Google Coral Edge TPU
Vendor: Google
Category: TPU Tensor Processing Unit
Architecture: Edge TPU
Introduction
Google Coral Edge TPU is a low-power AI accelerator designed specifically for edge devices. Available in multiple form factors including USB accelerator, M.2 module, and Dev Board, for real-time inference of TensorFlow Lite models.
Specifications
| Model | Compute | Memory | Interface | TDP | Process |
|---|---|---|---|---|---|
| Coral Edge TPU | 4 TOPS (INT8) | External DDR | USB 3.0 / M.2 / PCIe | 2W | 9nm |
| Dev Board Micro | 4 TOPS (INT8) | 512MB DDR | USB-C | 2W | 28nm |
Official Website
Driver Downloads
Linux
Related Documentation
OS Support
| Windows | Linux | macOS | Android |
|---|---|---|---|
| ✅ | ✅ | ✅ | ✅ |
Version History
| Version | Release Date | Description |
|---|---|---|
| Edge TPU Compiler 16 | 2024 | Model quantization optimization |
Performance Benchmarks
| Model | Task | Performance Metric |
|---|---|---|
| Coral Edge TPU | MobileNet V2 | ~400 fps |
| Coral Edge TPU | YOLOv5s | ~30 fps (quantized) |
| Dev Board Micro | Multi-model inference | Supports concurrent multi-model |
Pricing Information
| Model | Reference Price | Notes |
|---|---|---|
| Coral USB Accelerator | ~$60 | USB 3.0 accelerator dongle |
| Coral Dev Board | ~$150 | Standalone development board |
| Coral M.2 | ~$35 | M.2 module |
Quick Setup
Linux / Windows / macOS / Raspberry Pi
# 1. Install Edge TPU Runtime
sudo apt-get install -y libedgetpu1-std
# 2. Install Python SDK
pip install tflite-runtime
# 3. Compile model to TFLite + Edge TPU format
# Using Edge TPU Compiler
edgetpu_compiler model.tflite
Code Examples
Python (Edge TPU)
from tflite_runtime.interpreter import Interpreter
# Load Edge TPU compiled model
interpreter = Interpreter(
model_path="model_edgetpu.tflite",
experimental_delegates=[load_delegate("libedgetpu.so")]
)
interpreter.allocate_tensors()
# Inference
input_details = interpreter.get_input_details()
interpreter.set_tensor(input_details[0]['index'], input_data)
interpreter.invoke()
output = interpreter.get_tensor(interpreter.get_output_details()[0]['index'])
Architecture Highlights
- Edge TPU: Coprocessor designed specifically for INT8 quantized inference, with only 2W power consumption
- Cross-Platform Support: The only AI accelerator that simultaneously supports Windows/Linux/macOS/Android/Raspberry Pi
- Coral Ecosystem: Google Coral offers a complete hardware lineup (USB/M.2/PCIe/Dev Board) and software toolchain
Model Compatibility
| Model/Framework | Support Status | Notes |
|---|---|---|
| TFLite | ✅ Native | Mainstream format |
| ONNX | ⚠️ | Use after converting to TFLite |
| YOLO / MobileNet | ✅ Officially Optimized | Most mature scenario |
| Image Classification/Detection | ✅ | Official pre-compiled models |
| LLM | ❌ | Only small models supported |
Related Products
If you are evaluating alternatives, the following products may also fit your scenario:
- Hailo-8 / Hailo-15 — Hailo (NPU Neural Processing Unit)
- TsingMicro TX510 — TsingMicro (ASIC dedicated accelerator)
- Intel NPU (Neural Processing Unit) — Intel (NPU Neural Processing Unit)
- AMD Ryzen AI NPU — AMD (NPU Neural Processing Unit)
- Qualcomm Hexagon NPU — Qualcomm (NPU Neural Processing Unit)
- Google Cloud TPU — Google (TPU Tensor Processing Unit)