MOLECULAR DYNAMICS PERFORMANCE GUIDE - Digital Research Alliance of CANADA

ADVANCED SEARCH

APPLIED FILTERS: OPENMM _____ _____ _____ _____ _____ _____ _____ _____

  • EXPLORING THE DATABASE

    Default view
    When this page is viewed no filters are initially applied. All benchmarks are selected and sorted by simulation speed. The chart on the right displays only top 30 benchmarks for clarity.
  • Selecting benchmarks
    A subset of benchmarks can be selected using a custom chain of filters. Selected database entries can be downloaded as CSV files for further analysis or viewed in the Benchmark Details table at the bottom of the page.
  • Detailed views
    A detailed view of each database entry can be accessed from Benchmark ID and Software ID search forms. Detailed views include submission commands and simulation input files. View example: PMEMD @Narval (benchmark ID=46).
  • Parallel efficiency
    Efficiency is computed as PS/(SS * N) where PS is speed of the parallel program, SS is speed of the serial program, and N is the number of CPUs or GPUs.
  • Viewing parallel speedup and efficiency
    To view the graph of the dependence of parallel speedup and efficiency on the number of CPU/GPU equivalents select only one software and one cluster. View example: GROMACS @Narval .

  • Viewing QM/MM benchmarks
    To view QM/MM benchmarks select simulation system 4cg1 .

Performance Chart For Selected Benchmarks

*Data updated March 14, 2026

Cost Of GPU-accelerated Simulations

*Data updated March 14, 2026

  • OPTIMIZING GPU USAGE

    Parallel scaling to multiple GPUs
    Parallel scaling to multiple GPUs strongly depends on the compibation of software, hardware and simulation parameters. Often simulations do not run faster on multiple GPUs (PMEMD @Cedar example). Simulations on nodes with direct interconnect between GPUs (NVLink) are more likely to benefit from multiple GPUs, but efficiency decreases and cost goes up with the number of GPUs (NAMD3 @Cedar example ).
  • Benchmarking GPU accelerated MD Engines
    For benchmarking we use the optimal number of cores per GPU (the number needed for the fastest simulation time but not exceeding the maximum number of CPU cores per GPU in a GPU equivalent).

BENCHMARK RESULTS

CPUY: CPU years per 1 microsecond long simulation. GPUY: GPU years per 1 microsecond long simulation. | T: tasks | C: cores | N: nodes. Speed is in ns/day. Integration step = 1 fs. Measured with dataset 6n40 (239,131 atoms).

*More information is available by clicking ID in the table above
ID Software Module Toolch Arch Data Speed CPU CPUeff CPUY GPUY T C N GPU NVLink Site
139 OPENMM.cuda openmm/7.7.0 gofbc avx2 6n4o 1.26e+01 Xeon E5-2650 61.4 0.0 0.435 1 2 1 2P100-PCIE No Cedar
216 OPENMM.cuda openmm/8.0.0 gofbc avx2 6n4o 1.25e+01 Xeon E5-2650 39.3 0.0 0.656 1 3 1 3P100-PCIE No Cedar
142 OPENMM.cuda openmm/7.7.0 gofbc avx2 6n4o 1.24e+01 Xeon E5-2650 40.4 0.0 0.661 1 3 1 3P100-PCIE No Cedar
304 OPENMM.cuda openmm/8.4.0 gofbc avx512 9naw 1.21e+01 Xeon Gold 6448Y 74.9 0.0 0.452 1 2 1 2H100-HBM3 Yes Rorqual
314 OPENMM.ROCm container/8.5.0 rocm avx512 9naw 1.18e+01 AMD Instinct Accelerator MI300A 65.7 0.0 0.463 2 24 1 2MI300A No Nibi
217 OPENMM.cuda openmm/8.0.0 gofbc avx2 6n4o 1.15e+01 Xeon E5-2650 27.1 0.0 0.954 1 4 1 4P100-PCIE No Cedar
140 OPENMM.cuda openmm/7.7.0 gofbc avx2 6n4o 1.14e+01 Xeon E5-2650 27.8 0.0 0.96 1 4 1 4P100-PCIE No Cedar
214 OPENMM.cuda openmm/8.0.0 gofbc avx2 6n4o 1.06e+01 Xeon E5-2650 100.0 0.0 0.258 1 1 1 1P100-PCIE No Cedar
136 OPENMM.cuda openmm/7.7.0 gofbc avx2 6n4o 1.03e+01 Xeon E5-2650 100.0 0.0 0.267 1 1 1 1P100-PCIE No Cedar
313 OPENMM.ROCm container/8.5.0 rocm avx512 9naw 9.00e+00 AMD Instinct Accelerator MI300A 100.0 0.0 0.304 1 24 1 1MI300A No Nibi
Date Updated: March 14, 2026, 2:13 p.m.