The Raspberry Pi has been the go-to platform for IoT prototyping and edge node deployment for over a decade. The Pi 4 could handle lightweight ML inference with TensorFlow Lite, but anything involving real-time computer vision — object detection, pose estimation, face recognition — pushed the CPU-only platform to its limits and well beyond acceptable latency for production use.
The Raspberry Pi 5 with Hailo M.2 HAT+ changes that. It’s not a replacement for Jetson hardware on demanding multi-camera pipelines, but for a large class of edge AI deployments — single-camera object detection, local inference on a production line, presence detection in a smart building node — it delivers capable inference at a price point that was previously unachievable.
The Hardware
The Raspberry Pi 5 uses the Broadcom BCM2712 SoC, a significant step up from the Pi 4. The integrated VideoCore VII GPU gives the Pi 5 useful capability for some ML workloads, but the more interesting development is the M.2 HAT+ — a PCIe adapter board that connects an M.2 device to the Pi 5’s PCIe 2.0 interface.
The Hailo-8L is an edge AI accelerator available in the M.2 2242 form factor, designed specifically for use with the Pi 5’s HAT+. It delivers 13 TOPS (trillion operations per second) of dedicated neural processing — purpose-built for inference, not general compute. For comparison, the Raspberry Pi 5 CPU delivers roughly 0.1-0.3 TOPS for typical ML inference workloads. The Hailo-8L changes what’s computationally feasible by roughly an order of magnitude.
The full kit — Pi 5 8GB, M.2 HAT+, Hailo-8L M.2 module, cooling, and a decent power supply — comes in around £130-150. That’s notably cheaper than a Jetson Orin Nano Developer Kit at around £300-400, though the Jetson offers significantly more AI performance headroom for complex or multi-stream workloads.
What You Can Actually Run
Raspberry Pi Ltd maintains a repository of models pre-compiled for the Hailo-8L: hailo-ai/hailo_model_zoo on GitHub. The practical starting point is their inference application examples in hailo-rpi5-examples.
Object detection at real-time framerates is the primary capability unlock. YOLOv8n (the nano variant) runs at 25-30fps on a single 1080p camera stream using the Hailo-8L, compared to 5-8fps on the Pi 5 CPU alone. YOLOv8s at 15-20fps. For presence detection, vehicle counting, or simple surveillance applications, that’s production-ready performance.
Pose estimation using HigherHRNet or similar models works at 15-20fps — adequate for movement monitoring, occupancy detection, and health and safety compliance applications where you’re checking that people are wearing PPE or in correct positions.
Instance segmentation (YOLACT or similar) runs at lower framerates but is usable for applications where you need pixel-level object boundaries rather than just bounding boxes.
Classification workloads (single-image classification rather than detection in a scene) run efficiently and are well within the platform’s capability for applications like product inspection or document sorting at the edge.
Getting Started
The Raspberry Pi OS (Bookworm, 64-bit) with the Hailo software stack is the supported development environment. Installation has been simplified considerably since the initial HAT+ launch — the Hailo HailoRT runtime and the PCIe driver are now available via the standard apt repository.
# Install Hailo dependencies
sudo apt install hailo-all
# Verify the Hailo device is detected
hailortcli fw-control identifyIf that returns device information, you’re ready to run inference. Raspberry Pi’s own rpicam-apps has been updated to include Hailo post-processing stages, meaning you can pipe camera frames directly through a Hailo model with a single command-line invocation — no Python required for initial testing.
For production applications, the Python API (hailo-python) provides the inference integration layer. Models need to be compiled to Hailo’s HEF format using their compiler toolchain (available in Docker) or downloaded pre-compiled from the model zoo.
The ONNX-to-HEF compilation path handles most standard architectures well: YOLO variants, ResNet, MobileNet, EfficientDet. If you’re using a custom model, the Hailo compiler documentation describes quantisation requirements and any layer constraints — it’s more restrictive than ONNX Runtime on generic hardware but well-documented.
Practical Use Cases
Smart building occupancy sensing is a natural fit. A Pi 5 with Hailo running person detection on a corridor camera, connected to Home Assistant or a building management system via MQTT, delivers presence data with sub-100ms latency at under £150 per node. That’s economically viable for multi-room deployments in a way that Jetson hardware isn’t.
Agricultural monitoring — livestock counting, crop inspection — where edge inference needs to work without cloud connectivity and on solar/battery power. The Pi 5’s power consumption (5-8W active with Hailo inference running) is meaningfully higher than a microcontroller but still manageable for solar-powered edge nodes.
Production line inspection for small-scale manufacturing where you need a single-camera defect detection or measurement system. The Pi 5 can capture, run inference, trigger a reject mechanism, and log results locally without a cloud round-trip.
Retail analytics — footfall counting, queue detection — for independent retailers or smaller hospitality venues where Jetson hardware is financially unjustifiable.
Limitations to Know
The Hailo-8L is limited to 13 TOPS compared to 26 TOPS for the full Hailo-8. For single-stream applications, that’s rarely the bottleneck — model complexity and the application code around inference typically limit throughput before the accelerator is saturated. For multi-stream applications or complex model pipelines, you hit the ceiling faster.
The HAT+ uses PCIe 2.0 x1, which limits data transfer bandwidth. For very high-resolution inputs (4K cameras) or high-framerate workloads, the interface can become a constraint.
The software stack is under active development. Some operations between the Pi 5 and the HAT+ that should work do so unreliably on older firmware versions. Keeping both the Pi OS and the Hailo runtime updated matters more here than on more mature platforms.
But for the use case it targets — practical, affordable edge AI inference at the single-camera, single-model level — the Pi 5 with Hailo M.2 HAT+ is the most capable option in its price bracket.