AI Deployment on Google Coral Edge TPU
Google Coral — platform for highly efficient ML inference on edge. Edge TPU — specialized ASIC for INT8 inference. 4 TOPS at 0.5–2 W consumption. Ideal for battery-powered or low-power applications.
Coral Form Factors
- USB Accelerator: connects to Raspberry Pi / x86. Plug-and-play
- PCIe M.2 Accelerator (A+E): for embedded systems
- Dev Board: SoC NXP i.MX 8M + Edge TPU. Standalone edge computer
- Dev Board Mini: compact version
Optimal Applications
Object detection (MobileNet SSD, EfficientDet-Lite): MobileNet SSD on Coral USB → 400 FPS at 28 mW. Image classification: MobileNetV2 → 400 FPS. Pose estimation, face detection — Coral Model Zoo has models.
Model Requirements
Edge TPU executes only supported operations. Unsupported operations → CPU (slow). Rules for full TPU support: INT8 quantization, supported operations (Conv2D, DepthwiseConv, FC, BatchNorm, ReLU, etc.), model < 8 MB (else partial fallback reduces acceleration).
Workflow
- Train model (TensorFlow / TFLite)
- Post-training INT8 quantization with representative dataset
-
edgetpu_compiler model_quant.tflite - Deploy + PyCoral API
Limitations
Edge TPU efficient for standard CNN. For transformers, RNN, non-standard architectures — better Jetson or x86 with OpenVINO.