Dense optical flow is slow on Jetson Orin NX 16GB

[xuang]

Tested on Jetpack 5.1.2, VPI 2.3, with the provided example

https://docs.nvidia.com/vpi/algo_optflow_dense.html

However, I can only get around 30FPS

grid size = 4, num_levels=1, quality = medium, input video is 1080P ~ 30FPS
grid size = 8, num_levels=1, quality = medium, input video is 1080P ~ 50FPS

However, it shows that in the page, AGX Xavier is way faster, I can’t see any results from AGX Orin though…

Is this results expected?

[AastaLLL]

Hi,

Dense optical flow on Orin uses a new hardware called OFA.
You can find some performance on the below doc.

Please note that OFA on Orin is slower than the desktop due to the lower clock value.

NVOFA Application Note - NVIDIA Docs

Thanks.

[proxylullaby]

Thanks for the reply. I am aware that it’s new hardware and it is also different from desktop GPU and Xavier(which uses nvcodec for the optical flow). I have read the doc you sent but it does not contain any information regarding the OFA on Orin.

I also tested this same code on AGX Xavier, but got only 50FPS(medium. 4x4, NVENC backend) , which is not even close to what you have in the doc.

[AastaLLL]

Hi,

We don’t have Orin OFA perf currently.
But you can check the perf on the NVOFA doc since it’s the same hardware just with a lower clock rate.

On Xavier, please check if you have followed the perf guide below:
https://docs.nvidia.com/vpi/algo_performance.html#benchmark

Do you want to use dense optical flow on Orin or on Xavier since the workflow is quite different.
Thanks.