January 2021

The benchmark runs below used the Amber 20 Benchmark suite, downloadable from here.

Older benchmarks: [Amber 18] Amber 16] [Amber 14]

Hardware:

• CPU runs: 28 CPUs, Intel(R)_Xeon(R)_CPU_E5-2680_v4_@_2.40GHz
• k80: Intel(R)_Xeon(R)_CPU_E5-2680_v4_@_2.40GHz, Tesla K80 -- benchmark runs with 1-4 GPUs on single node
• p100: Intel(R)_Xeon(R)_CPU_E5-2680_v4_@_2.40GHz, Tesla_P100-PCIE-16GB -- benchmark runs with 1-4 GPUs on single node
• v100: Intel(R)_Xeon(R)_CPU_E5-2680_v4_@_2.40GHz, Tesla_V100-PCIE-16GB -- benchmark runs with 1-4 GPUs on single node
• v100x: Intel(R)_Xeon(R)_Gold_6140_CPU_@_2.30GHz, Tesla_V100-SXM2-32GB -- benchmark runs with 1-4 GPUs on single node

Amber 20 with all patches as of Nov 2020.
CPU runs: Amber 20 built with Intel 2020.0.166 compilers, CUDA 10.2, OpenMPI 4.0.4 (Biowulf module 'amber/20.intel')
GPU runs: Amber 20 built with gcc 7.4, CUDA 10.1, and OpenMPI 4.0.4 (Biowulf module 'amber/20-gpu')

Based on these benchmarks, there is a significant performance advantage to running Amber on the GPU nodes, especially the P100s and V100s, rather than on a CPU-only node.

Explicit Solvent

Implicit Solvent

The benchmark above was run on a Intel Xeon E5-2680 v4 @2.5GHz with 28 cores (56 hyperthreaded cores). As with other MD applications, the performance drops when more than 28 MPI processes (i.e. 1 process per physical core) are run. Thus, if running on CPUs, it is important to use the '--ntasks-per-core=1' flag when submitting the job, to ensure that only 1 MPI process is run on each physical core. If it is possible to run on a GPU node, you will get significantly better performance as shown in the benchmarks below.

Implicit Solvent (GB)

Explicit Solvent (PME)