How to increase GPU utilization

[Tolga Gürcan]

Hi,

When I check the GPU utilization while running openpiv GPU I get values between 20% and 30%. Is there a way to increase this value or does it have any role in the speed of the algorithm?

Best,

Tolga

[Alex Liberzon]

Please create a GitHub repo with some image examples - we need to learn the problem. I do not use the GPU version very often, and we have not checked it with recent Python and CUDA versions. Please explain which hardware/software you use, and which versions of openpiv-python-gpu you use. We have at least two on GitHub. 

Alex

[Tolga Gürcan]

Hi Alex,

Sorry for the late reply.

I’ve created a small GitHub repository with image examples to demonstrate the problem:
https://github.com/tg5941/piv_debug/tree/main

For the GPU implementation, I am using the openpiv-python-gpu package from:
https://github.com/OpenPIV/openpiv-python-gpu

Below is a summary of the hardware and software environment I am using.

CPU
AMD Ryzen Threadripper 2950X (16 cores)

Memory
125 GiB RAM

GPU
NVIDIA TITAN RTX (×2)
24 GB memory each
Driver version: 570.195.03

Operating System
Ubuntu 20.04.6 LTS

CUDA / Driver
NVIDIA Driver: 570.195.03
Driver-supported CUDA version: 12.8

CUDA Toolkit
nvcc: CUDA compilation tools, release 12.4 (V12.4.131)

Please let me know if additional details would be helpful.

Best regards,
Tolga

5 Aralık 2025 Cuma tarihinde saat 06:52:59 UTC-5 itibarıyla alex.l...@gmail.com şunları yazdı:

[Alex Liberzon]

Thanks, Tolga

I do not have access to the GPU unit right now. We'll try to find something in the lab. 

I tried using Colab for this task, but I do not know how to test the utilization:

https://colab.research.google.com/github/tg5941/piv_debug/blob/main/PIV_multi_gpu_github.ipynb

We need a toolbox that shows the results like a cprofile for GPU. 

I suggest is also to try: https://github.com/ali-sh-96/openpiv-python-gpu 

[Alex Liberzon]

Please see if this repo is useful - it seems that they were very much concerned about the GPU utilization on OpenPIV https://github.com/MechMicroMan/openpiv-xl

On Friday, December 19, 2025 at 7:48:20 PM UTC+2 tgur...@gmail.com wrote:

[Tolga Gürcan]

Hi Alex,

Thanks for sharing the openpiv-xl repository — I’ll take a closer look at it and try to evaluate performance on my system using the repos you suggested.

While searching for other GPU-accelerated PIV implementations for comparison, I also came across this project:
https://github.com/NikNazarov/TorchPIV

I realize it’s a separate code base and not directly related to OpenPIV, but I was wondering whether you have any experience with it. Do you think it could be a useful point of comparison?

Thanks again for the help!

Best,
Tolga

Alex Liberzon <alex.l...@gmail.com>, 22 Ara 2025 Pzt, 03:23 tarihinde şunu yazdı:

--
You received this message because you are subscribed to the Google Groups "openpiv-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to openpiv-user...@googlegroups.com.
To view this discussion, visit https://groups.google.com/d/msgid/openpiv-users/97ea386d-e836-4141-aace-3468b03b7400n%40googlegroups.com.

[Alex Liberzon]

Hi Tolga

I do not have any experience with TorchPIV. I recall that I contacted the author, and we even chatted on Zoom; he was positive, but this didn't materialize in adopting his work into OpenPIV. 

Please find a way to consolidate all the packages into a better GPU version for OpenPIV. My hands are too busy right now, and I'm not an expert in CUDA. 

I just came across this project https://github.com/exaloop/codon and am thinking of starting to test it on a CPU that apparently also works on a GPU. 

Alex

[William Thielicke]

Hi, sorry for commenting here without being really involved, but I was curious about the speed of TorchPIV. The author says:

This method allows processing 4 thousand pairs of images with a size of 4 MP each with a search window of 64, overlap of 50%, two iterations with re-arranging (an increase in the number of vectors by 4 times) in less than 10 minutes. 

I quickly tested this with PIVlab (no GPU processing, only parallel CPU processing on a i9-14900K CPU with 8 cores). And I get 13 minutes for the same task. So "GPU processing" alone doesn't really seem to mean too much in terms of speed. It would be interesting to see a comparison between TorchPIV and OpenPIV GPU. Is there similar information on speed and parameters for OpenPIV GPU?

[Alex Liberzon]

I do not have the hardware to run the tests. I think we need NVIDIA to be on our side to test various PIV algorithms w/o GPU. The papers by the colleagues from Uni. Toronto has shown some quantitative results - but on clusters of GPUs, not on a workstation machine. I look for some colleagues with Mac machines with their hardware - maybe they can do some test runs. 

[Erich's Lab]

I can do a test on one of my GPUs (RTX 5090) on Sunday. I never really checked the GPU utilization on the few occasions that I used a GPU, so this would be something of interest. 

[Erich's Lab]

I SSH'd into my server and ran a few tests. I am only getting a 6x increase in performance and 8% GPU utilization for 1920x1200 px 12 bit PIV images processed with 128>64>32>16 and at 75% overlap. On larger PIV images (e.g., 4504x4504 px 16 bit images), the performance increases are more noticeable when using smaller window sizes and high overlap. This is probably due to processing PIV images serially which underutilizes hardware resources. Processing PIV images asynchronously would likely be the best way to parallelize PIV processing and take better advantage of GPU resources.

[William Thielicke]

Can you also give an order if magnitude for the general speed like the guy from TorchPIV did?

[William Thielicke]

Are this numbers compared to single core CPU processing? Btw. we got the latest AMD R9 as Lab Computer yesterday, and when I disable hyper threading I get another significant speed increase. Strange.

Erich_...@hotmail.com schrieb am Mittwoch, 18. Februar 2026 um 01:47:03 UTC+1:

[Erich's Lab]

This is a single thread with SSE2 SIMD. Using 24 Threads on SSE2 SIMD (required a recompilation of the c++ Meson port), the performance difference is not much different compared to the GPU implementation. The overhead sort of kills the performance gains from using CUDA cores especially since so little resources are used to process a single PIV image pair.