Sixfab AI HAT+ Brings a DEEPX NPU and Local Edge AI to the Raspberry Pi 5

Local AI Lands on the Pi 5 — No Cloud Required

Here is the spec sheet I have been waiting to write up. On June 26, 2026, Sixfab and AI-chip company DEEPX formally launched the Sixfab AI HAT+, a PCIe accelerator board that bolts a dedicated neural processing unit onto the Raspberry Pi 5. The pitch is simple and, for once, the numbers back it up: real-time machine vision running entirely on-device, for the price of a nice dinner.

The board comes in two flavors built around the DEEPX DX-M1 NPU. The entry model delivers 13 TOPS for $63, and the step-up version pushes 25 TOPS for $90. Both connect through the Pi 5's PCIe interface using the HAT+ form factor, so you stack it, screw it down, and you are off — no external box, no GPU, no recurring cloud bill.

What 25 TOPS Buys You on a Single Board Computer

TOPS (trillions of operations per second) is one of those headline numbers that only matters if the silicon can actually feed it, so let's talk real workloads. On an 8GB Pi 5, the AI HAT+ runs YOLOv8n object detection at 640×640 resolution and hits roughly 30 to 35 frames per second. That is comfortably real-time for a single-board computer, which means smooth bounding boxes on a live camera feed rather than a slideshow.

The DX-M1 is purpose-built for vision: object detection, image segmentation, classification, and the kind of always-on perception that edge AI deployments depend on. Sixfab is positioning it squarely at "Physical AI" — robotics, smart agriculture, factory automation, and any project where a camera needs to understand what it sees without phoning home. I appreciate the honesty in the positioning, too: this is a vision accelerator, not a transformer-decoder engine, so it is not the board for running a local large language model. Matched to the right task, that focus is a feature.

Why On-Device Inference Matters

The reason I keep coming back to boards like this is privacy and latency. When inference happens locally on the NPU, your camera frames never leave the device. There is no round trip to a data center, which means lower latency for time-sensitive tasks like a robot avoiding an obstacle, and no exposure of a video feed to anyone else's servers. For makers building smart-home sensors, wildlife cameras, or assistive tools, that combination of private, fully local AI is genuinely empowering.

It also keeps costs predictable. A maker project that processes video around the clock would rack up real money on a cloud inference API. A one-time $63 to $90 for an accelerator that draws a few watts is the kind of math that turns "someday" projects into weekend builds.

A Friendly On-Ramp to Edge AI

What I like most is the accessibility. The Raspberry Pi 5 is already the most approachable computer in the maker world, with a huge community and endless tutorials. Bolting a capable NPU onto that ecosystem lowers the barrier to edge AI from "buy a dev kit and learn a new toolchain" to "stack a HAT on the Pi you already own." That is exactly the kind of democratization that gets more people building, and as we often note in our AI coverage, the most exciting progress is the kind ordinary tinkerers can actually touch.

If you have been curious about putting real machine vision into a robot, a bird feeder, or a workshop camera, this is one of the cleanest and cheapest on-ramps I have seen. The Pi 5 finally has an affordable, local brain for its eyes.

Sources: PR Newswire — "DEEPX and Sixfab Launch DEEPX AI HAT to Drive Edge Physical AI on Raspberry Pi" — June 26, 2026; CNX Software — "Sixfab AI HAT for Raspberry Pi 5 integrates DEEPX DX-M1 AI accelerator" — June 2026.