[Support]: Running Coral TPU process on a different server from the GPU/Stream decode

[velaar]

Describe the problem you are having

As formulated in the subject line. I have Coral TPUs in an M.2 format available on Server A with no additional expansion slots. I have an Nvidia GPU in a Server B that can be used for stream decoding and movement detection. Is this setup possible?

As the overhead of transferring 320x320 picture 2 times over 20 Gbps network should be insignificant at best.

Version

Not applicable

Frigate config file

Not applicable

Relevant log output

Not applicable

FFprobe output from your camera

Not applicable

Frigate stats

No response

Operating system

UNRAID

Install method

Docker Compose

Coral version

M.2

Network connection

Wired

Camera make and model

Not applicable

Any other information that may be helpful

No response

[blakeblackshear]

It's not possible, and it will be more significant than you think. If copying the data and sending over the network adds 10ms, then you might as well just use openvino or tensorrt on the same machine.

[velaar]

I see. Thank you for the clarification. I will explore options for bringing in a hardware decoding to the server. Sadly can't find any M.2 or usb solutions for that.

I will close the issue for safekeeping and hope it helps someone!