Is it possible to have a workstream supporting multi-device inference using both the Xaviers A and B on DRIVE AGX? We’re possibly running into some speed issues while running an application on Xavier A itself, and were wondering if we can also incorporate Xavier B in our inference pipeline.
Running at lower precisions (FP16 or INT8) and utilizing TensorRT is also being looked at in the meantime, but would love to know if this parallel workstream is supported or not.
Dear raul_91,
The reason why the design is divided into Xaiver A / B is because of redundancy.
Regarding connection of XaiverA and B, please refer to the Topology of the following link document.
https://docs.nvidia.com/drive/active/5.1.0.2L/nvvib_docs/index.html#page/DRIVE_OS_Linux_SDK_Development_Guide%2FNetwork%2520Setup%2Fnetwork_topology_av.html%23
Since XavierA / B is connected to PCI switch, I am not sure if you can solve the speed issue you are talking about.
Rather than using multi devices, I’d like to suggest to do DL network or pipeline optimization rather than multi-divce. Thanks.