New GPU to improve speed or useless

[lecaramel]

Hi,

I'm currently looking about improving the photoscan process by giving a boost to my computer. I don't want to change my workstation since it's a 6 months old one...

Right now, I have a i7 4770L + GTX780 and 32 Gb of RAM.

I found in an old thread (dec 2010...) that the GPU was used only on the geometry building stage, but not in camera alignment, dense cloud and texture creation.
Then if the GPU is still used for only the geometry stage, I may prefer spending some money in a new lens or an extra DSLR, but if the GPU is used and in almost or all the steps, perhaps it's worth buying a second GPU.

And about the CPU, if I disable a total of 4 cores in the preferences (2 per card since my hyper threading is enable), how things are working? only 4 physical cores will be use and not the HT ones? or it's an "unknown"?

Thanks!

Does someone have the info?

[David Cockey]

I use Open Hardware Monitor to observe CPU, GPU and memory usage. http://openhardwaremonitor.org/

GPU is only used during for creating the dense point cloud. It is not used for aligning photos, building the mesh, decimating the mesh, texturing or aligning chunks.

I have two computers with I7 CPUs (four physical cores, eight effective cores in total). One has an AMD GPU, the other has a Nvidia GPU. 2 cores are set as disabled in the OpenCL preferences on both machines. My testing indicated performance is not very sensitive to the number of cores set as disabled. All 8 cores are used, mostly at 100% during align photos and mesh generation, and all 8 cores are used at over 90% during dense point cloud generation.

[lecaramel]

Thank you David for this reply and thee information.

Since dense cloud is one of the longuest step, perhaps an extra GPU could help a little more than a little bit...

Thanks again!

[tommyboy]

Our computing machine has an i7-4930K with two R9 280X cards in it, and we are using 60 pictures at 18MP apiece (Canon 600D).  Timing a 'high' dense cloud reconstruction using the various optional numbers of GPUs we got:

0 GPU: 67 minutes
1 GPU: 23 minutes
2 GPU: 16 minutes

So while the first GPU got us a massive speedup, the second GPU got us just 30% faster, and sped up the entire process (photos->textured model) by 15%.  Given that each card is 12% of the system cost, it barely makes sense to add the second GPU.  What's more interesting though, is dividing the cores and GPUs amongst two PS instances, and running them at the same time.  Then you get a real nice speed up!  Assuming you have the need to run multiple PS instances simultaneously.

[lecaramel]

Tommyboy:
Thank you for this feedback and more important, this tip about running two instances of PhotoScan by dividing the resources.
I'm more in the process of increasing my current workflow since I'm doing mainly tests and not using PhotoScan in a production environment, then I'm processing just photos after shooting, but since one of my co worker starting to take photos for scans too, perhaps it could be a smart move to try this two instances of PhotoScan.

[tommyboy]

http://www.agisoft.ru/forum/index.php?topic=1330.45

This thread is worth checking out, there is much talk as far as running multiple PS instances to increase CPU core utilization.  I haven't done a lot of careful testing in that regard, or on enabling/disabling hyperthreading, but it did give me the idea to run multiple PS instances for making better use of the two GPUs.  I will work on getting more hard data to support this approach, since my test data so far has just been for one PS instance, but I assure you setting two instances going over the weekend got me many more ultra high dense cloud solves than I got normally..

[lecaramel]

Thank you for the link, I'll look that closely!