Using GPU + DLA for maximum no. streams

• Hardware Platform: Jetson
• DeepStream Version: 5.0.1
• JetPack Version: 4.4
• TensorRT Version: 7.1
• Issue Type: Question
• Requirement details
Leverage both DLA’s and GPU on Jetson Xavier AGX to increase the number of streams processed without reducing the inference interval

Lets say we have a 4xCCTV camera setup, if we can process 2 cameras on the GPU at 30FPS how could we specify to run 1 camera on DLA1 and the other camera on DLA2?

The end result would be a tiled output of 4 cameras.

Specifically using the python bindings, would help.

Can we have 3 pgie elements where 2 are using the DLA and 1 uses the GPU

Regards Andrew


To run DNN on a different process(DLA or GPU), you will need to use a separate engine file.

For example, camera-1 on DLA0 and camera-2 on DLA1:
You will need to create two separate nvinfer: one for camera1+DLA0 and the other for camera2+DLA1.

Since only one pgie is allowed in the deepstream pipeline, the other one need to be defined as sgie.
Here is an dual detector example for your reference:

The nvinfer component by default is running on the GPU.
You can specify it to use DLA with the below configure:

use-dla-core=0  # 0 for DLA-0, 1 for DLA-1