Puget Systems print logo


Read this article at https://www.pugetsystems.com/guides/1475
Article Thumbnail

Threadripper 2990WX Photogrammetry Performance: Windows 10 1809 vs 1903 & CorePrio

Written on June 11, 2019 by William George


When the performance of AMD's Threadripper processors is brought up, the discussion can often turn toward some of the oddities of this CPU's architecture and how it interacts with thread scheduling in Windows 10. That is something which AMD has attempted to improve with their Ryzen Master software, and which has been partially addressed by Microsoft as well (via Windows updates). Some folks in the wider AMD enthusiast community have tried to tackle this issue too.

While testing various photogrammetry programs recently this topic came up again in the comments on our articles, and it was suggested that Windows 10's 1903 update might improve Threadripper performance. CorePrio was also brought up as a potential boost for this processor family. Since we already had a testbed system set up, I decided to take a look at how the latest major version of Windows 10 impacted performance of the top-end Threadripper 2990WX, as well as whether CorePrio provided any improvement.

Test Hardware

Here is a list of the hardware we tested RealityCapture on. The CPU, RAM capacity, and video card were kept the same across all test runs, to avoid any of those throwing off the comparison. The only thing that changed was the version of Windows 10 - from 1809 to 1903 - and the addition / enabling of the CorePrio application (only tested on 1903).

When CorePrio was not present, AMD's Ryzen Master software was installed and was set to enable Dynamic Local Mode (DLM).

Test Platforms
CPU AMD Threadripper 2990WX
CPU Cooler Corsair Hydro Series H80i v2
Motherboard Gigabyte X399 AORUS Xtreme
RAM 8x DDR4-2666 16GB (128GB total)
Video Card NVIDIA GeForce RTX 2080 Ti 11GB
Hard Drive Samsung 960 Pro 1TB
Software Windows 10 Pro 64-bit (1809 & 1903)
Pix4D 4.4.12
Metashape 1.5.1

Benchmark Details

For testing photogrammetry applications, we have four image sets that we own the rights to - covering both smaller and larger size model and map projects. The exact photos from each set that are used in each application vary slightly, but if you want to check them out the smaller image sets are included in our public benchmarks which you can download and run if you want to compare your system's performance to what we have measured in various articles.

Each image set was processed 2-3 times per software configuration, and the fastest overall result was used for the comparisons below.

Results - Pix4D

Here is a chart showing the Threadripper 2990WX's performance in the latest version of Pix4D:

Pix4D 4.4 Performance with AMD Threadripper 2990WX in Windows 10 1809 Versus 1903 and with CorePrio

There is very little difference across all three tested software configurations. If anything, it looks like CorePrio slightly slows things down - but the numbers are probably within the margin of error.

Results - RealityCapture

And here is the 2990WX's performance in RealityCapture:

RealityCapture 1.0.3 Performance with AMD Threadripper 2990WX in Windows 10 1809 Versus 1903 and with CorePrio

RealityCapture shows a little bit wider difference between these software setups, but not in a consistent direction. It is also worth mentioning that, of the three photogrammetry programs we tested, this one has by far the most variance from one run to the next.

Results - Metashape

And lastly, a graph of 2990WX performance in Metashape:

Metashape 1.5.1 Performance with AMD Threadripper 2990WX in Windows 10 1809 Versus 1903 and with CorePrio

Here we do finally see some tangible gains in speed with 1903. Of the six tests here (there are two for each of the Model image sets, because Metashape offers different methods of generating 3D meshes) two are so close as to be almost identical, one shows a small slowdown, one shows a slight improvement, and two show substantial improvement. That is still a very mixed bag, but in particular with larger map projects it looks like making sure you are on Windows 1903 is a good idea. CorePrio didn't make as big of a difference, though: its results were slightly above or below the plain 1903 results, and probably within the margin of error of our testing.


Unfortunately, it doesn't look like improvements Microsoft made in Windows 10 version 1903, or those from the folks behind CorePrio, really make much difference in this type of application. It is generally a good idea to move to the latest version of Windows once it is stable - and when you have plenty of time for the update and to deal with any knock-on effects it might have - but I wouldn't rush to it in the hope of gaining performance in photogrammetry. If you use Metashape you might see some gains with large projects, but that is about it. And there is definitely no need to install the third-party CorePrio software for these workloads.

Even the gains that we saw in Metashape are not enough to make AMD's Threadripper 2990WX (or any other models in that family) a good choice for that application, though. Intel's Core i9 9900K is far faster in both Metashape and RealityCapture. The 2990WX does okay in Pix4D, but even there Intel offers a model that gives better overall performance consistency for a little bit higher price.

Looking for a
Photogrammetry Workstation?

Puget Systems offers a range workstations tailored for Pix4D, Metashape, and RealityCapture. Even in the most demanding situations, our workstations and designed to minimize downtime and allow you to work as efficiently as possible.

Tags: photogrammetry, RealityCapture, Pix4D, Metashape, AMD, AMD Threadripper 2nd Gen, Threadripper, 2990WX, Windows 10, CorePrio, Dynamic Local Mode

"... definitely no need to install the third-party CorePrio software for these workloads."... That is probably true, assuming they are using AMD's DLM instead.

However, there is utility in having an alternate, open-source, DLM implementation that can be tuned and targeted to specific applications. AMD's Ryzen Master also has other deficiencies, including not running with Virtualization Based Security enabled, a default (required?) setting of Hyper-V.

And, of course, there is Coreprio's distinct NUMA Dissociater feature, which continues to offer massive performance gains for some specific applications. The effect seen there is still mysterious and being researched.

I created the F/OSS software Coreprio as a intellectual challenge and exercise in thread management. I am not trying to convince anyone of its utility, nor profiting from its use. In short, maybe the bold text stung a little. I hope the community appreciates that Coreprio exists. In any event, I have enjoyed developing it!

Posted on 2019-06-12 16:55:43

The work you have done on this topic is fantastic, and it was not my intention to knock it or make it seem any less important - just to see whether CorePrio was beneficial in these specific workloads. It is a question we get a lot here at Puget, almost any time that Threadripper processors are included in our tests / articles, and that alone is a testament to how much your work is valued in the enthusiast community.

I've seen publications online showing CorePrio improving performance substantially in some applications, as well as others where it didn't seem to help (mirroring the results in these three programs) - and the only way to know when it will be beneficial is to try it out. Unfortunately, it doesn't look like the sort of workload that photogrammetry applications put on the CPU is one of those situations. I don't think that diminishes the value of your research, though, and I am sorry if the language in my article was painful. It was certainly not meant to be :)

Posted on 2019-06-12 17:13:28

I didn't mean to sound chaffed, as I'm really not. I just couldn't help but proffer some general defense, maybe unnecessarily. You know how it is; hard not to be emotional about one's work. Keep up the good work at Puget!

Posted on 2019-06-12 17:22:41

I'm keeping my 2990wx on Win10 Pro for Workstations 1803 Build. I once let auto update to 1809 and it broke the Task Manager, the CPU stats kept saying 100%, I had to temporarily rely on Bitsum's Process Lasso task utility. I put up with that for a good 5 months before I decided to just back up my data and clean install Win10 Pro for Workstations 1803.
I bet 1803 Build gets better performance than 1809 and 1903.

Posted on 2019-06-13 02:33:08

Thanks for this article, I did some searching and looks like the scheduler fix is there in 1903 but it's not activated yet- we need to wait for the upcoming chipset driver that will activate it- so maybe only after Jun 7th.
Wendell From CorePrio told this to Steve from Hardware unbowed:
Edit: looks like the AM4 chipset driver is out:
You can test Ryzen 2700X, lets see when TR4 driver will be out.

Posted on 2019-06-13 05:14:08

Oh, interesting... hmm. I don't recall any scheduler issues with the 2700X (or the other Ryzen / non-TR chips) and I don't have performance numbers on that CPU from pre-1903... so I'm not sure that we'd see anything helpful or informative there. It is a pity the video you linked to didn't specify what driver revision we are waiting on to enable the fixes in 1903 (so far as I heard when watching it). Do you happen to know what driver revision it is that is supposed to be?

Posted on 2019-06-13 19:41:42

You don't recall because the Ryzen 7 are AM4 CPUs, the chips are all on the same die. The Threadripper, at least 2990wx I use are broken into 4 different NUMA die cores, which supposedly Windows 10 in its current state doesn't properly manage processes beyond 2 NUMA dies unlike the robust scheduler used in Linux.
AMD supposed workaround to this since this past October has been the latest Ryzen Master DLM feature.
I watched that video of Windel with Steve and he seemed to react the improvements of 1903 and Ryzen that AMD were touting at E3 were to ZEN 2 and not for ZEN+.

Posted on 2019-06-14 16:25:25