Sixfab AI HAT+ for Raspberry Pi 5: 25 TOPS NPU From $63

If you have ever wanted to bolt a real neural accelerator onto a Raspberry Pi 5 without juggling USB sticks or eating half your power budget, the new Sixfab AI HAT+ for Raspberry Pi 5 is exactly the kind of board I get excited about. Announced June 1, 2026, it is a HAT+ compliant add-on built around the DEEPX DX-M1 NPU, and it does the one thing I always want from edge AI hardware: it talks to the Pi over the fast lane and sips power while doing serious work.

What the Sixfab AI HAT+ actually is

This is a dedicated NPU board, not a general-purpose coprocessor. The DEEPX DX-M1 silicon handles the inference, and Sixfab wires it into the Pi 5's native PCIe Gen 3 x1 interface using a 16-pin FFC ribbon cable. That detail matters. Instead of being throttled by USB 3.0 like a lot of older AI accelerators, you get the Pi 5's real PCIe link feeding the NPU, which is the right way to keep an object detection pipeline fed with frames.

Power comes straight from the 40-pin GPIO header, so there is no separate brick or auxiliary connector to manage. One cable for data, the header for power, and you are running.

Two tiers, two price points

Sixfab is shipping two configurations, and the spec gap is clean and easy to reason about:

- 13 TOPS (INT8) with 512MB LPDDR4X — $63

- 25 TOPS with 1GB LPDDR4X — $90

That is genuinely aggressive price-to-performance. The entry board lands at $63 for 13 trillion INT8 operations per second, and stepping up to 25 TOPS plus a full gigabyte of dedicated LPDDR4X only costs $27 more. For anyone building a fleet of camera nodes, that delta is small enough that I would just buy the 25 TOPS version and give myself headroom on model size and batch flexibility.

Real on-device performance and power draw

Here is the number that sells it for me: roughly 30 to 35 FPS on YOLOv8n while drawing only 2.5 to 3 watts. Object detection, segmentation, and classification all run entirely on-device, so nothing leaves the Pi and no cloud round-trip is involved.

Let me put that power figure in context. A Pi 5 on its own can pull several watts under load, and plenty of USB accelerators add meaningful overhead. An NPU that delivers smooth real-time YOLOv8n inference for under 3W is the kind of efficiency that makes battery-powered and solar-powered builds realistic. Thirty-plus FPS is comfortably above the ~24 FPS threshold where detection feels live rather than choppy, which is the sweet spot for a smart camera or a robotics perception loop.

Frames per watt is the real metric

I always come back to frames per watt for edge devices, because that is what determines what you can actually deploy in the field. At ~30 FPS for ~3W, you are looking in the neighborhood of 10 inferences per watt-second on a small detector. That is the difference between a project you can run off a power bank and one that needs to stay tethered.

Setup and what you can build with it

Software is the part that usually kills these boards, so I am glad Sixfab kept it boring. After you install the dxrt-runtime package, Raspberry Pi OS auto-detects the HAT+. No exotic kernel patching described in the launch materials, no manual device-tree surgery on the user's part — install the runtime, the OS sees it, you start running models.

So what do you actually build? The obvious candidates:

- A privacy-first smart security camera that does on-device person and object detection, with nothing streamed off-box.

- A robotics perception module that runs real-time segmentation for navigation and obstacle avoidance.

- A wildlife or retail counter doing continuous classification at the edge.

- A low-power monitoring node running on solar or battery thanks to that sub-3W envelope.

My take on the price-to-performance

The Sixfab AI HAT+ for Raspberry Pi 5 hits the trifecta I look for in edge AI hardware: a fast native PCIe link, low power draw, and a runtime that just works. Starting at $63 for 13 TOPS and topping out at $90 for 25 TOPS, this is one of the cleanest ways to add real machine-vision horsepower to a Pi 5 right now. It is available on the Sixfab store today, and if you have a perception project sitting on your bench, this is the board I would slot in first.

Sources: CNX Software, June 1 2026; LinuxGizmos, 2026; Sixfab product page