Hi Harry,
Results may be somewhat different between SP and DP. In particular the
maximization (reconstruction) step is (somewhat) better (and definitely
more reproducible) in DP mode. I would advice only using SP if the memory
is really needed. We've also seen a bit worse speed for SP. Probably
internally (some C-libs?), some floats are still cast to doubles.
HTH,
Sjors
> Hi,
>
> I am puzzled by the much lesser performance from the single precision
> Relion 1.4. Here is the info:
>
> CPU: Intel Xeon CPU E5-2667 v3
> System Ram: 512GB
>
> Number of particles: ~45K
> Window size: 400x400
>
> 3D Auto Refine time:
>
> Single Precision: 58 hours (44 iterations)
> Single Precision (with pre-read option): 53 hours (34 iterations)
> Double Precision: 36 hours. (37 iterations)
>
> Memory usage under SP was halved as expected. But it seems the floating
> point operation runs in emulation mode although the CPU has integrated
> FPUs, or did I not compile the code optimally for single precision?
>
> Also: SP final resolution for both runs before post-processing is 4.09804
> vs 4.05825 of DP.
>
> Thank you.
>
> Harry
>
--
Sjors Scheres
MRC Laboratory of Molecular Biology
Francis Crick Avenue, Cambridge Biomedical Campus
Cambridge CB2 0QH, U.K.
tel: +44 (0)1223 267061
http://www2.mrc-lmb.cam.ac.uk/groups/scheres
|