For the clearest documented route to hardware-accelerated CNN inference, use a Raspberry Pi 5 with a Raspberry Pi AI HAT+ or AI HAT+ 2, then install the supported software and a model compatible with that route. The add-on’s neural processing unit (NPU) accelerates supported workloads; it does not make every CNN run automatically or guarantee a particular speedup.
Which Raspberry Pi setup supports CNN acceleration?
Raspberry Pi’s documented Hailo acceleration path is for Raspberry Pi 5. The AI HAT+ family adds a Hailo NPU, and Raspberry Pi describes uses including object detection, camera post-processing, robotics, and other moderate neural workloads. The boards are not a generic accelerator promise for every Raspberry Pi model.
| Hardware | Accelerator specification | What to know |
|---|---|---|
| AI HAT+ (Hailo-8L) | 13 TOPS | One of two AI HAT+ variants for supported inference workloads. See Raspberry Pi AI HAT documentation and the AI HAT+ product page. |
| AI HAT+ (Hailo-8) | 26 TOPS | The higher-specification AI HAT+ variant. TOPS is an accelerator specification, not a CNN benchmark result. See Raspberry Pi AI HAT documentation and the AI HAT+ product page. |
| AI HAT+ 2 (Hailo-10H) | 40 TOPS (INT4); 8 GB onboard memory | A distinct product with generative AI capabilities as well as supported vision workloads; those additional capabilities are not required for ordinary CNN inference. See Raspberry Pi AI HAT documentation. |
The older Raspberry Pi AI Kit contains a Hailo-8L accelerator but is no longer in production. Raspberry Pi recommends AI HAT+ or AI HAT+ 2 for new designs; see its AI software documentation.
What you need before installing the software
The current official Hailo setup route requires Raspberry Pi 5 running 64-bit Raspberry Pi OS (Trixie), an AI HAT+ or AI HAT+ 2, and the documented software dependencies, drivers, and a supported model. For camera-based vision applications, you also need a supported camera. A board being detected by the system is not enough to make an arbitrary model compatible. Raspberry Pi lists the setup requirements in its AI software documentation.
The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →#1 Best Overall
- Includes Raspberry Pi 5 with 2.4Ghz 64-bit quad-core CPU (8GB RAM)
- Includes 128GB Micro SD Card pre-loaded with 64-bit Raspberry Pi OS, USB MicroSD Card Reader
- CanaKit Turbine Black Case for the Raspberry Pi 5
- CanaKit Low Noise Bearing System Fan
- Mega Heat Sink - Black Anodized
- Host: Raspberry Pi 5.
- Operating system: 64-bit Raspberry Pi OS (Trixie), as required by the current documented setup.
- Accelerator: AI HAT+ or AI HAT+ 2.
- Software and model: Install the relevant documented dependencies and drivers, and choose a supported model and runtime path.
- Camera, if needed: Use a supported camera for camera-based applications.
Set up the HAT and choose a supported inference path
- Assemble the hardware. Fit the AI HAT+ or AI HAT+ 2 to Raspberry Pi 5 according to Raspberry Pi’s assembly documentation. Raspberry Pi recommends an Active Cooler for the host Pi 5, but it is optional. The AI HAT+ 2 package includes a heatsink, which Raspberry Pi recommends installing alongside the Active Cooler.
- Install the documented OS and software. Use 64-bit Raspberry Pi OS (Trixie) and follow Raspberry Pi’s current AI software setup instructions for dependencies and drivers.
- Select a supported model and runtime. Confirm that your model is supported by the chosen software route; do not assume that an export from any CNN framework will run unchanged on the NPU.
- Connect a camera if the application needs one. Raspberry Pi says camera applications such as
rpicam-appsand Picamera2 can use the Hailo NPU for supported tasks, including image recognition and object detection. The supported integration is described in the AI HAT documentation. - Check the current LiteRT path if it fits your project. Raspberry Pi also documents a LiteRT workflow that can offload inference to AI HAT+ and AI HAT+ 2. Review the AI software documentation and LiteRT getting-started guide for the supported workflow.
What kind of speedup should you expect?
The published 13, 26, and 40 TOPS figures describe accelerator specifications, not measured latency or a speedup for a particular CNN. Raspberry Pi’s cited product and software material does not provide a controlled CPU-versus-NPU comparison for a named model, input size, runtime, and quantization setup. No specific latency, throughput, accuracy, or power improvement can therefore be inferred from those TOPS figures alone.
Application performance depends on whether the model’s operations and conversion path are supported, as well as input dimensions, preprocessing and postprocessing, thermal conditions, and any CPU work that remains in the pipeline. A fast NPU stage may not materially improve an application if image preparation, data movement, or other CPU tasks dominate end-to-end time.
Rank #2
- Includes Raspberry Pi 5 with 2.4Ghz 64-bit quad-core CPU (4GB RAM)
- Includes 128GB Micro SD Card pre-loaded with 64-bit Raspberry Pi OS, USB MicroSD Card Reader
- CanaKit Turbine Black Case for the Raspberry Pi 5
- CanaKit Low Noise Bearing System Fan
- CanaKit Mega Heat Sink - Black Anodized
How to benchmark a CNN fairly
Compare CPU-only inference with NPU inference using the same application conditions. Record enough detail for another developer to reproduce the result rather than reporting a bare “times faster” figure.
Quick Recap
Best Value
- Includes Raspberry Pi 5 with 2.4Ghz 64-bit quad-core CPU (8GB RAM)
- Includes 32GB EVO+ Micro SD Card pre-loaded with 64-bit Pi OS, USB MicroSD Card Reader
- CanaKit Turbine Black Case for the Raspberry Pi 5
- CanaKit 45W PD Power Supply for the Raspberry Pi 5
- Display Cable - 6 foot (Supports up to 4K 60p)
Rank #4
- Includes Raspberry Pi 5 16GB with 2.4Ghz 64-bit quad-core CPU (16GB RAM)
- Includes 128GB Micro SD Card pre-loaded with 64-bit Raspberry Pi OS, USB MicroSD Card Reader
- CanaKit Turbine Black Case for the Raspberry Pi 5
- CanaKit Low Noise Bearing System Fan
- Mega Heat Sink - Black Anodized
Rank #3
- CanaKit Raspberry Pi 5 Essentials Starter Kit
- Model and conversion: Name the CNN, runtime, model version, and any conversion or quantization used. Check accuracy after conversion against the original model.
- Input and workload: State the input dimensions, batch size if applicable, and whether the test uses a live camera stream or preloaded images.
- Timing boundary: Report both inference-only timing and end-to-end application latency where relevant. Specify whether preprocessing, postprocessing, and data movement are included.
- Throughput and latency: Measure the rate of completed inferences and the time per result under the intended workload; do not treat either as a substitute for the other.
- Power and thermals: Record the measurement method and operating conditions, and observe the system long enough to detect thermal effects.
- Software and hardware: Identify the Pi, HAT variant, operating system, drivers, and runtime so the comparison has a defined configuration.
When the AI HAT is—and is not—the right choice
Use an AI HAT when
- Your target is Raspberry Pi 5 and your model is supported by the documented Hailo software path.
- Your workload benefits from NPU offload, such as supported object detection or camera-based vision processing.
- You can validate end-to-end performance, converted-model accuracy, power, and thermal behavior on the actual application.
Do not assume it will help when
- Your model or framework export has not been confirmed as supported by the required runtime.
- The application is limited by CPU-side preprocessing, postprocessing, or another pipeline stage rather than inference.
- You need a guaranteed speedup, latency, or accuracy figure before benchmarking the exact model and configuration.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
