Hailo-8 / Hailo-8L / Hailo-15 Edge AI Accelerator
Overview
Hailo is an edge AI inference accelerator from Israel's Hailo Technologies, designed for computer vision and multimodal LLM. Hailo-8 (released 2019, 26 TOPS) has been commercially deployed as an alternative to NVIDIA Jetson Orin, in Raspberry Pi AI Kits, smart cameras, industrial vision, and more.
The latest product is Hailo-15 (released 2024, integrating VPU + AI accelerator, single-chip vision AI processor).
Product Line Comparison
| Item | Hailo-8 | Hailo-8L | Hailo-15-15 | Hailo-15-30 |
|---|---|---|---|---|
| INT8 Compute | 26 TOPS | 13 TOPS | 20 TOPS | 40 TOPS |
| Process | TSMC 14nm | TSMC 14nm | TSMC 7nm | TSMC 7nm |
| TDP | 2.5 W | 1.5 W | 6 W | 9 W |
| Efficiency | 10.4 TOPS/W | 8.7 TOPS/W | 3.3 TOPS/W | 4.4 TOPS/W |
| Memory | 4GB LPDDR4 | 4GB LPDDR4 | Integrated VPU + 4GB | Integrated VPU + 8GB |
| Form Factor | M.2 / mPCIe / BGA | M.2 / BGA | SoC (incl. ISP/VPU) | SoC |
| Launch | 2019 | 2020 | 2024 | 2024 |
Hailo-15 integrates ISP (Image Signal Processor) + VPU (Video Encode/Decode) + AI accelerator, replacing multi-chip solutions with a single chip.
Architecture Highlights
Dataflow Architecture
- Cerebras WSE-like spatial architecture, weights resident in on-chip SRAM
- No external memory access latency (CV models fully resident)
- Deterministic latency: Single-image inference latency < 5 ms (ResNet-50)
Model Compression
- Hailo Dataflow Compiler auto-quantization (INT8/INT4/mixed precision)
- Supports pruning, layer fusion, memory optimization
- Model compression rate ~3-5x (FP32 -> INT8) with < 1% accuracy loss
Multi-Model Parallelism
- Single card can run multiple models simultaneously (e.g., detection + classification + segmentation)
- Hardware resource partitioning with strong isolation
Software Ecosystem
Toolchain
- Hailo Dataflow Compiler (HDF -> HAR conversion)
- Hailo Model Zoo (30+ pre-trained models, including YOLOv5/v8, ResNet, EfficientNet)
- HailoRT (runtime, C/C++/Python API)
- Hailo Python API (gRPC-style)
Frameworks
- TensorFlow / PyTorch / ONNX (export then compile)
- GStreamer (video stream integration)
- ROS (robotics)
Operating Systems
- Linux (ARM/x86)
- Windows (select models)
- Integrated with NVIDIA JetPack (GPU acceleration alternative)
Deployment Examples
- Hailo-8 M.2 + Raspberry Pi 5 — 13 TOPS AI Kit ($70)
- Smart cameras (Axis, Qihoo 360) — edge video analytics
- Industrial vision (production line defect detection)
- Autonomous driving (Tier 1 supplier ADAS forward-facing)
- Robotics (service robot vision SLAM)
Performance Benchmarks (YOLOv5s 640x640)
| Device | FPS | Latency | TDP |
|---|---|---|---|
| Hailo-8 | 600 FPS | 1.7 ms | 2.5 W |
| Hailo-8L | 300 FPS | 3.3 ms | 1.5 W |
| NVIDIA Jetson Orin Nano 8GB | 125 FPS | 8 ms | 7-15 W |
| NVIDIA Jetson Xavier NX | 90 FPS | 11 ms | 10-20 W |
| CPU (i7-12700H) | 25 FPS | 40 ms | 45 W |
Hailo-8 on YOLOv5s is 6.7x faster than Jetson Xavier NX, with 4-8x lower power.
Vendor Information
| Item | Details |
|---|---|
| Vendor | Hailo Technologies (Israel) |
| Product Page | https://hailo.ai/ |
| Pricing | Hailo-8 M.2: $200 / Hailo-8L: $150 / Hailo-15 Kit: $300+ |
| Funding | Series C $136M (2023-04, total $340M) |
| Target Market | Edge AI, smart cameras, industrial vision, robotics |
Use Cases
- ✅ Edge vision AI inference (cameras, robots)
- ✅ Industrial vision (production line inspection)
- ✅ ADAS (Advanced Driver Assistance Systems)
- ✅ Edge LLM (Phi-3, Llama 2 7B quantized)
- ❌ Large model training
- ❌ Data center inference (NV H100 better perf/watt)
Related Cards
- NVIDIA Jetson Orin - Comparable edge AI
- NVIDIA Jetson Thor - Next-gen edge
- Google TPU v6e - Data center comparison