We just got a grant approved and it comes with a decently beefy budget to build a new machine.
For our purposes, we mostly run enhanced sampling heuristics with lots of short simulations running in parallel. This creates a little confusion when deciding how to build the system (4 high end GPUs vs 8 weaker ones etc).
I was wondering if there’s any sort of GROMACS benchmark using different GPUs in terms in ns/hour in a basic system, so that I can better spend this budget.
My intuition is that v100s won’t be 4x faster than RTX 3090s to justify the 4x increased price tag, but I might be mistaken.
I was able to find information for older cards but the architecture changed so much lately i wanted to ask again just in case.
Many cloud providers will have GPU hardware, including recent hardware.
You may wish to run a few benchmark tests for system sizes of interest
to you. Communication between GPUs may be more difficult to evaluate in
the commercial cloud, but if this not important, then you are probably
ok.
Probably also relevant is whether you need double precision performance
and if you want to tune the code for it to run efficiently for your use
case.
Finally, may wish to consider how much data you will need to store and
who will need to access the data.
Not sure how much do the V100s cost these days (since the A100 has superseded it), but it is unlikely that in terms of ns/day/cost professional will be more cost-effective than RTX cards.
Note however that depending on your budget you may want actually go to 6x or 8x slightly lower-end GPUs.
Also make sure that the rest of the system is up to the task including power supply and cooling as these cards are quite hungry.
Unless you use an enhanced sampling algorithm which is not compatible with GPU-resident steps in GROMACS, you don’t need worry about bus or GPU-to-GPU communication.