Using PCIe Endpoint Mode on Orin for GPU

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

When using an Orin “A” in Endpoint Mode, has anyone done any work or created a library to allow the root machine (which might be a different Orin, “B”) to run CUDA code on Orin “A” but have the CUDA driver resident on Orin “B”?

We could create some kind of user-mode process which, on Orin “A” receives a networking packet from Orin “B” and dispatches some work to CUDA on Orin “B”. I’m guessing that’s very slow due to context switching/etc.

I’m wondering if there is a fast & supported method here?

2

Please refer to PCIe Endpoint Mode — NVIDIA Jetson Linux Developer Guide 1 documentation to see if able to help on your case.

4

Hi Kay - I have checked the link above and I’m hoping for a software library or driver which allows receiving commands from the root device and running the commands on the local Orin GPU. I think we have the hardware portion figured out.

I’m investigating what level of support might exist for dispatching commands on behalf of the root device to the endpoint’s GPU.

5

A potential solution might also be some software method/library/driver to remotely host GPUs in a cluster and return results back to the root node.

6

Hi,
This setup is not supported. It may work to connect two AGX Orin boards through PCIe. One runs in RP and the other runs in EP for communication. And tasks on each GPU are run individually. Two Orin hosts are run individually and can communicate through PCIe.

IGX developer kit may be closer to the setup. It is a system with Orin SoC + one dGPU card(such as RTX6000 ADA card)

8

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