Machine learning on a Raspberry Pi — it sounds like something that would only become feasible in a few years. But that moment is already here. Google's Coral USB Accelerator makes it possible to run trained ML models locally, without an internet connection, without the cloud, and with a processing speed that a Raspberry Pi alone could never achieve.
What is the Coral USB Accelerator?
The Coral USB Accelerator is a USB accessory that contains an Edge TPU — a chip from Google designed specifically for machine learning inference. Inference means applying an already trained model to new data, not training the model itself.
| Specification | Value |
|---|---|
| Processing speed | 4 TOPS (tera operations per second) |
| Power consumption | 2 W via USB |
| Connection | USB 3.0 (also usable via USB 2.0) |
| Compatible with | Raspberry Pi, Linux x86, macOS, Windows |
What is the difference compared to inference on the CPU?
A Raspberry Pi 4 can run TensorFlow Lite models on the CPU, but with more complex models — image classification, object detection, facial recognition — inference times can increase to 500 ms or more. The Coral moves those calculations to the Edge TPU: for optimized models, inference time drops to just a few milliseconds — the difference between barely real-time and 30 frames per second.
What can you use it for in practice?
- Image classification — a camera recognizes objects, anomalies, or animals in real time, without an external server.
- Object detection — presence detection, counting cameras, surveillance systems.
- Keyword recognition — run audio classification and wake-word detection locally.
- Anomaly detection — analyze sensor data for deviations to support predictive maintenance.
Limitations
- Only TensorFlow Lite, compiled for the Edge TPU. Large models are partly executed on the Edge TPU and partly on the CPU.
- Inference, not training. Intended only for running models.
- Limited model size. The Edge TPU has 8 MB of on-chip SRAM.
Getting started
Google and Coral offer pre-trained models that work directly with the Edge TPU: MobileNet, EfficientDet, PoseNet, and more. The PyCoral library provides access from Python, with example scripts that can be used right away.
Ask your question