GPU Freq impact on glass-to-glass latency

Robotics & Edge Computing Jetson & Embedded Systems Jetson Orin NX
,
1

Hi Nvidia Support,

We are using Jetson Orin NX with IMX485 camera to stream 4K60 (H265) video over Wifi.

Pipeline used.
gst-launch-1.0 nvarguscamerasrc ! ‘video/x-raw(memory:NVMM), width=(int)3840, height=(int)2160, format=(string)NV12, framerate=(fraction)60/1’ ! nvvidconv ! nvv4l2h265enc bitrate=20000000 control-rate=1 insert-sps-pps=true num-B-Frames=0 maxperf-enable=1 ! h265parse ! rtph265pay config-interval=1 ! queue ! udpsink host=192.168.47.1 sync=false async=false port=5000

After disabling GPU 3D Scaling using below command , GPU frequency increased to Max.
echo 0 > /sys/devices/17000000.ga10b/enable_3d_scaling.

which helped in reducing glass-to-glass latency for about 1 frame.(16 ms).
We have tested this multiple times, & it showed the same results each time.

we know that nvvidconv is using VIC & encoder is using nvenc.
Could anyone please tell us which component of the above pipeline is using GPU & why it affected latency?

Thanks & Regards,
Kaushal Dekivadia

3

hello kaushal.dekivadia,

your pipeline did not have GPU involved, or, do you use Vision Programming Interface (VPI) in the background?

4

Hi Jerry,

We know that our pipeline doesn’t use GPU,
Then how increasing GPU Freq reduces latency.

Please let us know how it is related.

Thanks & Regards,
Kaushal Dekivadia

5

hello kaushal.dekivadia,

there’re some post-processing which using CUDA service. you may disable it for your own use-case.

6

Hi Jerry,

Thanks for the valuable input.

Could you please share more info on this post-processing using CUDA?
what is it doing ?
& How to disable it?

Best Regards,
Kaushal Dekivadia

7

hello kaushal.dekivadia,

there’re some pre-ISP stages and some WDR specific functions handled by CUDA, sorry, I cannot reveal more details…
it’s using makefile to control the service within the pre-built library, you cannot disable that by yourself.

may I know which Jetpack release version you’re now using?
I can re-built another library for your testing if there’s strong request to disable it for verification.

8

Hi Jerry,

Thanks for the response.

This behaviour is also seen with a USB camera which doesn’t involve an ISP pipeline or nvarguscamerasrc plugin.

10

Hi,

The pipeline doesn’t have a CUDA task involved.
To find the underlying CUDA implementation, would you mind profiling the pipeline with nsys so we can check which function runs on the GPU directly?

Thanks.

12

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