Hi all,
We are trying to spec a major extension for our “boutique” cluster and we have been given very decent prices on boxes featuring four RTX 5000 ADA cards each. We have been trying to look at Gromacs benchmarks for the new architecture, but there is nothing for Gromacs. There is however this for AMBER. Could someone comment on whether ADA would yield similar performance with Gromacs? All other comments are very welcome, especially from Szilard.
Many thanks!
I would expect similar relative performance in GROMACS. These look to have very similar cores to the previous generation.
Hi Sasha,
I don’t have data at hand, but as Berk noted, I would expect similar behavior as AMBER. Clocks are higher on Ada I think so it is likely that you could get better performance for smaller simulation systems that don’t really saturate a large GPU (e.g. a few tens of thousands of atoms on a 4070Ti vs 3080/3090).
You can look for NVIDIA L40 benchmarks there may be some data out there (note that those have more functional units and slightly higher TDP).
Cheers,
Szilárd
PS there are some benchmarks here (which luckily also include ns/day not just relative perf):
Just to be clear, the data on that site was gathered independently by NVIDIA I can not vouch for it personally (but I have no reason to believe that it is flawed in any way).
@hess note that the the Ada cores (as well as Hopper, so all 8.6, 8.9, 9.0 devices) have 2x FP32 throughput, 128 FMA ops/clock vs previous gen which have only 64 FMA/clock.
Thank you both so very much. I have convinced my managers to always use your opinion as a guiding principle for our upgrades! :)