Skip to main content

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

ItemHailo-8Hailo-8LHailo-15-15Hailo-15-30
INT8 Compute26 TOPS13 TOPS20 TOPS40 TOPS
ProcessTSMC 14nmTSMC 14nmTSMC 7nmTSMC 7nm
TDP2.5 W1.5 W6 W9 W
Efficiency10.4 TOPS/W8.7 TOPS/W3.3 TOPS/W4.4 TOPS/W
Memory4GB LPDDR44GB LPDDR4Integrated VPU + 4GBIntegrated VPU + 8GB
Form FactorM.2 / mPCIe / BGAM.2 / BGASoC (incl. ISP/VPU)SoC
Launch2019202020242024

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)

DeviceFPSLatencyTDP
Hailo-8600 FPS1.7 ms2.5 W
Hailo-8L300 FPS3.3 ms1.5 W
NVIDIA Jetson Orin Nano 8GB125 FPS8 ms7-15 W
NVIDIA Jetson Xavier NX90 FPS11 ms10-20 W
CPU (i7-12700H)25 FPS40 ms45 W

Hailo-8 on YOLOv5s is 6.7x faster than Jetson Xavier NX, with 4-8x lower power.

Vendor Information

ItemDetails
VendorHailo Technologies (Israel)
Product Pagehttps://hailo.ai/
PricingHailo-8 M.2: $200 / Hailo-8L: $150 / Hailo-15 Kit: $300+
FundingSeries C $136M (2023-04, total $340M)
Target MarketEdge 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)