External GPU (eGPU) experience with AgiSoft Photoscan

[ArnauCM]

Hi community,

Today I finally received a Sonnet eGFX Breakaway (eGPU enclosure). After some hours of scripts running up, uncountable reboots, crashes, etc. I managed to have it all working and wanted to share some amazing preliminary results I got.

My basic setup:

-   Macbook Pro 15’ (Late 2017)
-   CPU: Intel Core i7-7920HQ (4 Cores, 8 Threads @ 3.1 GHz, TB @ 4.1 (real 3.7))
-   GPU: AMD Radeon Pro 560 4GB
-   RAM: 16 GB 2133 MHz LPDDR3

Added:
-   Sonnet’s eGFX Breakaway Box 550 with Thunderbolt 3
-   ZOTAC GeForce GTX 1060 6GB

I used sample data from Parrot®. The sample contains 356 pictures at 4608x3456 resolution. I tried the following setups:

-   AMD + CPU
-   Nvidia + CPU
-   AMD + Nvidia + CPU
-   AMD + Nvidia

And just had time to Align and built de Dense Cloud in the lowest accuracy in boths. Here are some results.

TIME IN SECONDS (LOWER IS BETTER)


The processing times between the discrete Radeon GPU and Nvidia are jaw dropping. A 253% faster! As you can see, the best setup would be the dual GPU + CPU but, the different between this and just running Nvidia+CPU or NVidia+AMD are insignificant.

As you can see, for those of you who don't want to give up on your laptop, here it's a powerfull alternative.

That being said, if you want I can try to find some time to do some other benchmarks with Agisoft Demo sets like the statue or any other in different accuracies. I will try to do it so that those considering in buying one have a little more information.

Thanks for reading and hope you found this useful!

[SAV]

Hi ArnauCM,

Thanks for sharing. Very interesting results.

I would also be interested in the performance of Razorzone's hardware:
https://www2.razer.com/au-en/gaming-systems/razer-core-v2

Maybe there is someone who has access to it and is willing to run a few tests 

Regards,
SAV

[RHenriques]

I use a 2013 Mac Pro Octacore with two external Nvidia GTX 1080 Ti. Results are very good. Photoscan is one of the best apps in using all available resources. Both internal AMD fire pro 500 and external Nvidias are easily detected and used. However there are some processing steps where the hardware is not fully used.

In the Alignment step everything is used. Results are better if the CPU is not used.
In the build dense cloud step, the results are better if only the external GPU's are used. If the internal AMD's are active, there's about a 10% increase in processing times.
In the mesh building, no GPU's are used.

[ArnauCM]

If the internal AMD's are active, there's about a 10% increase in processing times.

Does it also happen to you? That internal AMDs seem to slow the process instead of increasing computability leading to higher processing times?

[RHenriques]

Yes ArnauCM, it happens that the internal AMD's really slow the dense cloud building process. It's strange because they do improve the alignment calculation but, with the external GPU's, this process is often very quick. With even 1000 photos it takes less than 5 minutes. The dense cloud stage, by contrary, takes longer to finish, often 10% or more, if the internal AMD's are used. I could not find any obvious explanation for that until now. The times should got better with more GPU's involved in the processing.
Cheers

[PolarNick]

Input data for computations on GPUs are generated on CPU, so when you have very fast eGPU - your CPU can became the bottleneck (especially when computations are in low quality - because CPU work part to GPU work part ratio will be bad).

When slow GPU added (AMD 560) - it adds pressure in input data preparation part, i.e. CPU tries to feed both GPUs and so fast eGPU will be used less.

P.S. this is just typical reason for the described behavior, and if my speculation is correct - you will see speedup with AMD used in case of higher depth maps quality. (because of better CPU work part to GPU work part ratio)

[ArnauCM]

Input data for computations on GPUs are generated on CPU, so when you have very fast eGPU - your CPU can became the bottleneck (especially when computations are in low quality - because CPU work part to GPU work part ratio will be bad).

When slow GPU added (AMD 560) - it adds pressure in input data preparation part, i.e. CPU tries to feed both GPUs and so fast eGPU will be used less.

P.S. this is just typical reason for the described behavior, and if my speculation is correct - you will see speedup with AMD used in case of higher depth maps quality. (because of better CPU work part to GPU work part ratio)

Interesting! Thanks a lot! So... I'll turn it on everytime I use depth filter. If disabled, I'll let Nvidia take the whole part!