Support for box sizes above 1100 pixels

Features and Functionality
1

Hi @sarulthasan,

THP “sort of” helped - the timing isn’t so asymmetric any more. Haven’t run more than 2 NU Refines in parallel yet to really stress it for that (waiting on RBMC!)

Wanted to query about something else, though…

Per here, box sizes seem to top out around 1100 pixels. Can confirm with latest CryoSPARC… 1200 pixel box fails to both refine or to reconstruct - even on 48GB GPUs. 1150 fails too. 1100 works. Anything you can suggest to push into 1200/1400 pixel boxes? 1480 fits (just, need to disable Xorg if running) into 48GB in RELION, or is this a PyCUDA limitation or something like that?

Seems my hopes regarding being able to work with some of our bigger targets in CryoSPARC have been dashed. :frowning: Can do focussed refinements after symmetry expansion, but that opens up a whole new can of worms which I’ll make a new post for when I’ve got a bit more of the processing done (symmetry expanded icosahedral reconstructions of 800 pixel boxes is slooooow…)

1 Like
2

Hi @rbs_sci, thanks for bringing this up. I can’t think of an obvious workaround for the time being, but we will look into this and consider some optimizations which will raise the box size limits in the future.

2 Likes
3

OK, thanks @hsnyder :slight_smile:

4

Hi, I would second the feature for boxes up to 1400 although I think I read in one of the post that pyem can go to 1290.

Cheers,
Arnaud

5

Related query - are the GPU memory requirements any different (less, hopefully) for particles extracted in fp16 format?

6

No, all the mathematics are still done in FP32 as far as I know (same as RELION) so memory requirements are the same. Just 16-bit MRC takes up less space. :slight_smile:

edit: grammar.

3 Likes
7

I can confirm that @rbs_sci is correct.

2 Likes