Google Edge TPU (Coral 4 TOPS Edge Inference)
Overview
Google Edge TPU (branded Coral) is an edge AI inference ASIC released by Google in March 2019, designed specifically for TensorFlow Lite model inference on edge devices with low power consumption and low latency. Delivering 4 TOPS INT8 at 2 TOPS/W (0.5W/TOPS), it comes in multiple form factors: Coral USB Accelerator (USB 3.0 dongle), M.2 Accelerator A+E key (22×30mm), M.2 Accelerator B+M key (22×80mm), Mini PCIe Accelerator (30×26.8mm), PCIe Accelerator (standard PCIe card), Dual Edge TPU M.2 (2× 4 TOPS = 8 TOPS), and Dev Board (complete SoC + Edge TPU Raspberry Pi-sized development board). The Coral USB Accelerator at $59.99 is the world's most affordable and accessible entry-level AI inference device. It achieves 400 FPS on MobileNet V2 (at 0.5W/TOPS).
Core Specifications
| Item | Spec |
|---|---|
| Architecture | Edge TPU ASIC (Google proprietary edge inference) |
| Process | Estimated 28nm |
| TOPS | 4 TOPS INT8 (single chip) / 8 TOPS (Dual TPU module) |
| Energy Efficiency | 2 TOPS/W (0.5W/TOPS) |
| Data Types | INT8 fixed-point (no FP16/FP32 training) |
| Model Format | TensorFlow Lite (only, PyTorch/ONNX require conversion) |
| MobileNet V2 Performance | 400 FPS (INT8 quantized) |
| PCIe Interface | PCIe Gen 2 x1 (single TPU) / 2× PCIe Gen 2 x1 (dual TPU) |
| USB Interface | USB 3.0 (Coral USB Accelerator) |
| M.2 Form Factor | A+E key (M.2-2230, 22×30mm) or B+M key (M.2-2280, 22×80mm) |
| Mini PCIe | 30×26.8 mm half-size |
| PCIe Card | Standard PCIe short card |
| Dual M.2 | 22×30 mm M.2-2230-D3-E (dual TPU sharing M.2 slot) |
| Coral Dev Board | NXP i.MX 8M SoC (4-core A53) + Edge TPU + GPU + Wi-Fi + Bluetooth |
| Operating Temperature | 0-50°C (Coral USB) / -40-85°C (industrial) |
| Operating Systems | Linux (Ubuntu, Raspbian, Linux Mendel for Dev Board), Windows 10, macOS |
| Launch Date | March 2019 (Coral Dev Board) / July 2019 (Coral USB Accelerator) |
| Coral USB Accelerator | $59.99 |
| Coral M.2 Accelerator A+E | $24.99 |
| Coral M.2 Accelerator B+M | $24.99 |
| Coral Mini PCIe Accelerator | $24.99 |
| Coral Dual Edge TPU M.2 | $49.99 |
| Coral Dev Board | $149.99 (discontinued, 2024) |
Comparison with NVIDIA Jetson Nano
| Metric | Edge TPU Coral | Jetson Nano | Note |
|---|---|---|---|
| TOPS (INT8) | 4 | 0.47 (FP16) | 8.5× |
| Energy Efficiency (TOPS/W) | 2 | 0.05 (FP16) | 40× |
| Price | $24.99 | $149 (4GB) | 6× cheaper |
| Power | 0.5W | 5-10W | 10× lower |
| Model Format | TF Lite | TF/PyTorch/ONNX | TPU limited |
| Training Support | None | Supported | Nano general-purpose |
| Interface | M.2/USB/PCIe | SoC (full board) | Coral accelerator card |
| MLPerf MobileNet V2 | 400 FPS | 64 FPS | 6.3× |
Measured Performance
| Model | Edge TPU (4 TOPS) | Performance (FPS) | Latency (ms) |
|---|---|---|---|
| MobileNet V1 (224) | 4 TOPS | 700+ | < 2ms |
| MobileNet V2 (224) | 4 TOPS | 400 | 2.5ms |
| MobileNet V3-Small | 4 TOPS | 250 | 4ms |
| EfficientNet B0 | 4 TOPS | 130 | 7.7ms |
| Inception V4 | 4 TOPS | 23 | 43ms |
| ResNet-50 (v1) | 4 TOPS | 50 | 20ms |
| SSD MobileNet V2 (object detection) | 4 TOPS | 60 | 17ms |
| PoseNet (pose estimation) | 4 TOPS | 100 | 10ms |
| DeepLab V3 (semantic segmentation) | 4 TOPS | 30 | 33ms |
Limitation: The Edge TPU only supports INT8 quantized models in TensorFlow Lite format. PyTorch / ONNX / MXNet native models are not supported (must be converted to TFLite first). Training is not supported, inference only.
Use Cases
- Embedded device prototyping (Coral USB Accelerator $60 plug-and-play)
- Industrial IoT edge gateways (M.2 form factor integrated into industrial PCs)
- Smart home (home voice assistants, pet recognition, motion detection)
- Retail smart cameras (footfall counting, heat map analysis)
- Smart cameras (person detection, license plate recognition, intrusion alerts)
- Wearable devices (low-power AI inference)
- Education/maker (Raspberry Pi 4 + Coral USB = complete AI development platform)
- Smart agriculture (pest/disease recognition, fruit counting)
- Assisted driving (blind spot detection, driver monitoring assistance)
Vendor Information
| Item | Detail |
|---|---|
| Vendor | Google LLC (Mountain View, USA) |
| Product Brand | Coral (coral.ai) |
| Design | Google Research edge AI team |
| Foundry | Estimated 28nm TSMC or Samsung |
| Software Stack | TensorFlow Lite Runtime, Edge TPU Compiler (quantize TFLite → Edge TPU compilation) |
| AI Framework | TensorFlow / TensorFlow Lite (no native PyTorch/ONNX support) |
| Price Tiers | USB $59.99 / M.2/Mini PCIe $24.99 / Dual M.2 $49.99 / Dev Board $149.99 |
| Availability | 30+ countries: US, Canada, EU, Japan, Korea, Taiwan, India, etc. |
| Coral Dev Board | EOL 2024 (replaced by Raspberry Pi + USB solution) |
| 2025 Status | Still purchasable, but Google's AI focus has shifted to Cloud TPU + Gemini Nano on-device cloud collaboration |
Key Features
- 4 TOPS INT8 single chip / 8 TOPS dual TPU module
- 2 TOPS/W excellent energy efficiency (0.5W/TOPS)
- 6 product form factors (USB / M.2 A+E / M.2 B+M / Mini PCIe / PCIe / Dual M.2)
- Plug-and-play (Linux auto-detects as PCIe device)
- TensorFlow Lite first-class citizen (best compatibility)
- Coral USB $59.99 (world's most affordable AI inference device)
- Low power (entire board 2-5W)
- Industrial grade (-40-85°C operating temperature, industrial version)
- Complete ecosystem (coral.ai provides model library + compiler + runtime)
- PCIe Gen 2 x1 (standard interface, works even on legacy industrial PCs)
Related Cards
- Google TPU v4 (Data Center Training) — Same-vendor cloud training
- Google TPU v5e (Data Center Training-Lite) — Training lite
- Google TPU v6e (DC Inference) — Data center inference
- Google TPU 8i (DC Inference Flagship) — Latest inference
- Hailo-15 (7-20 TOPS Smart Camera) — Competitor
- Hailo-8L (13 TOPS Edge) — Same-gen competitor
- NVIDIA Jetson Orin (40-275 TOPS Edge AI) — High-end edge comparison
- Blaize Xplorer (160 TOPS) — Edge competitor
- Mobilint Regulus (32 TOPS Korea) — Same-tier edge
- Architecture: TPU Edge Edition — Edge TPU concept
- Coral Official