Freesurfers,
I'd like to add my two about a recent experiment I ran with recon-all -use-gpu. (I'm using v5.1)
I took 6 T1 scans, imported them into two subjects a piece (gpu_SUBJID, cpu_SUBJID) and ran 'recon-all -all -use-gpu' on the gpu_ subject and the normal 'recon-all -all' on the cpu subjects. All gpu subjects ran on a single GTX 480.
On average, I saw about a 16% speedup using GPU hardware, which translates to about 4 hours on the hardware I'm using.
Qualitatively, the final inflated surfaces and {w,g}m boundaries are very similar.
Two questions:
1) Would you recommend -use-gpu for "production" analyses using FS v5.1?
2) Separately from the above experiment, I ran one gpu subject using 4 GTX 480s and saw no significant time difference. Is this to be expected, ie your CUDA code doesn't scale across multiple GPUs? Just curious, certainly not a deal-breaker if it doesn't.
Thanks for your input.
Scott Burns Neuroimaging Analyst Education and Brain Science Lab Kennedy Center Vanderbilt University
On Fri, Apr 20, 2012 at 12:52 PM, Burns, Scott S scott.s.burns@vanderbilt.edu wrote:
- Separately from the above experiment, I ran one gpu subject using 4 GTX 480s and saw no significant time difference. Is this to be expected, ie your CUDA code doesn't scale across multiple GPUs? Just curious, certainly not a deal-breaker if it doesn't.
Nothing I added (primarily mri_ca_register and mri_em_register) makes use of multiple GPUs, any more than Freesurfer makes use of multiple cores. I don't know about any additions since. If you have more than one GPU, you can run more than one subject at once (if you point each run to a different GPU). Also, make sure you're using the Fermi binaries on that GTX 480.
Richard
freesurfer@nmr.mgh.harvard.edu
-
Burns, Scott S -
R Edgar