GROMACS version:
GROMACS modification: Yes/No
Here post your question
I’m looking to purchase a workstation specifically for running GROMACS, but I am new to GROMACS. Is the following configuration balanced?
Any insufficiency or waste? Also, is RTX5880 Ada or RTX4090 better here?
Thanks in advance.
Jesse
Hi Jesse,
if you are looking to purchase a workstation mainly or exclusively for running GROMACS MDsimulations I would personally advice against using Threadripper unless you are forced into the platform for other reasons.
ryzen 9950x3d (if you really need the 16 cores/32 threads) would be “better“ bang for you buck and more than enough (since almost all calculations would be offloaded to the GPU anyway) although it is a pricey chip still.
in terms of graphics card it depends wether you want raw performance or if you are contraints in terms of power those two cards have vastly different TDPs 4090 being 450W and the 5880 285W
as a rule of thumb more cuda core count for nvida cards in GROMACS is better not taking into account tensor cores or others because GROMACS dont use them at all (to my knowledge)
following that line of thought pure performance the 4090 I beleive would be better, VRAM is not really a contraint in simulations as it is in AI workloads so most of the 48gb of the 5880 would stay unused. (again rule of thumb)
also although it is possible to use multple gpu setups unless you have a system with MILLIONS of atoms I wouldnt necesseraly build a system with multiple gpus just for GROMACS since it doesnt scale the best across gpus (even with nvlink) as one would expect
depending on how long the simulations you are going to run are and how frequently you are saving to the trajectory files maybe 32gb of ram might be just cutting it (I have personally had analysis that required at least 64gb to load the trajectories) but there are workarounds for that
wish you all the best with the build
Klaus
I completely agree with @Klubinus. As long as only one GPU is available, expensive chipsets like the Threadripper aren’t worth their cost. More standard 16-cores CPUs like AMD Ryzen 9 9950X or the 7950X can handle very well the workload and provide quite some power in analysis. GROMACS doesn’t need much RAM, but 64GB are suggested for analysis purposes (and also nowadays RAM is cheap enough that staying with 32GB is a useless risk).
GPU-wise, in our tests the RTX5090 performs incredibly well, and the RTX4090 does the same. The 5090 is stupidly expensive, but the 4090 is now out of production and very hard to find (and expensive as well). My take would be to try to put your hands on one of the two, the cheapest in case you (very luckily) find both. Consider that most of the VRAM is not utilized, so hardware like the 5880Ada is really not worth it. This might change if you know that you want to use the machine to do AI/ML as well, although both the 4090 and 5090 perform great, too, but that is problem dependent as you might want the memory of the 5880. Motherboards like the ProArt X870E work well, and should be cheaper than your options. Just be sure that the power supplier is sufficient, as these machines are quite demanding, especially if you go for a RTX5090.
In general, we observe that 16 good cores per GPU are more than sufficient most of the times for all the single simulations we run. Things change if you want to go towards multi-replica techniques (e.g. flavours of HREX, REST2, etc.) and you plan on running multiple simulations at the same time. In that case the performances are largely dependent on the number of replicas and how big the systems are, and having multiple GPUs tends to become more important than having a powerful but single one; the same holds true for CPUs, as having higher counts means more CPUs available per replica.
Dear Klaus,
Thank you very much for the suggestions. I will switch the 7955wx with 9950X, and replace the RTX5880ada with 5090. Now I understand that 7955wx is for server rather than personal use.
Regarding the RAM, I meant two 32G memory modules by “32G DDR5 *2", giving a total of 64G. Do you mean I need two 64G modules to obtain a total 128G RAM?
Thanks again for your time and advice.
Jesse
Dear obZehn,
Many thanks for your advice. It has been incredibly helpful. Now I have no question ragarding the CPU and GPU. You saved me a lot of money.
Just one quick question on the RAM, when you mentioning 64GB might be the safe choice, did you mean I should go with a pair of 64GB memory modules? Since I have already listed two 32G modules for a total of 64GB.
Best wishes.
Jesse
Glad to help :)
Sorry, I didn’t see the *2 in your original post. I would say that 64GB are sufficient for most cases, but if you expect very heavy analysis than 128GB are not going to hurt, given that the cost is reasonable nowadays.
Before spending more in RAM however be sure to have storage space. 2TB of SDD are okay for software/OS etc, but a secondary hard disk with > 4 TB is highly suggested if you need to storage simulations and have no backup somewhere else, but again here the amount of space you need depends on the simulations you intend to run!
Dear Jesse,
As @obZehn I also didnt see the *2 in the RAM configuration 64GB in dual channel is more than enough for all your needs (unless you have particularly heavy trajectories files) since RAM is quite cheap in comparison but at that point I also agree with @obZehn to look for some high capacity (even if slow) storage option
here is why:
when running simulations you usually end up with 2-10gb worth of trr and xtc trajectory files for 1 standard trajectory without analysis so you would likely want to have multiple trajectories with different initial velocities for analysis (lets say 10) you would quickly have 20-100gb worth of raw data for 1 group (you would have to consider if thats enough) and again only speaking about the trajectory files it quickly grows up from there once you want to evaluate multiple ligands for example or multiple mutations (just as examples)
I hope that helps you :)
Dear obZehn,
Thanks again for your help. I am planning to opt for 2TB SSD for the OS and 8TB SSD for storage.
Regarding RAM, I will likely get four 32GB DDR5 modules to yield a total of 128GB, assuming the price is acceptable.
Wishing you seamless progress with your simulations.
Dear Klaus,
Thanks for the detailed explanation. I would probably get 2TD SSD for the OS, and 8TB SSD for general storage.
32GB *4 RAM might be my choice, if a reasonable price is given.
Hope your simulations run smoothly.