Hailo-8 and Hailo-15: The Edge AI Accelerators Reshaping Industrial Computer Vision

TL;DR:

• Hailo-8 delivers 26 TOPS at 2.5W — enough to run YOLOv8n at 200+ FPS on a Raspberry Pi via M.2 HAT — without the cost or complexity of a Jetson Orin.
• Hailo-15 adds ISP integration for direct camera input, targeting smart camera and embedded vision applications at 40 TOPS.
• The Hailo AI Software Suite handles model compilation from standard ONNX/TFLite exports, making integration more accessible than comparable solutions.

Most discussions of edge AI inference default to NVIDIA Jetson for industrial applications or to microcontroller-class chips like the ESP32-P4 for embedded sensors. Hailo occupies a distinct middle tier: purpose-built neural processing units that deliver competitive inference performance within tight power and cost constraints, without the overhead of a full Linux system-on-module.

Since Hailo achieved silicon maturity with the Hailo-8, their chips have shipped in security cameras, traffic management systems, manufacturing quality control lines, and retail analytics platforms. The Raspberry Pi partnership — the Hailo-8L M.2 HAT ships as an official Pi 5 accessory — has dramatically expanded their developer reach beyond industrial OEMs.

The Hailo Chip Family

Hailo-8 (26 TOPS): The current mainstream offering. Available in M.2 form factor (B+M key and M key variants) and as a bare die for OEM integration. Power draw of 2.5W typical under inference load. Runs ResNet-50 at ~2,500 FPS, YOLOv8n at ~200 FPS, and more complex models like RT-DETR at 50–80 FPS. This is the chip in the Raspberry Pi AI Kit and the NVIDIA-compatible M.2 cards from several OEM board makers.

Hailo-8L (13 TOPS): A cost-reduced variant with half the compute of the Hailo-8. The official Raspberry Pi AI HAT+ uses the 8L variant at a lower price point. Suitable for simpler classification and detection tasks where the full 26 TOPS isn’t needed.

Hailo-15 (40 TOPS): Released for production in late 2024, the Hailo-15 adds an integrated Image Signal Processor (ISP) — allowing direct RAW sensor input without a separate image processing pipeline. This matters for smart camera designs where the chip interfaces directly with the CMOS sensor. The ISP handles demosaicing, noise reduction, and white balance on-chip, freeing the NPU to focus on inference. Available in Hailo-15H (high) and Hailo-15M (mid) SKUs.

Hailo-10H (40 TOPS): An M.2 form-factor chip aimed at PC and laptop integration, announced with Microsoft Copilot+ PC support. Different target market from industrial edge but uses the same architecture.

What Makes the Architecture Distinctive

Most competing edge NPUs use a matrix multiplication array with an internal SRAM scratchpad — the same basic structure as a miniaturised GPU or datacenter accelerator. Hailo’s architecture uses a “Dataflow Architecture” with a distributed processing element grid where data flows between processing nodes without a central memory bottleneck. The marketing claim is higher utilisation efficiency — more of the chip is actually computing at any given cycle.

In practice, the observable difference is efficiency at low batch sizes. Datacenter-focused architectures achieve their headline TOPS at batch sizes of 8, 16, or higher. Hailo’s architecture is designed for batch-size-1 inference — the relevant workload for a real-time camera feed — and their published FPS numbers reflect actual single-stream performance.

Software Stack: Hailo AI Software Suite

The compilation and runtime pipeline:

1. Export your model from PyTorch or TensorFlow as ONNX (preferred) or TFLite
2. Parse the ONNX model with the Hailo Model Zoo’s parser
3. Optimise using Hailo’s quantisation and graph optimisation toolchain
4. Compile to a .hef (Hailo Executable Format) file
5. Deploy the .hef using hailort on the target device

The Hailo Model Zoo provides pre-compiled .hef files for common architectures including YOLOv5, YOLOv8, ResNet-50, EfficientDet, and SSD-MobileNet. For these models, you can skip compilation and deploy immediately.

For custom models:

# Install the Hailo Dataflow Compiler (requires an account)
pip install hailo_dataflow_compiler

# Parse ONNX model
hailomz parse --net yolov8n --ckpt your_model.onnx

# Optimise and compile
hailomz optimize --net yolov8n
hailomz compile --net yolov8n

The compiler produces a .hef file for target deployment. Compilation happens on an x86 host, not on the target device.

Running Inference: Raspberry Pi + Hailo-8L

With the official Pi AI Kit (Hailo-8L + camera module):

import hailo_platform as hailort
import numpy as np

# Load and configure the HEF
target = hailort.VDevice()
hef = hailort.Hef("yolov8n.hef")
network_group = target.configure(hef)[0]

# Get input/output stream info
input_vstream_info = hef.get_input_vstream_infos()[0]
output_vstream_info = hef.get_output_vstream_infos()

# Run inference
with hailort.InferVStreams(network_group, input_vstream_info, output_vstream_info) as pipeline:
    input_data = {"input_layer_name": preprocessed_frame}  # numpy array
    results = pipeline.infer(input_data)

The rpicam-apps post-processing framework provides higher-level integration with less boilerplate for camera applications:

{
  "object_detect_tf": {
    "number_of_threads": 2,
    "refresh_rate": 1,
    "confidence_threshold": 0.5,
    "overlap_threshold": 0.5,
    "max_detections": 10,
    "labels_file": "coco_labels.txt"
  }
}

Performance Benchmarks in Practice

Real-world numbers from community benchmarks on Raspberry Pi 5 + Hailo-8L (13 TOPS):

On the full Hailo-8 (26 TOPS), roughly double these numbers for most models.

Industrial Deployment Patterns

Quality control cameras. The Hailo-8 paired with an industrial GigE or USB3 camera handles real-time defect detection on production lines at 30–60 FPS with complex detection models. Multiple Hailo-8 chips can be installed in a single PCIe carrier for parallel stream processing.

Smart traffic cameras. Vehicle detection, classification, and counting at intersections. The low power draw makes battery or solar-supplemented installations viable at remote locations.

Retail analytics. People counting, queue detection, and product shelf monitoring. The Hailo-15’s integrated ISP simplifies smart camera OEM designs for this application.

Agriculture and precision farming. Weed detection on autonomous sprayers, crop monitoring on drones with power budgets too tight for Jetson modules.

Hailo-8 vs Jetson Orin Nano

For industrial deployments where the tradeoffs matter:

The Hailo-8 is the better choice when you already have a host compute platform (Raspberry Pi, industrial PC, network camera processor) and want to add efficient inference without the Jetson’s cost and heat. The Jetson Orin Nano is better when you want a complete standalone compute module that runs Linux and handles its own camera I/O.

For new embedded vision designs targeting low power and cost, the Hailo-15’s integrated ISP makes it genuinely competitive with NVIDIA’s embedded lineup at a meaningfully lower system power and cost point.

The developer tooling has improved substantially since the early Hailo-8 days — the Model Zoo pre-compiled models, improved ONNX parser support for modern architectures, and the Raspberry Pi integration have removed most of the integration friction that previously made Hailo a specialist’s choice.