Hardware Provisioning and CPU/GPU Settings?

[jazzyj]

I'm trying to determine the most efficient hardware provisioning for a cloud computing workstation for Photoscan.  I typically do mapping jobs generating Orthomosaics, DEMS, and sometimes models, with typically 1000-3000 cameras (each a 20MP photo  - yes these are large 100acre-300acre areas of land). 

This doc seems to give decent guidelines as far as memory: http://www.agisoft.com/pdf/tips_and_tricks/PhotoScan_Memory_Requirements.pdf

If I read it right it seems to indicate for my typical use I want minimum 32GB RAM, but 64MB would be more optimal.  However I noticed the RAM for generating a model is way higher.  I may need to so something like a model of several buildings on a 75 acre property.  I'm guesstimating, I'd end up with maybe 750 nadir photos and another 500 obliques, so 1,200 photos total.  The chart doesn't even go beyond 500 photos on the model so is processing something with 1,200 photos essentially impossible if say I had dual GPU, 244GB RAM, 32 processors, and Dual GPUs?

I guess you could always batch process the 20MP JPEGs with a higher compression rate to lower memory requirements as necessary.

This great article indicates the old "Law of Diminishing Returns" as far as CPU cores, so when given a choice, obviously you want to choose a "RAM optimized" workstation setup as it appears the RAM is the limiting factor in the size of the job that can realistically be processed and the GPU is more important than the CPU once you get above 6 CPUs.


Also I saw something about reserving one CPU for every GPU but I didn't see that setting in Photoscan so maybe that was a setting in an older version and now the software automatically does that for you?

And finally, are there steps in the workflow that don't benefit from gobs of RAM and GPUs.  I'm thinking I could cut processing costs by moving the project files to a lower cost platform for those steps that doesn't have multi-GPU and tons of RAM.

Thanks