AMD or NVIDIA gpu, which is faster at same price level?

[UltraJessi]

GROMACS version: 2020.3
GROMACS modification: Yes: AMD-ROCM version: GitHub - Biu-G/gromacs-rocm: Gromacs that can be accelerated at AMD GPU
it seems AMD’s Instinct series are designed partically for HPC not AI, so the speedup aces nvidia, anyone got it tested?
By the way, at which condition will Gromacsd run faster (assuming same total price / same cpu)?: single expensive GPU(e.g MI100X1 / A100X1) versus multiple cheap GPUs(e.g V100SX2 / P40X4 /…)

[pszilard]

@UltraJessi

GROMACS modification: Yes: AMD-ROCM version: GitHub - Biu-G/gromacs-rocm: Gromacs that can be accelerated at AMD GPU
it seems AMD’s Instinct series are designed partically for HPC not AI, so the speedup aces nvidia, anyone got it tested?

Not sure what that code does, but the V100 performance is likely off (hence the extrapolated A100 number as well) as those results do not match my experience; the recent data we published shows the main performance-dominating kernel being >2x slower on an Mi50 than a V100 (see Fig 7 of https://aip.scitation.org/doi/pdf/10.1063/5.0018516) and the gap in total simulation performance is similar.

@UltraJessi

By the way, at which condition will Gromacsd run faster (assuming same total price / same cpu)?: single expensive GPU(e.g MI100X1 / A100X1) versus multiple cheap GPUs(e.g V100SX2 / P40X4 /…)

Depends on the simulation system, settings, single vs multi-trajectory, and hardware (also, the V100 is “cheap”?). Look at Figs 10-12 in https://aip.scitation.org/doi/pdf/10.1063/5.0018516, that should give a rough idea about the performance (of the current release).

Cheers,
Szilárd