Table 8. Performance of Various Accelerator Configurations to Run a Single Simulation of Dihydrofolate Reductase (DHFR)a.
| accelerator | engine | time scale (ns/day) |
|---|---|---|
| Anton 3 (64-node) (ASIC) | Custom139 | 212 200 |
| Anton 2 (512-node) (ASIC) | Custom140 | 85 800 |
| Intel Stratix 10 (FPGA) | Custom98 | 630 |
| 2x Nvidia Titan-RTX (GPU) | Amber154 | 629.03 |
| NVIDIA V100 SXM (GPU) | Amber154 | 522.20 |
| NVIDIA V100 PCIE (GPU) | Amber154 | 277.14 |
| NVIDIA TITAN X (GPU) | OpenMM155 | 393 |
| NVIDIA TITAN V (GPU) | OpenMM155 | 419 |
| NVIDIA RTX 3090 (GPU) | ACEMD156 | 1308 |
a
DHFR is a 159-residue protein (suspended in water) target for cancer therapeutics that has been used as a standard benchmark for MD simulation throughput. All simulations reported here employ NVE microcanonical constraints. NVE refers to the set of constraints on MD simulations in which moles (N), volume (V), and energy (E) are conserved in the simulation. All simulations reported here with the exception of Anton 2140 use the PME32 algorithm for non-bonded interactions. Anton 2 uses the μ-series142 algorithm for non-bonded interactions.