Help finding specs of motherboard PCI lanes on M2 connector for Coral TPU

Hi all.

I’ve been running HA on a small PC for about 3 years now, I’ve been using Frigate NVR and a USB Coral accelerator but need more TPUs and Frigate complains that “Coral is running slow”.

The PC is an MSI Cubi 5 10M i3 10110U, it has an M2 E-key slot with a WiFi/Bluetooth card installed which I don’t use. I’d like to use this slot for another Coral, preferably the dual TPU , but this requires a connector with “M.2 E-key (with two PCIe Gen2 x1 lanes)*”.

The only info I can find on the motherboard just lists Wifi and not specs on the slot itself.

Is there any way I can get the capabilities of this slot short of buying both modules and testing it? Are there any softwares out here that could probe the motherboard and check how many PCI lanes are connected on this slot?

Cheers

PC bios may show motherboard make/model I think. That may give more detailed spec.

USB coral shouldn’t be too slow

What is your inference speed?

How many camera?

Did you plug into USB-C port 3.x?

I think USB coral have about 20ms inference speed.

PCI dual is about 8-10ms x2.

Worst case only one tpu will work but it is still x2 faster than usb

Hi.

I’ll have a look at the BIOS if I can. The PC is headless and have only plugged in a monitor twice, once for initial setup and last year when HA updated and failed to boot.

I have 7 camera feeds and inference speeds are currently about 52ms.

The USB Coral is plugged in to a USB-A port 3.2 at the rear with the Coral supplied cable. Only the front has USB-C but I don’t have a short cable for this.

If I can’t confirm I can use a dual I’ll just go for a single TPU which will hopefully complement the USB one.

Cheers

Try long cable

52 is crazy slow. Doesn’t seem right.
Even at 20 with 7 cams you may have missed detections but it may be OK and you likely won’t notice

I’ve figured out why the inference speed is slow. My fault. :flushed:

I used a different model as the standard Frigate one was giving too many false positive ‘Person’ detections which was messing with automations so I used one from here which has a higher latency.

I’ve tried using different models but still get the same false positives, only the “efficientdet_lite0” model seems to work properly.

I’ll get an M.2 accelerator and see how both perform together.

was false positive > 75%
OR
was false positive < 70%

There are many things you can do to “tune” dectection.
And for automations I do find many false “person” detection < 72% occur so I generally only trigger automations when > 73-80% or so. Depends on camera view. Ultimately skipping a couple of bad detections doesnt matter since in many case the object is detected in multiple frames and if true one will be above xx%.


condition: and
conditions:
  - condition: time
    after: "21:00:00"
    before: "05:00:00"
  - condition: template
    value_template: "{{ trigger.payload_json[\"type\"] == \"new\" }}"
    enabled: true
  - condition: template
    value_template: "{{ trigger.payload_json[\"after\"][\"top_score\"] > 0.73 }}"
    enabled: true
enabled: true

USB is about 1/2 speed of single tpu pci. dual PCI is about 4 times the speed of USB

The false positives are high with the standard model. It constantly recognises garden furniture as a person with over 80-90% confidence :astonished:
This doesn’t happen with the new model but inference is higher.

I’ll try a second accelerator and try and find a better model.

Have you considered the possibility that people are disguising themselves as furniture and frigate has uncovered plot to take over your garden?

1 Like

:rofl: Perhaps. That would make training a new model interesting.