/frigate-configurator

Frigate Object Detector Comparison

Detector Types Overview

Detailed Comparison

USB Coral TPU

Pros:

• Best performance per watt
• Simple USB connection
• Works on any system with USB
• Handles 10+ cameras easily

Cons:

• Can be hard to find in stock
• USB bandwidth can be limiting with many cameras
• Generates some heat

Configuration:

detectors:
  coral:
    type: edgetpu
    device: usb

PCIe/M.2 Coral TPU

Pros:

• Same performance as USB
• No USB bandwidth limitations
• Lower latency
• Dual-edge TPU available for double throughput

Cons:

• Requires PCIe slot or M.2 slot
• Harder to install
• Limited to desktop/server systems

Configuration:

detectors:
  coral:
    type: edgetpu
    device: pci

OpenVINO (Intel)

Pros:

• Uses existing Intel GPU (free)
• Good performance on modern Intel CPUs
• Supports multiple model architectures

Cons:

• Requires Intel 6th Gen (Skylake) or newer
• GPU mode requires compatible integrated graphics
• CPU mode is inefficient

Configuration:

detectors:
  ov:
    type: openvino
    device: GPU  # or CPU for testing

ONNX Runtime

Pros:

• Automatically uses available GPU acceleration
• Works with AMD ROCm, Intel OpenVINO, NVIDIA TensorRT
• Flexible model support (YOLO variants)

Cons:

• Requires compatible GPU drivers
• Performance varies by hardware

Configuration:

detectors:
  onnx:
    type: onnx
    device: auto  # Automatically selects best available

TensorRT (NVIDIA Jetson)

Pros:

• Optimized for Jetson hardware
• Low power consumption
• Good for edge deployments

Cons:

• Only works on Jetson devices
• Requires model preprocessing on target hardware
• Limited to NVIDIA ecosystem

Configuration:

detectors:
  tensorrt:
    type: tensorrt
    device: 0

Hailo-8L (Raspberry Pi 5 AI Kit)

Pros:

• Designed for Raspberry Pi 5
• Low power consumption
• Good performance for edge

Cons:

• Limited to specific hardware
• Newer, less tested

Configuration:

detectors:
  hailo:
    type: hailo8l

CPU Detector

Pros:

• Works on any hardware
• No additional hardware needed

Cons:

• Very high CPU usage
• Not suitable for production
• Slow inference times

Configuration:

detectors:
  cpu:
    type: cpu
    num_threads: 3

Capacity Guidelines

Important Notes

1. Cannot mix detector types for object detection - You can't use Coral for some cameras and OpenVINO for others. Pick one detector type for all object detection.

2. Other tasks can use different hardware - Semantic search, face recognition, and other features can use different hardware than the main object detector.

3. USB bandwidth matters - If using multiple USB Coral TPUs, ensure they're on different USB controllers.

4. Temperature affects performance - Coral TPUs throttle at high temperatures. Consider cooling for sustained workloads.

Recommendations