Skip to main content

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

ItemSpec
ArchitectureEdge TPU ASIC (Google proprietary edge inference)
ProcessEstimated 28nm
TOPS4 TOPS INT8 (single chip) / 8 TOPS (Dual TPU module)
Energy Efficiency2 TOPS/W (0.5W/TOPS)
Data TypesINT8 fixed-point (no FP16/FP32 training)
Model FormatTensorFlow Lite (only, PyTorch/ONNX require conversion)
MobileNet V2 Performance400 FPS (INT8 quantized)
PCIe InterfacePCIe Gen 2 x1 (single TPU) / 2× PCIe Gen 2 x1 (dual TPU)
USB InterfaceUSB 3.0 (Coral USB Accelerator)
M.2 Form FactorA+E key (M.2-2230, 22×30mm) or B+M key (M.2-2280, 22×80mm)
Mini PCIe30×26.8 mm half-size
PCIe CardStandard PCIe short card
Dual M.222×30 mm M.2-2230-D3-E (dual TPU sharing M.2 slot)
Coral Dev BoardNXP i.MX 8M SoC (4-core A53) + Edge TPU + GPU + Wi-Fi + Bluetooth
Operating Temperature0-50°C (Coral USB) / -40-85°C (industrial)
Operating SystemsLinux (Ubuntu, Raspbian, Linux Mendel for Dev Board), Windows 10, macOS
Launch DateMarch 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

MetricEdge TPU CoralJetson NanoNote
TOPS (INT8)40.47 (FP16)8.5×
Energy Efficiency (TOPS/W)20.05 (FP16)40×
Price$24.99$149 (4GB)6× cheaper
Power0.5W5-10W10× lower
Model FormatTF LiteTF/PyTorch/ONNXTPU limited
Training SupportNoneSupportedNano general-purpose
InterfaceM.2/USB/PCIeSoC (full board)Coral accelerator card
MLPerf MobileNet V2400 FPS64 FPS6.3×

Measured Performance

ModelEdge TPU (4 TOPS)Performance (FPS)Latency (ms)
MobileNet V1 (224)4 TOPS700+< 2ms
MobileNet V2 (224)4 TOPS4002.5ms
MobileNet V3-Small4 TOPS2504ms
EfficientNet B04 TOPS1307.7ms
Inception V44 TOPS2343ms
ResNet-50 (v1)4 TOPS5020ms
SSD MobileNet V2 (object detection)4 TOPS6017ms
PoseNet (pose estimation)4 TOPS10010ms
DeepLab V3 (semantic segmentation)4 TOPS3033ms

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

ItemDetail
VendorGoogle LLC (Mountain View, USA)
Product BrandCoral (coral.ai)
DesignGoogle Research edge AI team
FoundryEstimated 28nm TSMC or Samsung
Software StackTensorFlow Lite Runtime, Edge TPU Compiler (quantize TFLite → Edge TPU compilation)
AI FrameworkTensorFlow / TensorFlow Lite (no native PyTorch/ONNX support)
Price TiersUSB $59.99 / M.2/Mini PCIe $24.99 / Dual M.2 $49.99 / Dev Board $149.99
Availability30+ countries: US, Canada, EU, Japan, Korea, Taiwan, India, etc.
Coral Dev BoardEOL 2024 (replaced by Raspberry Pi + USB solution)
2025 StatusStill 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)