Hailo-8 / Hailo-15
Vendor: Hailo
Architecture: Dataflow
Category: NPU Neural Processor
Introduction
Global leader in edge AI computing. Delivers up to 26 TOPS at extremely low power (just a few watts), widely used in autonomous driving, robotics, and smart camera applications.
Specifications
| Model | Compute | Memory | Interface | TDP | Process |
|---|---|---|---|---|---|
| Hailo-8 | 26 TOPS (INT8) | External DDR | PCIe 2.1 / M.2 | 2.5W | 16nm |
| Hailo-15 | 42 TOPS (INT8) | Internal SRAM | PCIe 3.0 | 11W | 10nm |
| Hailo-8L | 13 TOPS (INT8) | External DDR | M.2 | 1.5W | 22nm FDX |
Official Website
Driver Downloads
Linux
Related Documentation
OS Support
| Windows | Linux | macOS | Android |
|---|---|---|---|
| ✅ | ✅ | ❌ | ✅ |
Version History
| Version | Release Date | Description |
|---|---|---|
| HailoRT 4.x | 2024 | Hailo-15 support + multi-model orchestration |
| Dataflow Compiler 3.15 | 2024 | YOLOv8 optimization |
Performance Benchmarks
| Model | Task | Performance Metric |
|---|---|---|
| Hailo-8 | YOLOv8 Inference | ~26 TOPS, low latency |
| Hailo-8 | ResNet-50 | ~700 fps |
| Hailo-8L | Edge Detection | 13 TOPS, ultra-low power |
| Hailo-15 | Multi-channel Video Analytics | 20 TOPS, with ISP |
Pricing
| Model | Reference Price | Notes |
|---|---|---|
| Hailo-8 | ~$99-149 | Modular accelerator |
| Hailo-8L | ~$49-79 | Entry-level |
| Hailo-15 | Contact vendor | Chip-level (ODM-oriented) |
Quick Installation
Linux / Raspberry Pi
# 1. Install Hailo SDK
pip install hailo-sdk
# 2. Compile model to Hailo format (HEF)
hailo compiler model.onnx --output model.hef
# 3. Inference
hailo run model.hef --input image.jpg
Hailo-8 can be connected to x86 or ARM devices via M.2 / PCIe accelerator stick.
Code Examples
Python (Hailo Runtime)
from hailo_platform import HailoPlatform
# Initialize Hailo device
platform = HailoPlatform()
platform.init()
# Load compiled HEF model
platform.load_hef("yolov8.hef")
# Inference
output = platform.run(np_input)
Architecture Highlights
- Dataflow Architecture: Hailo proprietary dataflow processing architecture, optimized for CNN inference, no external DRAM required
- Ultra-Low Power: Hailo-8 only 2.5W, Hailo-8L only 1W, suitable for battery-powered devices
- Edge Positioning: Specializing in edge AI scenarios such as cameras, robots, and drones
Model Compatibility
| Model/Framework | Support | Notes |
|---|---|---|
| ONNX | ✅ Compile | Mainstream model format |
| TensorFlow | ✅ Compile | Compile after ONNX conversion |
| YOLO Series | ✅ Official Optimized | Most mature use case |
| Image Classification | ✅ | ResNet/MobileNet |
| LLM | ❌ | Not suitable for large model inference |
Related Products
If you're evaluating alternatives, the following products may also fit your scenario:
- AMD Ryzen AI NPU — AMD (NPU Neural Processor)
- Intel NPU (Neural Processing Unit) — Intel (NPU Neural Processor)
- Qualcomm Hexagon NPU — Qualcomm (NPU Neural Processor)
- Google Coral Edge TPU — Google (TPU Tensor Processor)
- Cambricon Siyuan 590 — Cambricon (ASIC Dedicated Accelerator)
- Groq LPU — Groq (LPU Language Processor)
- Nexchip Dragon Series — Nexchip (NPU Neural Processor)