Where's the Bottleneck (RAM)?

[Eggz]

I'm stitching together large panoramic photos in Photoshop with the machine in my signature. Since it always takes so long, I monitored my processes to see what was going on. My guess is RAM speed, that's only a guess, and I want to check.

CPU: 25% - 60% (spread fairly evenly across 12 threads)
RAM: 10 - 32 GB used (increases over course of render)
SSD: 0%-5% (not involved during render)
HDD: 0%-2% (not involved during render)
GPU0: 1%-8% (not involved during render)
GPU1: 0% (not involved during render)

From watching over the course of the render, I notice that only the CPU is working, but it isn't getting stressed, and it doesn't go up to 4.5 Ghz (max speed) very often. While that's going on, the RAM increases from just a few GB used to nearly all 32 GB.

My thought is that the CPU is generating stuff for the RAM to hold, but the RAM can't accept it as fast as the CPU can deliver it. That's just the best explanation I can come up with, though. What still doesn't make sense is that my RAM is relatively fast (2133 Mhz, CL9).

Any thoughts?

I'd love to reduce these pano render times. They can take nearly an hour sometimes.

 

[Eximo]

You can try an alternate solution, maybe this CUDA based one will work better for you. Not sure about memory usage though, but your GPU memory certainly has bandwidth to play with.

http://webuser.hs-furtwangen.de/~dersch/PTStitcherNG/PTStitcherNG.html

Maybe a very fast PCIe SSD used as a ram drive might help? (Might try out a fast SSD USB drive and try ReadyBoost as a quick test)

 

[Fokissed]

Your memory is capable of transferring at 17GBPS but your CPU is likely able to process at 300GIPS/100GFLOPS (which, assuming 64bit numbers, would require 2400GBPS integer and 800GBPS float). Of course, cache locality significantly reduces the need for high RAM speeds, but since you're using 32GB of RAM, the 12MB of cache on your CPU is unlikely to have the relevant data used for the calculations.

 

[Eggz]

Those sound like bandwidth numbers to me. There's no way its moving data at anywhere near that rate. Each stick has its own channel, so I don't think bandwidth is the issue.

If only I had some super-crazy fast RAM to test, I'd be able to see if it helps at all. Benchmarks for RAM I see never actually touch upon things like this, so I don't know.