Does Lan speed related with GPU?

Autonomous Machines Jetson & Embedded Systems Jetson AGX Orin
,
1

Our carrier board had 5 RJ-45 ports : 4 x 1Gbe PoE & 1 x 10 Gbe

To ensure carrier board Lan performance, we found out lan speeds can’t align during different roles.

Here’s our test result:

image
image732×231 7.13 KB

Due to the aforementioned reasons, the current stress test is conducted using the NV AXG DV Kit as the server side. The observations are as follows:
A. Different stress conditions result in varying network speeds:
。Without GPU stress, the speed is approximately 6-7G.
。With GPU stress, the speed can reach around 9G.

B. Subsequently, if all RJ-45 ports had connected with devices. All 1G LAN ports speed is about 685MB, and the 10G LAN port is reduced to only 7GB. However, when individually connected, each 1G port can achieve speeds of 7/800M.

Even though the 1G and 10G LAN chips are from different manufacturers, the overall maximum supported speed for Carrier board seems to be around 10G.

4

Hi,
For profiling ethernet performance, please refer to the posts:
Orin Unable to Reach 10 Gbps Networking - #6 by DaneLLL
[35.4.1] LAN Tx performace is getting lower when HDMI is plug out -EE - #22 by DaneLLL

Suggest use minimum rootfs and execute sudo jetson_clocks

5

Hi, Thanks for your feedback.
The 2nd post you shared with us had different test conditions with us.
The post you provided mentions that

  1. Lan speed rises if HDMI is connected
  2. Lan speed drop if plug out HDMI cable.

But our condition is on the contrary. And use minimum rootfs mode to replace GUI mode is not friendly for normal users. Would you help to advise more suggestions? Thanks.

6

Hi,
In out test we see worse performance with desktop GUI, so desktop GUI does impact performance. If you are looking for optimal throughput, it is a must to use minimum rootfs.

When sudo jetson_clocks is execute, we see identical performance in both HDMI is connected and disconnected. The reason is without jetson_clocks and HDMI is connected, clock of certain engines may drop since there is no loading from HDMI output. With jetson_clocks, the engines are in maximum clock always so performance is lasting.

7

Hi DarneLL,
Thanks for your reply. But the conclusion is different with us. Here’s the comparison table. Please let me know if any advice. Thanks.
image

8

Hi,
Please share the steps so that we can replicate it on AGX Orin developer kit with Jetpack 5.1.3.

9

Procedures for verify sudo jetson_clocks test result.

  1. To connect 10GbE ethernet port with NV DV kit. (Server : NV DV Kit/ Client : Our product)

  2. Connect monitor via HDMI port and Run  sudo jetson_clocks.

  3. Start iperf3 and throughput performance is 7.X G

  4. Disconnect HDMI port and check throughput performance on Server, throughput  test result is 9.X G

And we also find out another issue, 10GbE network port throughput performance will slow down while 1GbE network port be connected simultaneously.
Make sure 10GbE network port throughput performance can reach 9.X G, then connect 1GbE network port *4. 10GbE network port throughput performance will slow down to 7.X G.
Please advise suggestions. Thanks.

10

Hi DarneLL,
I wanted to follow up regarding the test procedures I updated a month ago. Despite sending the revisions, I have yet to receive any feedback from your end. If you have any suggestions or comments on the updated procedures, please do not hesitate to share them with me at your earliest convenience. Your input is highly valued and will contribute to the refinement of our testing protocols.

11

Hi,
For information, do you observe it on AGX Orin developer kit? Since it is DP port on developer kit, it seems like the steps are for custom board. It says HDMI port in the steps.

13

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