AGX Xavier to AGX Orin Migration questions

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

Hi,

We’ve been working on a project with the Jetson AGX Xavier for some time now and it was selected as our main processor as it complied with the performance that we wanted. It did not have much margin but so far it was enough. The issue we have now is the HW availability, it is very hard to find developer’s kits and our last purchase was the Auvidea X221. This resulted in a terrible user experience and very low reliability.

(This is out of the question in case you are not interested) We received the board in April and had to send it back, we started working with it just a week ago and we see future issues (it is not even possible to reboot it without having a black screen) and do not trust the flashing process of the board (this was the initial issue and we had to ask to get the “repaired” board already flashed).

Right now we think that migrating to ORIN may be the best option, get a couple of dev’s kits and keep working. The issue is that we have constraints on energy usage (it will be on a drone) and time to develop, and we would like to seek more information before doing the purchase. Regarding energy usage, we may need to limit it to 30W as the Xavier, we suppose that the performance of the ORIN should still be better or at least the same, is that right?

Regarding the migration effort, we are mainly concerned by the deep learning models that we use with TensorRT. Currently, we are using Unet and YOLOv4 (YOLOv3 could serve us). For YOLOv4 our previous repo is not working anymore (with the Auvidea X221 we have Jetpack 4.6.2) as it used already deprecated functions in TensorRT 7, and it will need a considerable customisation effort to solve it. With Unet we are having some issues but we are not certain it is related to the version as the migration is successful (from ONNX using trtexec) but the output of the model is nonsense. Are you aware if these architectures are implemented in TensorRT 8 or if someone has been able to run them in the ORIN?

To give a broader overview, we will have to interface with multiple USB cameras and a cube (Pixhawk 2) to recover telemetry; everything is coded in C++, with OpenCV and we may use ROS2 as middleware, mainly for peripheral management tasks. Any information regarding possible issues during migration or incompatibilities from Jetpack 4.5-4.6 to 5 is welcome!

Also, the ORIN NX is definitely the perfect option for us but we’ll have to finish most of the development for November, and even if it is released in September, when is it supposed to arrive if ordered?

Thanks in advance

2

Basically Yes, both Xavier and Orin module can be limited by the NVPmodel setting for power mode:
Xavier: NVIDIA Jetson Linux Driver Package Software Features : Clock Frequency and Power Management | NVIDIA Docs
Orin: Jetson Orin NX Series and Jetson AGX Orin Series — Jetson Linux
Developer Guide 34.1 documentation (nvidia.com)

That will be too rush due to carrier board for Orin and Orin NX are not pin campatible. I will suggest to bypass this consideration.

4

Hi kayccc, thanks for the reply.

I see that the maximal frequency for the GPU at 30W is 30% less for the Orin compared to the Xavier (900 MHz Xavier - 624.75 MHz Orin), it does seem like a considerable reduction in frequency, is any data available comparing inference performance between Xavier and Orin for the same neural network for the different power modes? I also see the maximal frequency is still lower for the Orin 32 GB (930.75 MHz vs 1377 MHz) so I’m a little bit confused regarding the real impact.

About the pin compatibility, I’m talking about using Nvidia’s full developer’s kit so I don’t see how it could be a problem. Are you not releasing a dev’s kit for the Orin NX?

5

Regarding the performance at differen watt, you cna check this blog Delivering Server-Class Performance at the Edge with NVIDIA Jetson Orin | NVIDIA Technical Blog to find those comparision table TOPS to do the assessment.

Regarding the devkit, it’s for developemnt, can’t be used to launch as end product.

6

That performance per watt is exactly what I needed!

It is only for development purposes, so do you expect to launch the 16 GB module and the dev’s kit together?

7

I can’t comment this for product plan and launch schedule.
Please wait for our announcement.

8

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