Does DLA support YoloV8?

Autonomous Machines Jetson & Embedded Systems Jetson AGX Orin
1

Set up info:

Jetson AGX ORIN
Deepstream version : 6.3
Jetpack version: 5.1.2
TRT: 8.5.2

Issue: I tried to generate DLA TRT engine for YoloV8 using ‘enable-dla’ flag. However, many layers were not supported and GPU fallback mode was enabled for the same. Still, I am able to run the inference using ‘deepstream-app’ but the fps rate is heavily affected and dropped from 130 to 20 (fps).
Will the support be given for yolov8 or yolo-nas in the near future?

Also, please let me know if there’ s any way to improve the fps rate.
Thanks.

Regards,
Lokaram Thabasu

4

Dear @le.lokaram.t,
When you use DLA, if there is a non DLA supported layer, it will pushed to GPU. So there will be intermediate data transfers across DLA and GPU which increases overall execution time. So the FPS is expected to drop. Please see Developer Guide :: NVIDIA Deep Learning TensorRT Documentation for details on supported layers and restrictions.

You may use just GPU to get more FPS.

5

Thanks for your response @SivaRamaKrishnaNV

Will there be any support for those unsupported layers in the near future?

6

Dear @le.lokaram.t,
Could you share the log to see the reported unsupported layers? If you use trtexec binary, log should report that information.

7

Hi @SivaRamaKrishnaNV,

Attached is the file with unsupported layers logs list.

file.txt (58.7 KB)

Thanks.

9

Dear @le.lokaram.t,
Currently there is no visibility on new DLA layers. If you have list of operators that needs to be added. You may raise a RFE at Deep-Learning-Accelerator-SW/operators at main · NVIDIA/Deep-Learning-Accelerator-SW · GitHub

11

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.