Ram per CPU recommendation

Dear AFNI experts:
I am designing a new neuroimaging cluster and wanted to see if there is a recommended RAM to core recommendation for AFNI. Generally I have purchased systems with 4GB ram per core but the current core count on this new system is about 52-64 cores per node (hyper threading disabled) so I am wanted to get more feedback as that can be a pricy endeavor. Is there a recommended minimum amount of ram/core to optimally run AFNI openmp versions?

Additionally, while I plan on using a docker to run AFNI, I am wondering if there is a strong potential for differing results between AMD based systems and Intel based systems. One of the existing clusters runs on Intel based processors (6230R) and so I wanted to make sure there are no issues in terms of reproducibility if we went with AMD vs Intel for this new cluster.

Thanks in advance for your help.


Hi Ajay,

Actually, 4 GB is often not enough, but it really depends on the resolution of the researchers’s data and the number of time points that they are analyzing. Non-linear registration will often need a bit more RAM, but the RAM usage for the linear regression step of model fitting the time series data varies wildly, and depends heavily on the input data. Sometimes 4 GB is enough, often not.

One thing that is done here is to have some nodes with 4 GB RAM, and some with more. But now the system is getting complicated. Having a separate system with more RAM is also a nice option.

Using different CPUs can have a small effect on the results. One would not expect binary equivalence when jumping between systems, but the results should certainly be close.

Does that seem reasonable?

  • rick

Hi Rick,
Thanks for the feedback. It makes sense regarding no major differences between AMD vs Intel.

In terms of RAM my question it was not 4GB total per node but 4GB/physical core/node. Since a dual socket node with 52 cores (intel 6230R x2) that would be 208 GB ram which is doable if we need 4GB/core/node. If I go with AMD which is 128 cores (64 cores x2 cpus) this would be 512GB which is much more expensive. I wanted to get a sense if AFNI has any hard or soft rules when it comes to ram per core since we would launch several jobs in parallel so I wouldn’t be using all of the ram to run a single subject. I hope this clarifies my question a bit more.