
b
Simulations were run on an Intel Xeon Gold 6240R CPU.