Initially it is myself, but the ultimate goal is for others to also be able to use it. For now, it is unlikely that a single job would be multinode. This may happen in the far future, but alot of things have to work out first :) — Wesley May 19 at 19:52

Do you already have selected the code to run? — Camps ♦ May 19 at 21:19

+1. Nice first question, welcome to our new community and thank you so much for contributing your question here! We hope to see much more of you in the future !!! — Nike Dattani May 19 at 23:14

Not really enough for an answer, and rather generic advice, but get the fastest RAM you can afford. Even simulation and modeling software that is not particularly memory hungry still accesses RAM a lot, and memory bandwidth is often overlooked as a limiting factor for such workloads. — Austin Hemmelgarn May 20 at 12:22

How much RAM per job do you expect? Since you basically say you run 8-32 jobs, lets assume each one with 32GB you need 256GB-1TB? — lalala May 20 at 17:58

@lalala great question, don't know offhand, but I would be running DFT at most, so that will only be a couple GB. I won't be running MP2 or higher on anything big. The MD is very light. — Wesley May 20 at 18:08

Many of the specifics are very much dependent on the size of systems / code. Also, if most of your calculations are small (6-8 cores), using a 128 cores workstation may save shelf space but not necessarily the most economical solution. If you do lot of MD, you may be better having a smaller workstation with graphics card just for MD, and keep it separate from DFT calculations. — Greg May 20 at 18:09

@greg Aye, I will be DFT heavy for the first 4-6 months, then possibly switching to being MD heavy. — Wesley May 20 at 18:43

If you are running any classical MD at all (Gromacs), you will want a GPU. For a few hundred dollars, you can get a lower-end GPU or one a few years old that will make classical MD an order of magnitude faster. — WaterMolecule May 20 at 18:57

I agree totally with @Greg. A 128-core machine is extremely "niche" hardware and therefore very expensive since only maybe 0.001% of the population will buy it. 4 machines with 32 cores each might be cheaper. It's just like RAM: one stick with 128GB can be more than the equivalent amount of RAM using 8GB sticks, though you then run into the problem of whether or not you have "space" to put them. — Nike Dattani May 21 at 1:37

@Wesley As Nike says, 128 cores (double AMD CPU I assume) are both expensive and most probably out of stock (so are other parts like large memory sticks needed for a 128 cores beast). The other practical part is that 3-4 workstations are more flexible (can upgrade one with GPU or more storage, mem ..), and less stress if one of them needs service. If you will share this with others (as you say in OP), you will also feel with people who just take all the computational capacity themselves - having separate machines can help on this, too. — Greg 2 days ago

Maybe you can find a cloud service that allows you to test different hardware setups without investing too much money. That would give you more data for your decision. — Gregor Michalicek yesterday