Here in our lab we have observed cryoSPARC v2 to be noticeably slower than v1 in refinement jobs across several types of datasets, running on the same machines we used for v1.
In particular, we notice it takes a long time in the “Computing FSCs” step.
Is this expected behavior, or any hints to what the problem could be?
Thanks for reporting - how much slowdown are you seeing? The FSC code is slightly different than in v0 but shouldn’t be much slower. And is this in the standard homogeneous refinement job type, or nonuniform refinement?
Hi @apunjani, thanks for answering. To put some numbers to it:
For a dataset with 42k particles, box size 360, the final “Computing FSCs” takes ~440s for standard homogeneous refinement and ~820s for NU-refinement. We are running on GTX-1080 Ti cards. Is this reasonable?
Actually the FSC computation is not GPU accelerated at the moment. But for a 360 box size, the final FSC calculation (which also does automatic mask adjustment) taking 440s is not unreasonable. We have also identified this as now the bottleneck in refinements so we will be accelerating this on the GPU soon.
OK, good to know. Thanks for replying and looking forward to the updates!
As of cryoSPARC v3.3.1 released December 1, 2021, FSC computation is now GPU accelerated, which should yield substantial speedups. Please see https://cryosparc.com/updates