CPU memory requirements

Is there a rough formula for CPU memory requirement vs. box size?

In release notes and forum posts, I’ve seen that a 640px homogeneous refinement should fit on <11GB of GPU memory and <55GB of CPU memory, while a 1024px refinement would also use <11GB of GPU and ~160GB of CPU memory.

I had trouble running refinements on a 768px box using 11GB of GPU RAM (GTX 1080 Ti) and 128 GB of system RAM (succeeds without CTF refinement, fails with defocus refinement). I’m trying 600px instead, but I’d like to know what limitations to expect.

Update 1 - 600px also fails if CTF refinement is on, with skcuda.cufft.cufftAllocFailed.
Update 2 - Same issue with local CTF refinement alone - what are the memory requirements for local CTF refinement?

1 Like

It seems I can get the jobs to complete by manually setting the GPU batch size to a smaller value. How are the batch sizes computed? I’m getting a batch size of 1000 for the 600px box (11GB 1080Ti, no other processes running) which crashes, but a manual size of 512 is OK.

1 Like