Sixfab AI HAT+ for Raspberry Pi 5
Request a quote- Ships from Germany in 1–3 business days
- 2-year warranty
- Secure checkout
Description
Build and test vision AI models directly on Raspberry Pi 5. Sixfab AI HAT+ delivers compact DEEPX AI acceleration for rapid prototyping and local inference.
Run vision AI workloads on Raspberry Pi 5 in real time. Locally, no cloud, no GPU. Plug in the HAT+, install one APT package, and ship inference on your own hardware. As an Official Raspberry Pi Design Partner, Sixfab integrates the DEEPX DX-M1M family directly onto a HAT+ compliant board, giving Pi 5 developers production-grade NPU acceleration over native PCIe without leaving the Raspberry Pi ecosystem.
Performance at a glance
- Up to 25 TOPSINT8 · DEEPX DX-M1M
- PCIe Gen 3 ×1Native Pi 5 PCIe
- ~3 WTypical NPU draw
Highlights
- 25 TOPS at INT8DEEPX DX-M1M, 2 GB LPDDR4X · 13 TOPS variant with DX-M1ML.
- PCIe Gen 3 ×1Native Pi 5 PCIe via 16-pin FFC cable. No USB hops, no bandwidth bottleneck.
- ~3 W typical3 W typical NPU draw · ~13-15 W combined Pi 5 + HAT+ under load on the official 27 W PSU.
- HAT+ spec compliantRaspberry Pi HAT+ EEPROM auto-config · 56.5 × 65 mm · stacking-friendly.
- Soldered NPUSoldered DEEPX silicon transfers heat into the PCB far more efficiently than a socketed M.2 card. No socket to fail, no module slop, no third-party variability.
- APT installSigned Sixfab repository ships dxrt-runtime, kernel driver, and tools. Update with apt update.
- DXNN SDKBring ONNX models from PyTorch, TensorFlow, or Keras. Compile with DX-COM. Deploy with the C++ or Python runtime.
Technical specifications
| AI accelerator | DEEPX DX-M1M (25 TOPS) or DX-M1ML (13 TOPS) at INT8 |
|---|---|
| NPU memory | 2 GB LPDDR4X (DX-M1M) · 1 GB LPDDR4X (DX-M1ML) |
| Host interface | PCIe Gen 3 ×1 over 16-pin FFC cable |
| Form factor | Raspberry Pi HAT+ · 56.5 × 65 mm · 6.56 mm tall |
| Power input | 5 V via the Pi 5 2×20-pin header and PCIe connector |
| NPU power draw | 3 W (typical) |
| Cooling | Passive |
| Operating temp | 0-70 °C commercial |
| Supported host | Raspberry Pi 5 |
| Host OS | Raspberry Pi OS (Trixie) |
| Runtime | dxrt-runtime · APT install · Python & C++ APIs |
| Model pipeline | ONNX → DXNN via DX-COM compiler |
| Hot-plug | Not supported. Power off Pi 5 before mounting |
| Compliance | |
| Certification | Certification in progress: CE, FCC, UKCA, RoHS, REACH |
Two SKUs · one board
Same PCB, same HAT+ form factor, same software stack. The NPU module is the only difference. Pick the variant that fits your workload and budget.
| AI HAT+ 13 TOPS | AI HAT+ 25 TOPS | |
|---|---|---|
| NPU | DEEPX DX-M1ML | DEEPX DX-M1M |
| Precision | INT8 | INT8 |
| NPU memory | 1 GB LPDDR4X | 2 GB LPDDR4X |
| Best for | Single-camera, single-model | Multi-model, multi-camera |
| Typical scenario | Low-power projects, cost-sensitive builds | High-resolution streams, headroom for growth |
Available by Q4, 2026.
How it works
- 01CaptureMIPI CSI, USB UVC, or RTSP IP cameras feed frames into the Raspberry Pi 5 - up to 4×1080p.
- 02HostPi 5's Cortex-A76 CPU runs your app: pre-processing, control flow, I/O, network.
- 03InferenceDEEPX DX-M1M or DX-M1ML NPU runs your compiled DXNN model over PCIe Gen 3 ×1 - 25 TOPS · INT8.
- 04Use resultsDetections, segments, and classifications return to your app. Display, log, trigger, stream - 30-35 FPS.
Two paths to deployment
Option 1
Sixfab Model Zoo
Pre-compiled DXNN models - ready to run, no training required. YOLOv8n, MobileNet, ResNet, and more, already compiled for the DEEPX NPU.
Option 2
DEEPX DXNN SDK
Full custom model deployment: export your PyTorch, TensorFlow, or Keras model to ONNX, compile to DXNN with DX-COM, and run it through the Python or C++ runtime. INT8 quantization is automatic, with ~2% accuracy delta vs the FP32 source.
What you can build
- Video analytics camerasOn-device object detection, counting, intrusion analytics, and retail insights on a single Pi 5 unit.
- Robotics & autonomous systemsReal-time perception, object tracking, and navigation assistance on AMRs, robot arms, and visual-inspection rigs.
- Smart city & infrastructureTraffic monitoring, facility management, and safety systems on roadside Pi 5 units.
- Industrial automationDefect detection, quality inspection, and process monitoring on the production floor.
- Drones & autonomous systemsOn-board perception with low weight and ~3 W typical NPU draw.
- Edge servers & AIoTCompact inference nodes for multi-camera deployments.
Sixfab edge AI ecosystem
One NPU, one SDK, three form factors for real-world deployment.
-
AI HAT+This product
13 / 25 TOPS AI acceleration for Raspberry Pi 5.
-
Sixfab Edge AI Expansion Board25 TOPS · LTE/5G
Add LTE/5G connectivity, NVMe local storage, and multi-camera support in a single under-board stack.
-
ALPON X5 AI25 TOPS · −20 to +60 °C
Fanless, rugged, always-online edge AI computer for fleets and distributed industrial sites.
What's in the box
- Sixfab AI HAT+ board13 TOPS · INT8 (DEEPX DX-M1ML) or 25 TOPS · INT8 (DEEPX DX-M1M), depending on the option×1
- PCIe FFC cable (16-pin)×1
- 16 mm stacking header (2×20, 2.54 mm)×1
- M2.5 × 16 mm F-F spacer×4
- M2.5 × 5 mm plastic screw×8
- Passive cooler (with thermal pad)×1
Both options include the exact same 6-item mounting kit and assembly hardware; the only difference is which DEEPX NPU is soldered on the board. For Raspberry Pi 5 only.
Not included (sold separately)
| Required | Raspberry Pi 5 (host board) · official 27 W USB-C PD power supply for the Raspberry Pi 5 · microSD card flashed with Raspberry Pi OS |
|---|---|
| Recommended | Raspberry Pi Active Cooler for the Pi 5 (for sustained 100% NPU utilization) |
| Optional | USB or CSI camera (for live vision inference workloads) |
Compatibility
Supported host
- Raspberry Pi 5
Not supported
- Raspberry Pi 4
- Compute Module 4
- Compute Module 5 on the CM5 IO Board
- Non-Raspberry Pi SBCs
| Cameras | Raspberry Pi Camera Modules (MIPI CSI) · USB cameras (UVC) · IP cameras (RTSP) · Multi-camera configurations |
|---|---|
| Frameworks | ONNX (primary) · PyTorch · TensorFlow · Keras · Ultralytics YOLO (native integration coming soon) |
Frequently asked questions
What's the difference between 13 TOPS and 25 TOPS?
13 TOPS (DEEPX DX-M1ML): single-model, single-camera deployments. Low power. 25 TOPS (DEEPX DX-M1M): multi-model pipelines, multi-camera, high-resolution. Recommended for most projects. Both share the same PCB, HAT+ form factor, software stack, and DXNN SDK. Only the soldered NPU module differs.
How do I decide whether 13 or 25 TOPS is right for my project?
Size it from three things: how many camera streams and models you run at once, your target input resolution, and whether you want headroom for future models. Pick 13 TOPS (DX-M1ML) for single-camera, single-model pipelines, prototypes, and budget-sensitive builds at common resolutions. Pick 25 TOPS (DX-M1M) for multi-stream 1080p, higher-resolution single-stream, multi-model pipelines, or larger compiled models - it is the safe default when you want room to grow. To estimate before you buy, check the per-model FPS figures in the Sixfab Model Zoo against your target frame rate and stream count.
How does it compare to the Raspberry Pi AI HAT+ and AI HAT+ 2?
All three are HAT+ form-factor accelerators that mount on a Raspberry Pi 5 and run inference on a soldered NPU over PCIe. The difference is the silicon: Sixfab AI HAT+ uses DEEPX (DX-M1ML at 13 TOPS or DX-M1M at 25 TOPS, INT8); the Raspberry Pi boards use Hailo. The Raspberry Pi AI HAT+ ships as Hailo-8L (13 TOPS) or Hailo-8 (26 TOPS); for vision, Sixfab's 25 TOPS at INT8 is competitive with the 26 TOPS Hailo-8 model, and YOLOv8n at 640×640 runs 30-35 FPS on a Raspberry Pi 5 with 8 GB RAM. The newer Raspberry Pi AI HAT+ 2 (Hailo-10H, 40 TOPS at INT4, 8 GB on-board RAM) adds on-device LLMs and VLMs; the Sixfab AI HAT+ is vision-focused today. If your project specifically needs local generative AI right now, the Raspberry Pi AI HAT+ 2 is built for that workload.
Can I run LLMs on this?
No. DEEPX DX-M1M and DX-M1ML are optimized for computer vision (object detection, segmentation, classification). The current generation doesn't support LLMs. LLMs are on the DEEPX roadmap and Sixfab will support them as the silicon enables.
How long does setup take?
Under 15 minutes: power off the Pi 5, mount the AI HAT+, connect the 16-pin FFC cable, install the dxrt-runtime APT package, verify with lspci | grep DEEPX, run a Model Zoo demo, and see YOLOv8n at 640×640 hit 30-35 FPS on Raspberry Pi 5 with 8 GB RAM. Custom DXNN SDK deployments take 1-2 hours for ONNX export, DXNN compilation, and application integration.
Does it work offline?
Yes, completely. Inference runs entirely on-device over the PCIe Gen 3 x1 link between the Raspberry Pi 5 and the DEEPX NPU. No cloud, no GPU, no external connectivity required.
Which Raspberry Pi models are supported?
Raspberry Pi 5 is the only supported host. Not supported: Raspberry Pi 4, Compute Module 4, the Raspberry Pi Compute Module 5 on the Raspberry Pi CM5 IO Board, and non-Raspberry Pi SBCs. Hot-plug is not supported - power off the Raspberry Pi 5 before mounting or removing the AI HAT+.
What camera formats are supported?
Raspberry Pi Camera Modules (MIPI CSI), USB cameras (UVC), IP cameras (RTSP), and multi-camera configurations. Cameras connect directly to the Raspberry Pi 5; the AI HAT+ does not obstruct the Pi 5's CSI connectors.
What AI frameworks are supported?
ONNX (primary), PyTorch, TensorFlow, Keras, and Ultralytics YOLO (native integration coming soon). Models are exported to ONNX, then compiled to DXNN via the DEEPX DXNN SDK for execution on the NPU. The Sixfab Model Zoo includes pre-compiled models (YOLOv8n, YOLOv8s, MobileNet, ResNet, and others) ready to deploy.
