AI-narrated version of this post using a synthetic voice. Great for accessibility or listening while busy.
If you want a Raspberry Pi 5 to recognise things rather than just record them, there are now three mainstream ways to get there. You can bolt a Hailo accelerator onto the PCIe connector, buy a camera that runs the neural network inside the image sensor, or hang a Coral Edge TPU off a USB port. All three do on-device inference, which means no cloud subscription and no video leaving your network.
They are not interchangeable. One is a camera, one is a HAT that needs the PCIe slot, and one is a USB stick with a long history in self-hosted NVR software. The right pick depends heavily on what you are building: a people counter in a doorway, a battery-conscious wildlife box in the garden, or a Frigate instance watching four PoE cameras.
This guide walks through how each option handles inference, what software currently drives it, and the power and cooling consequences of each. It is aimed at people who already have a Pi 5 or are planning one, and who would rather understand the trade-offs than guess.
How the three options actually differ
The split is about where the model runs. The AI HAT+ keeps the camera ordinary and moves inference to a Hailo chip sitting on the Pi’s PCIe lane. The AI Camera inverts that: the Sony IMX500 sensor runs the network itself and hands the Pi metadata, so the Pi barely works. The Coral USB Accelerator is the middle road — an external TPU that any Debian-based system can talk to over USB 3.0.
That architectural difference drives everything else. On-sensor inference saves bandwidth and CPU but locks you to one camera and one model format. PCIe acceleration is faster and model-flexible, but occupies the connector you might have wanted for NVMe storage.
| Option | Where inference runs | Interface | Vendor-stated throughput | Strongest fit |
|---|---|---|---|---|
| Raspberry Pi AI HAT+ (26 TOPS) | Hailo accelerator on the HAT | PCIe, HAT+ form factor | 26 TOPS | Multi-stream NVR detection, heavier models |
| Raspberry Pi AI HAT+ (13 TOPS) | Hailo accelerator on the HAT | PCIe, HAT+ form factor | 13 TOPS | Single or dual stream detection, pose estimation |
| Raspberry Pi AI Camera | Inside the IMX500 sensor | Pi camera ribbon (CSI) | Integrated low-power inference engine | Low-CPU smart cameras, counting, wildlife |
| Coral USB Accelerator | Edge TPU in the USB dongle | USB 3.0 Type-C | 4 TOPS at roughly 2 TOPS per watt | Frigate and TensorFlow Lite workloads |
| Pi 5 CPU alone | Arm cores | None needed | No accelerator | Occasional stills, experiments, learning |
Raspberry Pi AI HAT+: Hailo acceleration over PCIe
The AI HAT+ is the most capable of the three on paper. Both versions measure 65mm x 56.5mm, mount with standard HAT+ points, and are rated for 0°C to +50°C operation, which is a reminder that these are indoor-ish devices unless you build a proper enclosure. The PCIe link is what makes the difference — it moves frames to the accelerator far more comfortably than USB does.
13 TOPS or 26 TOPS?
The 13 TOPS board carries the smaller Hailo part and is plenty for one or two camera streams running an object detection model. The 26 TOPS version doubles the headroom, which matters when you want a larger model, higher input resolution, or several streams sharing the same accelerator.
- POWERFUL AI ACCELERATOR: Features 13 TOPS (Tera Operations Per Second) Hailo AI accelerator for enhanced machine lear…
- COMPATIBILITY: Specifically designed for Raspberry Pi 5, utilizing PCIe interface for high-speed data transfer and sm…
- COMPACT DESIGN: Measures 65mm x 56.5mm, perfectly sized to stack on top of Raspberry Pi board while maintaining effic…
Pick the 26 TOPS board if the Pi is doing detection for a multi-camera setup, or if you expect to experiment with segmentation and pose models rather than settling on one detector. Pick 13 TOPS for a single-purpose build where the accelerator will mostly idle between events.
- HIGH PERFORMANCE: Features 26 TOPS (Trillion Operations Per Second) AI acceleration capability through the Hailo AI A…
- COMPATIBILITY: Specifically designed for the Raspberry Pi 5, connecting via PCIe interface for optimal data transfer…
- COMPACT DESIGN: Measures 65mm x 56.5mm, offering a space-efficient solution while maintaining full functionality as a…
Software support
This is where the HAT+ has caught up quickly. Raspberry Pi’s own camera stack exposes Hailo post-processing stages through rpicam-apps, so object detection and pose demos run from the command line without writing inference code. Hailo’s runtime and model zoo cover the conversion path if you bring your own network, and Frigate’s detector options now include Hailo hardware alongside Coral. If you are weighing HATs generally, our roundup of Raspberry Pi expansion boards worth adding covers how the PCIe slot gets contested.
Raspberry Pi AI Camera: inference on the sensor
The AI Camera pairs a 12.3 MP Sony IMX500 sensor with a neural network accelerator and an RP2040 that handles firmware and model management. It ships pre-loaded with a MobileNet vision model, so it produces useful detections out of the box. Sensor modes are 4056×3040 at 10fps and 2028×1520 at 30fps.
Because the network runs on the sensor, the Pi receives results rather than having to churn through every frame. That keeps CPU load low and leaves the PCIe connector free for an NVMe drive — a genuine advantage on a Pi 5 that also needs fast storage. Integration goes through picamera2, which surfaces inference metadata alongside the image stream.
- 12.3 MP Sony IMX500 Intelligent Vision Sensor with a powerful neural network accelerator
- Integrated low-power inference engine
- Integrated RP2040 for neural network and firmware management
The catch is scope. The AI Camera accelerates its own feed and nothing else, so it cannot help a Pi that is processing RTSP streams from PoE cameras. Models must be converted and packaged for the IMX500 toolchain, which is more involved than dropping a file into a models folder. For comparison with conventional modules, see our guide to Raspberry Pi camera modules.
Coral USB Accelerator on a Pi 5
The Coral USB Accelerator is the veteran here. Its Edge TPU is rated at 4 trillion operations per second at roughly 2 TOPS per watt, it connects over USB 3.0 Type-C with the supplied cable, and it runs TensorFlow Lite models compiled for the TPU. AutoML Vision Edge is supported if you want custom classification without hand-building a model.
Its real strength is ecosystem maturity. Frigate has supported Coral for years, documentation is everywhere, and the power draw is modest. Expect some driver setup on a current Pi OS kernel — this is the one option where you should read the install notes before plugging anything in.
- A USB accessory that brings machine learning inferencing to existing systems. Works with Raspberry Pi and other Linux…
- Performs high-speed ML inferencing: the on-board edge TPU Coprocessor is capable of performing 4 trillion operations…
- Works with Debian Linux: connects to any debian-based Linux system with an included USB 3.0 Type-C cable
Many people eventually move Frigate off the Pi entirely. If that is your direction, our complete Frigate hardware setup guide explains why a mini PC plus Coral is the common endpoint for multi-camera recording and retention.
Power, cooling and the PCIe slot
A Pi 5 under sustained inference load runs hot, and adding an accelerator above the board traps heat. The official Active Cooler combines a temperature-controlled blower fan with an aluminium heatsink and takes 5V from the Pi’s four-pin fan header, mounting with spring-loaded push pins. Fit it before you start troubleshooting throttling.
- Official RPi 5 Active Cooler — This is Official RPi Active Cooler for the latest RPi 5 4GB/8GB Board
- Composition–The RPi 5 Active Cooler is composed of Temperature-controlled Blower Fan and Aluminium Heatsink and come…
- Input Voltage–5V DC (supplied via four-pin fan header on RPi 5)
Use a power supply that meets the Pi 5’s requirements, since an accelerator plus a fan plus USB peripherals adds up. For permanently installed outdoor cameras or any mains or in-wall wiring, follow local electrical code and bring in a licensed electrician where it is required.
One more constraint: the AI HAT+ occupies the PCIe connector, so you cannot run it and an NVMe HAT at the same time on a standard Pi 5. Decide early whether fast local storage or on-board acceleration matters more.
Trade-offs and who should choose something else
If your project is a single smart camera — people counting at a door, a bird box, a shop-floor tally — the AI Camera is the tidiest answer. It needs no HAT, no USB dongle, and very little CPU.
If you are building a serious NVR with several high-resolution streams, a Pi is the wrong base regardless of accelerator. Storage throughput, continuous recording and transcoding all favour x86. Our notes on NPU mini PCs for local AI and the broader NVR software and hardware comparison both point the same way.
Coral is the choice when you already have TensorFlow Lite models or a Frigate config built around it, and when you want the accelerator to survive a move to different host hardware. A USB device travels; a HAT does not.
What to buy
For general-purpose vision AI on a Pi 5 — multiple models, room to grow, standard camera modules — the 26 TOPS AI HAT+ with the Active Cooler is the strongest combination, and the 13 TOPS version is the sensible step down for one fixed task. Choose the AI Camera when the whole project is one camera and you want the PCIe slot free for an SSD. Choose the Coral USB Accelerator when Frigate and TensorFlow Lite compatibility outweigh raw throughput, accepting a little driver work up front.
How these picks were chosen: HomeNode compares products using manufacturer specifications, current Amazon listing details and published owner feedback. Products are not tested in-house, and listings change often, so confirm current specifications on the product page before buying.
Affiliate disclosure: As an Amazon Associate, HomeNode earns from qualifying purchases at no additional cost to you.