Hi everyone,
We would like to buy a workstation with one of the top CPU and GPU. We are considering to buy threadripper 3990x and one or two either RTX3090 or RTX3080. Could you please tell me if two GPUs will increase the simulation speed against to a single GPU? Also, would it be more profitable to run two simulation simultaneously with two GPUs?
Thanks,
Vlad
Hi.
32 cores is enough for two 3090, so you may be want to choose a processor with the highest possible clock rate - it is 3970X (if two GPUs is the maximum number you want). Two GPUs surely increase simulation speed, but how much it will speed up is strongly dependent on system you want to simulate (it must at least have > 200k atoms). You should read this forum and gather information about different cases.
Hi,
A 3990x is probably overkill, for typical simulation workloads you only need 2-4 CPU cores per GPU, though having more helps on some workloads that have tasks that can’t be offloaded to the GPU.
Yes, multiple simulations in parallel almost always has greater throughput than one simulation on multiple GPUs, unless (as @skyto alluded to) you have a giant system. In fact, you can sometimes benefit from running multiple simulations on the same GPU, particularly with older Gromacs versions.