Deepstream real-time performance

Hi

Is there a proper config file to run with deepstream 5.01 to get the highest performance on NX?

We tested NX real-time performance and got results below.
deepstream-app with 12 channels test result:
L4t r32.4.3 — got 28 fps
L4t r32.4.4 — got 30 fps

The performance is not close to the performance below.
The config file we used is “source12_1080p_dec_infer-resnet_tracker_tiled_display_fp16_tx2.txt.”

Already set nvpmodel to 2 and enabled jetson clock.

Thank you for any advice,

please propvide setup info as other tipics.

And, please check https://docs.nvidia.com/metropolis/deepstream/dev-guide/text/DS_Performance.html#jetson

Hi mchi,

Thank you for your prompt support.
The link helps.

The result we got as below.

  1. 16 channel (h264) worked, got 30fps.
  2. 23 channel (h265) only got 22-23 fps.

Attached the config file we modified.
Is there we missed on the config of 23 channel?

source23_1080p_dec_infer-resnet_tracker_sgie_tiled_display_int8.txt (5.6 KB)

Thank you,

Hi @HuiW,

  1. 16 channel (h264) worked, got 30fps. ==> total fps = 16 x 30 = 480
  2. 23 channel (h265) only got 22-23 fps. ==> total fps = 23 x (22-23) = 506 - 529

the fps you got with 23 channels is even higher than 16 channels, what do you expect ?

Thanks!

Hi mchi,

The performance of NX with H265 is 23 which is from the link you support.
We like to know, if it’s our setting problem or NX performance is not as the link said.
Sometimes our customer will ask what max performance of NX and how to test. We do need to provide proper data to our customer.

Thank you,

Hi @HuiW,
May I know where the image is from?

Is there a proper config file to run with deepstream 5.01 to get the highest performance on NX?

There are two notes:

  1. max NX capability
    $ sudo nvpmodel -m 0 // or sudo nvpmodel -m 2 , their difference is CPU core and clock.

    $ sudo jetson_clocks

  2. run some DL compute on DLAs after GPU gets fully utilized.