Agisoft PhotoScan 1.4.1 - CPU and GPU Preferences

[yowich]

I found an interesting article by William George about whether to use CPU when performing GPU accelerated processing.
   
https://www.pugetsystems.com/labs/articles/Agisoft-PhotoScan-1-4-1---CPU-and-GPU-Preferences-1149/

He is concluding as follows:
"So what does all this mean? Assuming you are using a higher-end CPU and video card combination, it looks like turning the "Use CPU when performing GPU accelerated processing" OFF (by un-checking it) is most likely to perform the best. With a lower-end video card, probably just leave it enabled. If you have a chance, though, it would be ideal to test your own hardware setup and image sets to see which way gives you the best performance. Either way, it is a good lesson not to assume that default settings are the best."

[Mak11]

I found an interesting article by William George about whether to use CPU when performing GPU accelerated processing.
   
https://www.pugetsystems.com/labs/articles/Agisoft-PhotoScan-1-4-1---CPU-and-GPU-Preferences-1149/

He is concluding as follows:
"So what does all this mean? Assuming you are using a higher-end CPU and video card combination, it looks like turning the "Use CPU when performing GPU accelerated processing" OFF (by un-checking it) is most likely to perform the best. With a lower-end video card, probably just leave it enabled. If you have a chance, though, it would be ideal to test your own hardware setup and image sets to see which way gives you the best performance. Either way, it is a good lesson not to assume that default settings are the best."

Yup I've noticed that using the CPU in some cases can have an adverse effect especially in the depth map generation phase where it's often the case where the GPU has finished it's assigned work and I notice that the CPU still take several minutes to finish its own assignements. An even more interesting test is the Nvidia vs AMD:

https://www.pugetsystems.com/labs/articles/Agisoft-PhotoScan-1-4-1-NVIDIA-GeForce-vs-AMD-Radeon-Vega-1194/

This is especially important because Photoscan is the only fully featured photogrammetry soft that support's Radeon GPUs via its OpenCL path. As we can see the CUDA path for Nvidia GPUs totally obliterates the OpenCL path during the alignment process :-(. The slower RX580 is even faster than the higher-end Vega GPUs there..But things get evened out during the Dense cloud generation phase.

[GrinGEO]

@Mark
So do you advice in purchasing the RX580 instead of a Vega GPU?

[Alexey Pasumansky]

The tests have been performed on the version 1.4.1, since them there were some improvements and optimizations implemented related to the GPU-supported stages.

Also those tests involved quite small sample datasets, which may not be fully exploiting high performance GPUs. Additionally note that each discussed stage (Align Photos and Build Dense Cloud) consist of two parts, where the first one in both cases is GPU supported and the second is not, so it might be reasonable to compare only GPU supported stages timing while "use CPU" flag is unchecked in the GPU Preferences dialog.

[jomax]

Ill bring this thread back to life

I'm currently running a gtx 1070 and an i9 9900k. My issue is that my GPU only hits 6-7% usage when processing a scan and when I check this option it puts all the stress on my CPU. Any thoughts on this?

-j

[Alexey Pasumansky]

Hello jomax,

If you are trying to monitor GPU usage with the help of Windows 10 Task Monitor utility, please open the GPU tab and switch one of the graphs to "Cuda" option, it will show the GPU load during the GPU-supported operation in Metashape, providing that your GTX card is enabled in the GPU Preferences tab of Metashape.