A Raspberry Pi 5 can run object detection on its own, but not fast enough for more than a camera or two. An AI accelerator fixes that, and the hard part is choosing one, because these boards do not compete on a single number. Pick the best AI accelerator for your Raspberry Pi 5 by the job you give it, not by the TOPS on the box.
We checked specs, prices and stock on October 21, 2026, against Raspberry Pi, Hailo and Sixfab documentation and live Amazon listings. We did not benchmark these boards ourselves, so every FPS figure below names its source and test conditions.
Pick by workload
You have five boards and one M.2 alternative to choose from. Each wins a different job:
| Board | Chip and rating | Best for | Frigate | Price band |
|---|---|---|---|---|
| AI HAT+ 26 TOPS | Hailo-8, 26 TOPS INT8 | Frigate and vision on several cameras | Official | $160 to $185 |
| AI HAT+ 13 TOPS | Hailo-8L, 13 TOPS INT8 | One or two cameras on a budget | Official | $125 to $170 |
| AI HAT+ 2 | Hailo-10H, 40 TOPS INT4, 8 GB RAM | Local LLMs and VLMs next to vision | Unofficial fork | about $340 as a kit |
| Sixfab AI HAT+ | DEEPX DX-M1M, 25 TOPS INT8, 2 GB RAM | Raw YOLO throughput | Dev branch only | about $95 direct |
| Raspberry Pi AI Camera | Sony IMX500, on-sensor | One camera, any Pi, PCIe lane kept free | No | $100 to $110 |
| Waveshare Hailo-8 M.2 | Hailo-8, 26 TOPS INT8 | A card you can move to a mini PC later | Official | $210 to $230 plus a HAT |
Raspberry Pi does not sell these boards on Amazon itself, so every listing is a third-party seller and prices move with stock. Check the current price before you order.
Check TOPS and the PCIe lane
You cannot compare the TOPS numbers directly. Raspberry Pi rates the AI HAT+ 2 at 40 TOPS in INT4 and the AI HAT+ at 13 or 26 TOPS in INT8. For vision, Raspberry Pi calls the AI HAT+ 2 comparable to the 26 TOPS board, while Hailo says it’s up to 100 percent faster. The Hailo Model Zoo shows the gap depends on the model.
The PCIe lane matters more. Your Pi 5 has one PCIe lane, and every HAT here takes it. You cannot run an NVMe boot drive and an accelerator together without a PCIe switch HAT, and the switch HATs we found run at Gen 2 only. The AI HAT+ and AI HAT+ 2 set the lane to Gen 3 automatically, so dtparam=pciex1_gen=3 is only needed on the older AI Kit or a bare M.2 card on the M.2 HAT+.
If you’d rather keep NVMe, the AI Camera is the one option that leaves the lane free. Our Raspberry Pi 5 kit guide covers the board and storage side.
Compare the measured numbers
No single benchmark covers every board. Three sources fill the gap, each with different test conditions:
- Frigate hardware docs: detector inference time per frame
- A Sixfab bench on one Pi 5 8 GB, run by a DEEPX partner
- The Hailo Model Zoo, measured on x86 with PCIe Gen 3 x4
The Frigate hardware docs list about 6 ms for the Hailo-8 and 10 ms for the Hailo-8L on SSD MobileNet v1. On the default YOLOv6n model they list about 7 ms and 11 ms. That works out to roughly 140 and 90 detections a second per detector. Jeff Geerling measured about 12 ms with two cameras on a Hailo-8 in a Compute Module 5 build.
Sixfab ran all four NPUs on the same Pi 5 in its benchmark. A DEEPX partner wrote it, so treat it as a vendor test. Its most useful finding still cuts against the vendor:
| YOLOv8l 640 | DEEPX DX-M1M | Hailo-8 | Hailo-8L | Hailo-10H |
|---|---|---|---|---|
| NPU throughput, async FPS | 50.1 | 19.6 | 15.0 | 40.0 |
| End to end with decode, FPS | 11.2 | 10.3 | 8.7 | 11.9 |
| Average system power | 6.73 W | 6.53 W | 6.08 W | 7.59 W |
End to end, every board landed between 7 and 21 FPS across the Sixfab models. For a single stream, video decode on the Pi 5 is the bottleneck, not the NPU. A faster NPU only pays off when you run several streams or batch frames.
The Hailo Model Zoo shows far higher numbers, such as 491 FPS for YOLOv8s on the Hailo-8, 110 on the 8L and 166 on the 10H. Those were measured with four PCIe lanes on a desktop CPU. Read them as an upper bound, because your Pi 5 gives the card one lane.
Pick the AI HAT+ 26 TOPS for Frigate
We’d put this board under a Pi 5 running Frigate. Its Hailo-8 is rated at 26 TOPS INT8, and Frigate supports it through its Hailo detector (type: hailo8l, which covers the Hailo-8 too). After you install the dkms and hailo-all packages, rpicam-apps and Picamera2 run the Hailo demos too.
In the Sixfab bench it narrowly had the best detections per watt, 1.74 against 1.69 FPS per watt, measured end to end. It doesn’t run LLMs, though, and it takes your PCIe lane. The Amazon listing we checked was a third-party seller at about $162, with only a few units left.
Our Coral to Hailo guide has the Frigate detector config. For the base setup, start with our Frigate install guide.
Save with the 13 TOPS AI HAT+
The 13 TOPS AI HAT+ costs about $125 to $170 from third-party sellers on Amazon; check the current price. A Hailo-8L runs Frigate at about 11 ms per frame on YOLOv6n, which is plenty for one or two cameras. It drew the least power in the Sixfab bench, 6.08 W average at 33 to 47 C, though Sixfab cooled the AI HAT+ 2 with its official heatsink and the others with an extruded one. Geerling, who tested a Hailo-8, notes the same Frigate setup works on the 8L.
Heavier models are where it falls off. The zoo shows 110 FPS on YOLOv8s against 491 for the Hailo-8. It costs only about $35 less than the 26 TOPS board, so buy it only if you know you’ll stay small.
Buy the AI HAT+ 2 for generative AI
It is the only board here with 8 GB of its own memory, LPDDR4X next to a Hailo-10H. That memory lets it run small LLMs through the Hailo Ollama server and vision-language models through hailo-apps. Hailo’s own figures, published on the Raspberry Pi blog, put time to first token at 320 ms on the board against 2,039 ms on the Pi 5 CPU with llama.cpp, for Qwen2.5-1.5B in 4-bit with a 96-token prompt.
It’s strong on large vision models too, with 40.0 FPS on YOLOv8l in the Sixfab bench against 19.6 for the Hailo-8. Our verdict is narrower than the spec sheet, though. After the first token, independent tests found generation slower than the Pi 5 CPU alone, so it helps prompt latency more than chat speed.
Frigate support is still an unofficial fork, not an official detector. It ran hottest in the Sixfab bench, up to 68.1 C on its official heatsink, and on Amazon its kits cost about twice the 26 TOPS board for similar small-model vision. Our Jetson Orin vs AI HAT+ 2 comparison has the full token tables.
On Amazon it sells only in kits, such as this Yahboom AI HAT+ 2 kit at about $340 from Yahboom. Its hailo-h10-all packages cannot sit alongside the hailo-all stack the AI HAT+ and AI Kit use.
Try the Sixfab DEEPX HAT
The Sixfab board carries a DEEPX DX-M1M rated at 25 TOPS INT8, with 2 GB of its own RAM on the same PCIe Gen 3 x1 link. In the Sixfab bench it posted the highest NPU throughput on YOLOv8l, 50.1 FPS against 19.6 for the Hailo-8. It also costs less than the 26 TOPS AI HAT+.
The only Pi 5 benchmark we found is Sixfab’s own. Frigate merged a community DEEPX detector into its dev branch in September 2026, but it is not in a stable release yet, and there is no LLM support. Until that detector ships, it suits people running their own detection code or the Frigate dev image. End to end it lands with the Hailo boards, at 11.2 FPS on YOLOv8l. Sixfab lists CE certification as pending and recommends the Raspberry Pi 27 W supply, because 15 W chargers trigger undervoltage. It sells the board directly, not on Amazon.
Use the AI Camera for one camera
Inference runs on the Sony IMX500 sensor in the camera, so it works on any Raspberry Pi and leaves the PCIe lane on your Pi 5 free for NVMe. The imx500-all package ships MobileNet SSD and PoseNet models ready to run.
Its limits are tight. The sensor has about 8 MB for firmware, weights and working memory, and it takes int8 models with a 640 by 640 input at most. Loading a model the first time can take minutes. It runs inference only on its own sensor feed, which rules it out as a detector for IP camera streams in Frigate.
The Amazon listing was a third-party seller at about $106.
Consider a Hailo-8 M.2 card
A bare Hailo-8 M.2 card on the official Raspberry Pi M.2 HAT+, about $25 on Amazon, gives you the same 26 TOPS chip. Sixfab measured it within 0.1 percent of the AI HAT+. The Waveshare card is a 2242 module rated at 2.5 W typical and 8.65 W maximum.
With the HAT it costs about half again as much as the AI HAT+, and the M.2 HAT+ is specified for PCIe Gen 2. Waveshare says Gen 3 must be enabled with dtparam=pciex1_gen=3 in /boot/firmware/config.txt, and the Raspberry Pi docs warn that the Pi 5 is not certified for Gen 3 speeds, so the link may be unstable. Buy it if you might move the card to an x86 mini PC later, as in our Frigate NVR appliance build.
Skip these three
Three options show up in every search, and none suits a new Pi 5 build:
- Google Coral USB: libraries archived, PyCoral wheels stop at Python 3.9
- Raspberry Pi AI Kit: discontinued, same chip as the 13 TOPS HAT
- Axelera Metis M.2: 2280 card, won’t mount on the official HAT
Frigate itself no longer recommends Coral for new installs, and Raspberry Pi OS now ships Python 3.13. If a Pi is the wrong host altogether, read our mini PC with NPU guide and the Pi 5 vs N150 comparison. Our Hailo vs Coral guide compares detectors across platforms.
