Antares, out latest generation of AMD-based cluster, has come online adding 1152-cores to bramble’s computing capacity.
We’ve been really impressed with the pace of this machine, so we wanted to share some benchmarking data with you.
Antares is built around second generation AMD EPYC chips and FDR Infiniband, all contained within Gigabyte hyper-converged chassis. Antares itself is built from four 2U chassis, each containing 128-cores and 1Tb RAM.
Our clusters have been designed to balance performance and cost, with prices starting around £25,000 for 128-cores and going to £183,000 for a full 1152-cores. (Prices valid 3rd March 2021 and include one-year’s warranty).
Many of the simulations that are run through bramble are full car models running steady-state RANS or transient DES solves. These models vary in size depending on the amount of detail but are typically between 80m and 150m cells.
For these benchmarks we’ll be comparing Antares back to our first generation AMD EPYC cluster, Hyperion. We’ve generated two versions of the standard OpenFOAM® motorbike test cases: one at 82m cells and the other at 148m cells. These have been meshed in snappyHexMesh and solved in simpleFoam (RANS) and pimpleFoam (DES) with both solvers being run single and double precision.
And now for the interesting part… the actual benchmarking data starting with the meshing times. Cache size and memory access are real bottlenecks for snappyHexMesh, but thankfully, this is where Antares has made big strides forward over the older Hyperion.
Meshing times are reduced by a massive 20% across the different core-counts and model sizes.
The next two tables compare average iteration speeds for the simpleFoam and pimpleFoam solves in single precision (SP) and double precision (DP) respectively.
Again Antares shows its speed with 4 to 20% reductions in solve time seen depending on core-count and precision.