Dear experts,
I am planning to purchase a GPU to accelerate the FSL’s eddy/bedpostx/probtrackx2 tasks on my personal PC. Do you have any recommendations for the GPU requirement to deal with a typical DWI data ( with 2mm isotropic resolution & 60-100 diffusion weighting directions)? Minimal information was found except for this post after a thorough web search. Thanks!
I did add some new data for FSL’s diffusion tools as well as brainchop AI inference. I originally wrote that page when we were evaluating what was then the latest hardware, and our hardware is no longer leading edge. I have included details for alllowing others to run the benchmarks on their own hardware, so perhaps we can compile feedback from others with more recent equipment.
1 Like
Honestly, I find that any Nvidia gaming GPU is good enough for eddy. Of course, the more cores, the better, but you don’t need to break the bank. Do aim for higher memory, as large DWI data sets can easily cause eddy to out of VRAM. My current GPU is a GeForce RTX3060 (12 GB memory) and I find it quite good for this purpose.
1 Like
@lconcha I agree that there are clear diminishing returns for faster processing, and the commodity gamer GPUs are just as fast as industrial GPUs for the FSL tools. Servers do allow you to run data from more individuals in parallel (though it would be cheaper to have a number of commodity computers) and industrial GPUs with huge RAM do aid AI model development (though have little impact on AI model inference for our field).
Would you be interested in contributing the performance of your hardware on the benchmarks?
Sure, I would be happy to help. We have a beowulf-style cluster, and some of the PCs have Nvidia GPUs, so I could run the benchmark tests on a few of them. Please provide the instructions.
@lconcha see the full instructions here. Briefly, assuming you have FSL 6.0.7.18 and brainchop-cli installed you can run:
git clone https://github.com/neurolabusc/CPUsForNeuroimaging
cd ./CPUsForNeuroimaging/bench
python dwi.py
python fmri.py
python ai.py