You are here

Help on building a PC to run rosetta

3 posts / 0 new
Last post
Help on building a PC to run rosetta
#1

Dear all,

I'm currently in conversation with my PI to get a PC mainly for protein modeling and MD simulations.

One of the software for protein modeling I want to use is rosetta, but unfortunately, I don't know what component I should focus on. RAM/CPU/GPU? all three of them?

I hope someone can help me.

Thanks in advance.

Best regards,

Carlos

 

ps- I don't have a specific budget yet, so feel free to suggest to me the best configuration you can think of.

Category: 
Post Situation: 
Thu, 2022-02-10 08:37
cnavarro

Hi!

I'm commenting mostly because I need to do the same and PC parts are also not my forte. I can give general guidelines that I'll be considering in an upcoming build...

  • More processors. Most protocols require multiple models to be built and this is best accomplished by utilizing multiple processors. It's convenient for local process development to have at least 8 but a greater number may obviate the need for a cluster (if you have a bit of patience to complete the task).
  • RAM usage is kinda high. A general rule of thumb is to have 2GB per instance of Rosetta running. This is probably on the max side of things as memory usage has improved. Utilizing MPI parallelization helps with this a lot.
  • Large Solid State Drive. This is less important for Rosetta use than for pipelines that use very large databases, such as RoseTTAFold's or AlphaFold's dependence on HHBlits and very large sequence databases. Multiple processors will help speed up these searches, too. (Note that there are also servers to provide these searches but access can be restricted to academic use or the sequence may be proprietary in nature. Application in an industry environment drives my need here.)
  • A moderate GPU. The base Rosetta protocols don't require a GPU (but might make use of one soon for certain tasks) but this could be useful in the Rosetta-DL pipelines. I'm unlikely to be training any DL models for the time being but, as mentioned, it would be useful still for model inference tasks. Having a local GPU is mostly for convenience - if I need something more I can use AWS/Colab/etc.

 

Hope this helps. Please share what you end up buying!

 

 

Thu, 2022-02-10 09:40
nannemdp

Rosetta minimums out at about 2GB per process when using it in MPI mode (and caps far higher depending on protein size) - RAM needs scale very strongly with protein size.  You can go a lot lower with multithreading but only a few applications use multithreading.  Almost nothing in base Rosetta uses the GPU, but the RosettaDL stuff mostly does use GPU.  

Unfortunately "Rosetta" is not a sufficient specification to know what compute resources you should use.  I've run jobs that used 1 processor and 200 GB of RAM; I've run jobs that used 10000 processors at once.  Some jobs can run on your laptop or even your phone, some should be run on dedicated supercomputers.  I'd recommend getting something big enough to run "test scale" jobs to make sure you are doing the experiment you mean to, and then farming out the real work to academic super compute time acquired via grants.

Thu, 2022-02-10 10:12
smlewis