Also known as: Coral, Edge TPU
Google Coral is a hardware platform built around the Edge TPU, a small AI accelerator that runs machine-learning models locally rather than in the cloud.1
Overview
The Edge TPU is a cut-down tensor processing unit that executes TensorFlow Lite models very efficiently at low power. It is not a general processor: it accelerates the tensor math at the core of neural-network inference and little else, which is exactly why it can hit high throughput on a couple of watts. Models must be quantised to 8-bit integers and compiled for the TPU before they will run on it.
Coral ships in several forms so the same accelerator suits different projects: a full single-board computer (the Coral Dev Board), a solder-down module for embedding in a product, and a USB Accelerator stick that adds the TPU to an existing host such as a Raspberry Pi.2 The USB stick is the most common entry point, because it upgrades a board you already have rather than replacing it.
Coral vs Jetson
| Google Coral (Edge TPU) | NVIDIA Jetson | |
|---|---|---|
| Core | Edge TPU (fixed-function) | ARM CPU + CUDA GPU |
| Runs | Quantised TensorFlow Lite | Broad frameworks, general GPU |
| Power | Very low (~2 W) | Higher (5–60 W) |
| Flexibility | Narrow, supported models only | General-purpose accelerator |
| Best for | Cheap, fixed inference tasks | Heavier or varied ML workloads |
Where it fits
Coral targets edge AI: on-device vision, audio, and sensor inference where sending data to a server is too slow, too costly, or impossible. It is a more specialised choice than an NVIDIA Jetson — the Edge TPU runs supported quantised models fast and cheap, but it is not a general GPU, so anything outside that lane belongs elsewhere. In a signal-processing project, a Coral USB Accelerator on a Pi could classify or flag patterns in decoded data at the edge while the Pi itself handles the radio, keeping inference off the CPU that is busy demodulating.