How to increase FPS in Nvidia Xavier

Autonomous Machines Jetson & Embedded Systems Jetson AGX Xavier
1

Hi,

I using Nvidia Xavier Jectpack 4.1.

Please suggest me the possibilities to improve FPS for YOLO V3 in the above mentioned board.

Regards,
Aravindan k

2

Improve compared to what?
Are you using CPU or CUDA YOLO?
Are you training your own models or using a pre-trained model?
Are you running the Xavier at the highest performance setting?

Making sure you use CUDA, and changing the resolution to a smaller image size (and then re-training your model to match) are useful steps.
(If you’re using an NVIDIA implementation of YOLO, rather than DarkNet, then you can presumably be sure that it’s already using CUDA, so reducing the resolution is your best bet.)

3

Hi,

Improve compared to what?
Compared to Jetson TX2. And anything maximum possible.

Are you using CPU or CUDA YOLO?
CUDA YOLO

Are you training your own models or using a pre-trained model?

Own model trained in TX2.

Are you running the Xavier at the highest performance setting?
Please tell me what are all the highest performance setting. And how to do highest performance setting.

Regards,
Aravindan K

4

Hi aravind24.ece,

I think you are looking for the following commands:

sudo nvpmodel -m 0
sudo jetson_clocks

In case you are wondering what these commands do, first nvpmodel is used to change the power management “profiles”, you can find more info about it here: https://www.jetsonhacks.com/2017/03/25/nvpmodel-nvidia-jetson-tx2-development-kit/. Then there is jetson_clocks that it is used to set max frequencies to CPU, GPU and EMC clocks.

Also you can monitor RAM usage, CPU/GPU usage (and among many other stats) using:

tegrastats

You can find more info about this tools in the L4T documentation: https://docs.nvidia.com/jetson/archives/l4t-archived/l4t-322/index.html

Hope this helps,

-Jafet