Puget Systems print logo

https://www.pugetsystems.com

Read this article at https://www.pugetsystems.com/guides/1194
Article Thumbnail

Agisoft PhotoScan 1.4.1 NVIDIA GeForce vs AMD Radeon Vega

Written on July 25, 2018 by William George
Share:

Introduction

PhotoScan is a photogrammetry program: an application that takes a set of images and combines them to create a 3D model or map. This article is part of a series looking at how different aspects of computer hardware affect PhotoScan performance. For more information on this software, or to see the other entries, check out our introductory article.

PhotoScan makes use of the video card in a computer to assist with the computation of certain steps. The primary GPU accelerated steps are aligning photos and building a dense point cloud (a 3D representation of physical points that are seen from multiple angles in different photos). Because the video card plays a big role in these calculations, the model of video card (or GPU) used can have an impact on the amount of time those steps take. In this article, we are taking a look at AMD's Radeon line of video cards - including the Vega 56 & 64 - to see how they compare to NVIDIA's GeForce 1000-series.

Methodology and Test Hardware

Even though these processing steps in PhotoScan are GPU accelerated, the CPU still plays a role in performance as well. To ensure that the different GPUs are the focus of this particular test we selected a processor and platform that is among the fastest available for this application: the Intel Core i9 7900X, a 10-core processor which combines quite a few cores with a reasonably high clock speed (thanks to turbo boost). Both that CPU and the various GPUs we tested were attached to a Gigabyte X299 Designare EX motherboard.

All tests were conducted with "High" quality settings. The only results we are interested in for this comparison are the Align Photos and Build Dense Point Cloud times. Those steps use the GPU, while the other processing steps (like Build Mesh and Build Texture) are entirely CPU driven. We did go ahead and test those as well, just to make sure, and the results across both AMD and NVIDIA cards were within a 2% spread - safely within the margin of error.

It is worth pointing out that two of the AMD cards we are testing, the Vega 56 and Vega 64 models, are overclocked variants manufactured by Gigabyte. Normally we try to avoid cards that are factory overclocked since it makes for a bit of an unfair comparison to stock-speed models, but in this case, we couldn't find the reference-style cards we prefer to use. As such, the performance of those two AMD GPUs in our results will probably be a little higher than you would get with non-OC parts. Moreover, the cooling necessary for that overclocking means these cards have dual fans and take up three PCI-E slots (rather than the usual two) - so they would not work well in multi-GPU configurations, and as such we are looking only at single-card performance in this article.

If you would like more details about the full hardware configuration we tested on, and the image sets we used within PhotoScan, simply .

Results

Here are results for the Align Photos workflow step using the Building Model image set:

Agisoft PhotoScan 1.4.1 Radeon vs GeForce Performance Comparison - Building Model - Align Photos

And here are results for the Build Dense Cloud workflow step, again using the Building Model image set:

Agisoft PhotoScan 1.4.1 Radeon vs GeForce Performance Comparison - Building Model - Build Dense Point Cloud

Moving on to the Monument Model image set, here are results for the Align Photos step:

Agisoft PhotoScan 1.4.1 Radeon vs GeForce Performance Comparison - Monument Model - Align Photos

And lastly, here are results for the Build Dense Cloud step with the Monument Model image set:

Agisoft PhotoScan 1.4.1 Radeon vs GeForce Performance Comparison - Monument Model - Build Dense Point Cloud

Analysis

In the graphs above it is clear that the AMD Radeon cards are slower in the Align Photos step, but make up for much of that time with good performance during Build Dense Point Cloud. It is hard to see with the data spread across four graphs, though, so here is a total of all four (both parts of the workflow, on both image sets):

Agisoft PhotoScan 1.4.1 Radeon vs GeForce Performance Comparison - Total Time

Now we can see clear performance bands that these cards fit into:

Radeon RX 580 ~ GeForce GTX 1060

Radeon RX Vega 56 ~ GeForce GTX 1070

Radeon RX Vega 64 ~ GeForce GTX 1070 Ti & 1080

The GTX 1080 Ti is well in the lead, with no tested Radeon cards approaching it - then again, it costs more than any of the other models in this roundup. Speaking of prices, GPUs fluctuate quite a bit these days... but it is still worth looking at how much performance will cost. Here are the lowest prices from Newegg.com for each of these GPU models, across different brands and variants, as of 7/24/18:

Radeon RX 580 8GB - $260

GeForce GTX 1060 6GB - $290

GeForce GTX 1070 8GB - $400

GeForce GTX 1070 Ti 8GB - $450

Radeon RX Vega 56 8GB - $480

GeForce GTX 1080 8GB - $490

Radeon RX Vega 64 8GB - $580

GeForce GTX 1080 Ti 11GB - $720

*All prices rounded up to the nearest dollar, with used and refurbished cards excluded

The price-to-performance ratio winner looks to be the GeForce GTX 1070 Ti, with faster overall speeds than any of the AMD cards we tested while having a lower price tag than either of the Vega models. Keep in mind that those Vega cards we used were also overclocked by their manufacturer (Gigabyte) - so if anything, stock-speed Radeon cards (if available) might be slower yet.

One thing to keep in mind with the results shown above is that we did our tests at "High" quality. There is an option above that for building the dense cloud, "Ultra High", which will take substantially longer. Given the Radeon cards' strong showing in that portion of the test, it is possible they would pull ahead of some of the GeForce models. The GTX 1080 Ti, though, is still the fastest from among the cards we tested - and if you are spending hours per image set on Ultra High quality, it is probably well worth spending extra for that GPU.

Conclusion

AMD's Radeon cards performed fairly well in PhotoScan, but the price:performance champion for this application looks to be the GeForce GTX 1070 Ti. If you want the fastest overall performance, spending a little more for the GTX 1080 Ti will net you about an 8% increase in performance. Multiple GPUs in a single system are also an option, so check out our article on that topic if interested.

Recommended Systems for Agisoft PhotoScan

Standard Workstation

Configure

Specialized Workstation

Configure

Tags: Agisoft, PhotoScan, GPU, Video Card, AMD, Radeon, Vega, RX, 64, 56, 580, NVIDIA, GeForce, Pascal, Performance, Comparison

Hi there,

Just wanted to point out that there something wrong with your Radeon results here because running exactly the same bench on a Radeon Fury X (and a i7 6700K) I was getting alignement speeds identical to the GTX 1080. That was more than 6 months ago with PS1.4.3.
You guys should do an article with the newest version of PS which is now Metashape 1.5. It's now fully GPU accelerated from the laingment to the mesh generation with the new Depth Maps based meshing method.

Cheers

Posted on 2019-01-11 01:54:24

I wonder if 1.4.3 improved AMD support compared to 1.4.1 that this article used? Not sure.

I do plan to test Metashape 1.5, but because of the branding change it will likely require a major re-work of our test scripts. I won't have time to work on that till I finish SOLIDWORKS testing for SWW, which is what I am starting in on now. Hopefully sometime in February we will have updated articles for 1.5.

Posted on 2019-01-14 20:05:55

William,
AFAIK nothing has changed with regards to AMD GPU improvement in 1.4.1 & 1.4.3/5. AMD dropped SPIR support after Radeon driver version 18.5.1 so GPU support was broken in Photoscan if you upgraded the drivers (errored out and didn't run) but Agisoft found a work around which they included in Metashape (1.5). One thing to make sure of is to never select "Use CPU when performing GPU accelerated processing" in preference in Photoscan/Metashape because it actually slows down overall processing time when you have a dedicated GPU in the system, especially when using an OpenCL card. Agisoft finally added this info just above the setting so that users don't enable it thinking that it would help them.

I would be great if you guys could get your hands on the Radeon VII. AMD is claiming some rather surprising +60% improvements compared to VEGA 64 in OpenCL (Luxmark bench) thanks to the 1TB/s of bandwidth.

Cheers

Posted on 2019-01-16 14:46:20

We just got a Radeon VII in, and are working on getting better image sets for our photogrammetry tests. I'll be updating our PhotoScan benchmark to Metashape soon too, so you can probably expect a review including the Radeon VII in the next month or so :)

Posted on 2019-03-06 18:16:10

William,
I also just got my hands on a Radeon VII and started some extensive testing with Metashape 1.5.1 (1.5.2 should be out this week). Feel free to ping me if you have any questions.

Cheers

Posted on 2019-03-20 12:00:26

Small update: Metashape 15.2 build 7838 has totally busted the new depth maps based mesh generation on all GPUs (it's slower & messes up the meshes)..Best to stick with 15.1 for now.

Cheers

Posted on 2019-03-25 18:32:46

Oh, thank you for the heads-up! I am starting in this week on updating our benchmarking from Photoscan to Metashape, so this is very timely information :)

Posted on 2019-03-25 18:56:30

Posted on 2019-03-26 11:15:24