[Detector Support]: Coral is very slow (219.85 ms) - any ideas why?

[Sauws]

Describe the problem you are having

Hi all,

Thanks for taking the time to read this.
I've searched the web but the people who had the same issue seem to have the USB version of Coral with the problems that come along with that (cable, power, port) or somehow the PCIE version overheating.

For me the temperature seems reasonable so I would like to double check if there are any indications where I'm making a mistake (logs or config). My setup is a Dell Thinclient with an Intel Celeron processor which has a PCIE slot on the motherboard, the Coral TPU is detected by Proxmox and successfully passed through into the LXC which I also works inside the Frigate LXC (/dev/apex_0 exists).

I'm using a Reolink RLC-811A camera. Fully aware now that both (LXC in Proxmox via the Tteck script+ Reolink camera) are not recommended, its the HW I had before I started on this journey, might this be the reason its slow than thats it. But can you take a look to see if there are any obvious mistakes to why its so slow?

thanks

Version

System 0.14.1-

Frigate config file

mqtt:
  enabled: false

detectors:
  coral:
    type: edgetpu
    device: pci 

cameras:
  Cam1:
    audio:
      enabled: false

    ffmpeg:
      hwaccel_args: preset-vaapi
      inputs:
        - path: rtsp://user:[email protected]:554//h264Preview_01_sub
          roles:
            - detect
        - path: rtsp://user:[email protected]:554//h265Preview_01_main
          roles:
            - record
    motion:
      mask:
        - 0.371,0.256,0.297,0.49,0.242,0.677,0.214,0.795,0.205,0.847,0.233,0.904,0.28,0.951,0.279,1,0,1,0,0.5,0.33,0.06,0.364,0.069
        - 0.56,0.914,0.589,0.914,0.589,0.949,0.56,0.949
        - 0.744,0,0.771,0.183,0.787,0.347,0.796,0.37,0.805,0.39,1,0.399,1,0
    zones:
      Zone1:
        coordinates: 
          0.645,0.996,0.279,1,0.279,0.95,0.23,0.897,0.206,0.845,0.214,0.795,0.242,0.684,0.3,0.48,0.372,0.254,0.371,0.238,0.402,0.22,0.483,0.191,0.541,0.177,0.543,0.198,0.565,0.193,0.562,0.316,0.585,0.477,0.617,0.728
        loitering_time: 0
        inertia: 3
      Zone2:
        coordinates: 
          0.369,0.139,0.371,0.234,0.405,0.219,0.481,0.191,0.545,0.17,0.546,0.157,0.564,0.157,0.564,0.122,0.499,0.12,0.463,0.121,0.434,0.125,0.399,0.129
        loitering_time: 0
      Zone3:
        coordinates: 1,0.699,0.871,0.748,0.844,0.847,0.88,0.999,0.999,0.997
        inertia: 3
        loitering_time: 0
    review:
      alerts:
        required_zones:
          - Zone 1
          - Zone 2
          - Zone 3

model:
  path: /cpu_model.tflite
version: 0.14

docker-compose file or Docker CLI command

I have used the default Proxmox helper script (formerly TTeck) to install this as an LXC inside Proxmox.

Relevant Frigate log output

2025-08-05 16:08:28.706539  [INFO] Preparing Frigate...
2025-08-05 16:08:28.706733  [INFO] Starting Frigate...
2025-08-05 16:08:31.235382  [2025-08-05 16:08:31] frigate.app                    INFO    : Starting Frigate (0.14.1-)
2025-08-05 16:08:31.235492  [2025-08-05 16:08:31] frigate.util.config            INFO    : Checking if frigate config needs migration...
2025-08-05 16:08:31.268769  [2025-08-05 16:08:31] frigate.util.config            INFO    : frigate config does not need migration...
2025-08-05 16:08:31.645329  [2025-08-05 16:08:31] frigate.util.services          INFO    : Automatically detected vaapi hwaccel for video decoding
2025-08-05 16:08:37.335512  [2025-08-05 16:08:37] peewee_migrate.logs            INFO    : Starting migrations
2025-08-05 16:08:37.337372  [2025-08-05 16:08:37] peewee_migrate.logs            INFO    : There is nothing to migrate
2025-08-05 16:08:37.349853  [2025-08-05 16:08:37] frigate.app                    INFO    : Recording process started: 12022
2025-08-05 16:08:37.363692  [2025-08-05 16:08:37] frigate.app                    INFO    : Recording process started: 12029
2025-08-05 16:08:37.372766  [2025-08-05 16:08:37] frigate.app                    INFO    : go2rtc process pid: 131
2025-08-05 16:08:37.420052  [2025-08-05 16:08:37] detector.coral                 INFO    : Starting detection process: 12050
2025-08-05 16:08:37.434022  [2025-08-05 16:08:37] frigate.detectors.plugins.edgetpu_tfl INFO    : Attempting to load TPU as pci
2025-08-05 16:08:37.434131  [2025-08-05 16:08:37] frigate.app                    INFO    : Output process started: 12052
2025-08-05 16:08:37.438015  [2025-08-05 16:08:37] frigate.detectors.plugins.edgetpu_tfl INFO    : TPU found
2025-08-05 16:08:37.457497  [2025-08-05 16:08:37] frigate.app                    INFO    : Camera processor started for Oprit: 12077
2025-08-05 16:08:37.475833  [2025-08-05 16:08:37] frigate.app                    INFO    : Capture process started for Oprit: 12083

Install method

Proxmox via TTeck Script

Object Detector

Coral

Screenshots of the Frigate UI's System metrics pages

Any other information that may be helpful

No response

[inkeep-ai-assistant[bot]]

I see several potential issues with your Coral TPU performance. Your 219.85 ms inference time is significantly slower than expected for a Coral TPU, which should typically achieve 5-10 ms inference speeds (1).

The main issue I notice in your configuration is that you're using the wrong model path. You have:

model:
  path: /cpu_model.tflite

(2)

This is the CPU model, not the EdgeTPU model. For EdgeTPU detectors, you should be using the EdgeTPU-optimized model. The Edge TPU detector uses a TensorFlow Lite model at /edgetpu_model.tflite by default (3). Remove the model path specification entirely to use the default EdgeTPU model, or explicitly set it to:

model:
  path: /edgetpu_model.tflite

(3)

Additionally, there are several common causes for PCIe Coral performance issues:

1. Driver Installation: The most common reason for PCIe Coral detection issues is that the driver has not been installed (2). The process varies based on OS and kernel. For most cases, the Coral documentation shows how to install the driver for PCIe-based Coral (2).

2. Device Communication: Running Frigate in a VM may cause communication with the device to be lost and need to be reset (2). Since you're using Proxmox LXC, this could be affecting performance.

3. Device Access: When running in VMs or containers, you must ensure both device IDs are mapped correctly (2). The USB coral has different IDs when uninitialized (1a6e:089a Global Unichip Corp.) and initialized (18d1:9302 Google Inc.) (2).

Your logs show the TPU is detected (TPU found), but using the CPU model instead of the EdgeTPU model would explain the extremely slow inference times. Fix the model path first, and if issues persist, verify the PCIe driver installation in your Proxmox environment.

📚 Sources:

1. Recommended hardware
2. Troubleshooting EdgeTPU
3. Supported Hardware

---

Have another question? Just tag @inkeep.

[Sauws]

The main issue I notice in your configuration is that you're using the wrong model path. You have:

model:
  path: /cpu_model.tflite

(2)

This is the CPU model, not the EdgeTPU model. For EdgeTPU detectors, you should be using the EdgeTPU-optimized model. The Edge TPU detector uses a TensorFlow Lite model at /edgetpu_model.tflite by default (3). Remove the model path specification entirely to use the default EdgeTPU model, or explicitly set it to:

model:
  path: /edgetpu_model.tflite

Yeah, this is very impressive by the AI (to me). But yes, this was my mistake, its at 8MS now. Many thanks who maintains the AI.

[wjdp]

@Sauws & Mr AI bot, thanks for this post. I was about to make a new discussion but spotted this one.

In my case I'm using a plus model where the config looks like

model:
  path: plus://<hex>

It's not clear without checking the plus website which hash is CPU or edgetpu and I'd copied the wrong one 🤦