Amber is a suite of biomolecular simulation programs. Amber developers have been working on GPUs for more than a decade, and a set of benchmarks have been produced for a number of graphics cards as shown in the tables at the links:
We executed the Amber22 benchmarks on the VMs of CloudVeneto hosting various models of GPUs. The VMs had CentOS Stream release 8 as operating system, the NVIDIA Driver Version 520.61.05, the CUDA Version 11.8 and the flavors available in CloudVeneto (VCPUs ranging from 4 to 36 and RAM from 20 to 360 GB).
Amber22 benchmarks run on docker containers whose image was build from Ubuntu 20.04 and CUDA Version 11.6.2.
The table below shows the values of the benchmarks for the CloudVeneto GPUs, directly comparable with the ones published at ambermd.org and exxactcorp.com.
The first row shows the GPU architecture, the second row the Floating Point single-precision (FP32) performance in TeraFlops (TF) as declared by NVIDIA specifications, the following rows shows the 12 benchmarks in the same order as published at exxactcorp.com.
The 6 GPU models tested were hosted on servers with different Intel Xeon CPU models. The table below shows the same benchmarks when executed in parallel on 8 VCPUs and without GPU.
The first row here shows the HEPSPEC06 benchmark of the server hosting the CPUs divided by the number of threads (all CloudVeneto servers have Hyper-Threading enabled). The following rows shows the 12 benchmarks and can be compared with the 8X CPU row of the tables at ambermd.org.
