Decrease performance of GPU when DLAs run?

Autonomous Machines Jetson & Embedded Systems Jetson Xavier NX
1

Hi all,
I run multi-stream(H.264-720p) with peoplenet34 on Jetson xavier nx . I reached the below results:

The FPS of system reach to 135 and EMC 50% with only GPU running .
When I run both GPU and two DLA at the same time, the EMC reach to 95% and FPS of GPU and DLA are 45, 30 and 30 respectively.
That means overlay fps is 45+30+30=105.

Q1- What is the EMC?
Q2- Why has the frame rate dropped from 135 to 105 with the running of DLAs along with GPUs?
Q3- ‌Under what circumstances do DLAs reduce GPU performance?

3

Hi,

You can find the details of tegrastats in our document below:

https://docs.nvidia.com/jetson/l4t/index.html#page/Tegra%20Linux%20Driver%20Package%20Development%20Guide/AppendixTegraStats.html

EMC is related to memory bandwidth.

Percentage of EMC memory bandwidth in use relative to the current running frequency

First, it’s recommended to check if your model can fully run on DLA.
Since DLA doesn’t support all the TensorRT layers, some non-supported operations will be fallback into GPU.

As a result, the GPU resource might be shared between some layers in DLA model and the original GPU model.
And generates overhead from resource sharing and the memory transfer(DLA<->GPU).

Thanks.