Low efficiency of AI application: RTX 3080 vs GTX 1060

I am seeking an advice to solve my problems with RTX 3080. I am running an OpenPose demo app. (AI based human pose detection kit) on Ubuntu 18.04.5 LTS with the latest driver 455.23.04, CUDA V11.1.105, cudNN 8.0.5 and getting very low FPS (~8 FPS). For comparison, the same app running with the same OS but GTX 1060 managed to achieve similar FPS performance.

My question:

Could the inferior performance of RTX3080 be due to the linux drivers or else? Any clues or advice will be very much appreciated.

Hi @dradamPK,
Please allow us some time, we are checking on this.

Thanks!

Can you please point us to the code base of the app, or post the API log for a few iterations?

Thanks!

So do I. 01_body_from_image.py is ok, but 02_whole_body_from_image.py got a segment falut.

Hi, sorry I haven’t replied so far. This is the point to the application source code: https://github.com/CMU-Perceptual-Computing-Lab/openpose

mayicome I have the same problem with 02_whole_body_from_image.py example.

In the meantime, I reinstalled Ubuntu with the latest nVidia driver 455.45.01 and CUDA 11.1. I saw that there is also a new CUDA 11.2 but I am not installing it because at the moment there is no cudNN for 11.2 yet.

Connecting the GPU directly to the motherboard PCI-E allows me to achieve 10 FPS. Previously, I only had to use it with a riser - then I got 8 FPS. To reduce GPU consumption, I run Ubuntu in text mode and connect it through a VNC server.

Unfortunately, the performance of the RTX3080 is still 10 FPS. It’s very strange that the same computer and the same installation (just a different hard drive) with GTX 1060 also gives 10 FPS. But the biggest question is that the GTX 1060 with CUDA 10.2 gives 12fps !? Is there any problem with CUDA 11.1 for AI applications?