DEEPX AI HAT Lands on Sixfab Raspberry Pi 5 Starting at $63

DEEPX’s NPU now lives on Raspberry Pi 5. The Korean AI chip company has partnered with Sixfab, an Official Raspberry Pi Design Partner, to ship the Sixfab AI HAT+ for Raspberry Pi 5, co-branded “Intelligentized by DEEPX,” with two SKUs priced at $63 and $90.

The board mounts on any Pi 5 and delivers up to 25 TOPS at INT8 inside a 3 W envelope, with inference running entirely on-device after a single APT install. DEEPX is framing the launch as the opening move in a longer play: turn a single-board computer with a global developer community into the default front end for Physical AI.

What the Sixfab AI HAT+ Actually Ships

The Sixfab AI HAT+ is a HAT+ specification compliant add-on board that mounts onto the Raspberry Pi 5’s 40-pin GPIO header for power and connects to the Pi 5’s external PCIe Gen 3 x1 port through a 16-pin FFC cable. The 40-pin header is used only for power, ground, and the HAT+ ID EEPROM, which leaves every other GPIO free for stacked HATs and user peripherals. Both variants share the same 65 × 56.5 mm PCB, the same software stack, and the same connectors.

The cheaper board carries DEEPX’s DX-M1ML NPU with 13 TOPS at INT8 and 1 GB of LPDDR4X memory. The higher-end board carries the DX-M1M with 25 TOPS at INT8 and 2 GB of LPDDR4X. Sixfab lists typical NPU draw at 3 W, with combined Pi 5 + HAT+ system power of 13 to 15 W on the official Raspberry Pi 27 W USB-C PD supply, which the company says is required because standard 5 V / 3 A supplies will trigger under-voltage warnings and throttling.

Sixfab sells the 13 TOPS version for $63 and the 25 TOPS version for $90, both available direct through its store. Full specs and ordering details sit on Sixfab’s AI HAT+ product page with full specs.

How the AI Pipeline Splits Between Pi and NPU

The HAT+ keeps the Pi 5 in charge of everything except the neural math.

Camera frames from a CSI, USB, or IP source enter the Pi 5, where the Cortex-A76 CPU handles capture, color conversion, tensor packing, and any pre-processing the model needs. The prepared tensor crosses the PCIe Gen 3 x1 link to the DEEPX NPU, which runs the compiled DXNN model in INT8. Results come back over PCIe, and the Pi 5’s CPU runs any post-processing (such as non-maximum suppression) and routes outputs to a display, log, or RTSP stream.

Sixfab’s documentation describes the board as shipping with a vision focus: object detection, classification, and segmentation at real-time rates. The product page lists inference performance at 30 to 35 FPS, with the higher-end DX-M1M option on a Raspberry Pi 5 with 8 GB of RAM able to sustain up to four 1080p streams through the same pipeline before CPU work becomes the bottleneck.

After the runtime is installed, the system runs fully offline. No frames or model outputs leave the Raspberry Pi 5 unless the application explicitly forwards them, which Sixfab positions as a fit for privacy-sensitive vision deployments. Models are brought in through two paths: the Sixfab Model Zoo for pre-compiled DXNN files such as YOLOv8n, or the DXNN SDK for custom ONNX models converted through the DX-COM compiler. INT8 quantization is automatic, and Sixfab expects approximately 2 % accuracy loss compared with the original FP32 model. The full pipeline and software paths are spelled out in Sixfab’s technical overview for the AI HAT+.

The Strategic Play Behind “Intelligentized by DEEPX”

DEEPX is not selling the HAT+. Sixfab is. DEEPX supplies the NPU, the DXNN compiler, and the runtime that Sixfab integrates into a HAT+ spec compliant board as an Official Raspberry Pi Design Partner. The product is co-branded, with Sixfab’s name on the box and DEEPX’s mark stamped as “Intelligentized by DEEPX” alongside the Raspberry Pi “Built on Raspberry Pi” designation.

The joint announcement positioned the move as a deliberate counterweight to cloud-centric AI. Lokwon Kim, CEO of DEEPX, said the future of Physical AI “belongs to the billions of devices, robots, and industrial endpoints in the real world,” not the data centers. Kim framed the Pi 5 launch as a long game: turn the global developer base for low-cost embedded computing into the on-ramp for Physical AI hardware at scale.

By pairing Raspberry Pi’s versatility with our low-power NPU, we aim to drive the democratization of Physical AI so that anyone can build and deploy advanced AI solutions. This launch is a cornerstone of our long-term strategy to cultivate future talent and developer communities. DEEPX will expand collaborations with governments, academia, and industries worldwide to ultimately emerge as the definitive global standard platform for Physical AI.

Kim made the comments in the joint announcement of the Sixfab AI HAT+ on June 26, 2026, covered in the DEEPX and Sixfab joint announcement on the AI HAT+.

Where DEEPX Already Has Real-World Deployments

DEEPX is not new to embedded AI silicon. The Pi 5 launch sits inside a broader push that already has named production partners.

At DEEPX’s CES 2026 Foundry session, the company brought together senior executives from Hyundai Motor Group Robotics Lab, Baidu, Edge AI Foundation, Ultralytics, and Wind River to discuss Physical AI deployment at scale. The panel context and partner roster sit on DEEPX’s news page covering the CES 2026 partners.

Hyundai’s Dongjin Hyun, Executive Director of the robotics lab, said the company had validated DEEPX on-device AI in real-world conditions. Hyun added that Hyundai plans “to deploy these solutions in our next-generation robots and security systems starting in 2026.” That kind of named-OEM commitment is what DEEPX can now point to when pitching its Pi 5 board to industrial customers.

Beyond robotics partners, DEEPX has built the tooling layer that the HAT+ leans on. Francesco Mattioli, Lead Partner Engineer at Ultralytics, said the partnership “simplifies the journey from training to deployment, delivering real-world performance without requiring developers to become hardware experts.” That same gap is what Sixfab is targeting by putting pre-compiled YOLOv8n and MobileNet models in its Model Zoo.

Sixfab lists the workloads its Pi 5 customers are running on the AI HAT+. The categories match the deployment scope DEEPX has been selling into:

• AI cameras and on-prem security
• Smart-city and infrastructure
• Robotics perception
• Drones and autonomous platforms
• Industrial inspection
• Edge inference nodes

What the Board Cannot Do Yet

The Sixfab AI HAT+ is vision-first silicon, and the company is explicit about the limits.

LLMs are not supported today. Sixfab’s documentation says “LLM support is on the DEEPX roadmap and Sixfab will support it as the silicon enables,” and it gives no date for that change. The NPU runs INT8 only, with no FP16 or FP32 paths on the chip itself, so any application that needs higher precision at inference time has to keep that step on the Pi 5’s CPU. Training is also off-board: models are trained on a host machine, exported to ONNX, then compiled with DX-COM before they run on the HAT+.

The supported host list is narrow. The board is built specifically for the Raspberry Pi 5 and is not electrically or mechanically compatible with the Raspberry Pi Compute Module 5 on the CM5 IO Board, the variant most industrial system builders gravitate to for headless deployments. Sixfab points buyers of the CM5 IO Board toward its separate Edge AI Expansion Board for Raspberry Pi 5 if they need NVMe SSD expansion, LTE/5G, or extra I/O.

• Vision only: LLM support is on DEEPX’s roadmap with no date set.
• INT8 only: DXNN quantization is automatic, with approximately 2 % accuracy loss vs FP32.
• Pi 5 only: Compute Module 5 on the CM5 IO Board, Pi 4, and non-Raspberry Pi boards are not supported.

The Wider Physical AI Race DEEPX Is Joining

The DEEPX AI HAT lands at a moment when Physical AI has moved from keynote to deployable product at major OEMs. At the same CES 2026 panel, Pete Bernard, CEO of Edge AI Foundation, said the data-center-centric AI approach “faces clear constraints regarding cost, power, and latency,” and pointed to Physical AI as the response. Sandeep Modhvadia, Chief Product Officer at Wind River, framed the same gap from the mission-critical side: “For AI to serve as infrastructure in defense, aerospace, and industrial sectors, security, predictability, and long-term stability are non-negotiable.”

DEEPX’s wager is that those constraints favor distributed inference at the edge, where a 3 W NPU on a Pi 5 can run vision models without touching the cloud. Whether 13 or 25 TOPS at INT8 is enough horsepower for the maker-to-production pipeline DEEPX is chasing is the open question. The same edge-AI race is reshaping other vendors’ hardware roadmaps, and ZOTAC’s push at COMPUTEX 2026 to ship Jetson T5000-based industrial modules shows how much bigger the field has become (see ZOTAC’s enterprise AI hardware push at COMPUTEX 2026).

DEEPX has framed its longer-term goal as becoming “the definitive global standard platform for Physical AI.” Kim added that the Pi 5 launch is “a cornerstone of our long-term strategy to cultivate future talent and developer communities.” Whether the 25 TOPS top end is enough to anchor industrial workloads at scale will be tested by the Hyundai robotics deployments already on the books for 2026.

Frequently Asked Questions

How much does the Sixfab AI HAT+ cost?

The cheaper DX-M1ML version lists at $63 and the higher-end DX-M1M version lists at $90 on Sixfab’s store, both priced as standalone add-on boards for the Raspberry Pi 5.

Does the AI HAT+ work with Raspberry Pi 4 or Compute Module 5?

No. Sixfab’s documentation lists the Raspberry Pi 5 as the only supported host platform. The Raspberry Pi 4, Compute Module 4, and the Raspberry Pi Compute Module 5 on the CM5 IO Board are not supported, and non-Raspberry Pi SBCs are explicitly excluded.

Can the AI HAT+ run large language models?

Not today. Sixfab’s documentation states that LLM support is on DEEPX’s roadmap, with no date set. The board ships with a vision focus covering object detection, classification, and segmentation.

What power supply does the board need?

The official Raspberry Pi 27 W USB-C PD power supply. Standard 5 V / 3 A supplies are insufficient and will trigger under-voltage warnings and throttling, per Sixfab’s documentation.

Is the AI HAT+ hot-pluggable?

No. Sixfab’s documentation is explicit: the Raspberry Pi 5 must be powered off and the USB-C cable disconnected before mounting or removing the AI HAT+.

Where does inference actually run?

Entirely on the device. After the runtime install, frames stay on the Raspberry Pi 5 unless the application explicitly forwards them, and no internet connection is required at runtime.