Sixfab AI HAT+ Brings 25 TOPS and 3W Power to Raspberry Pi 5

Sixfab and DEEPX have launched the AI HAT+ for Raspberry Pi 5, a third-party accelerator that provides 25 TOPS of dedicated AI acceleration at approximately 3 W of typical sustained NPU power. (NPU: A dedicated accelerator for AI workloads designed to handle the heavy math of neural networks efficiently.) This hardware enables the Raspberry Pi 5 to run vision-language models (VLM) and small language models (SLM) locally, which matters to developers building autonomous systems where cloud latency and data privacy are primary constraints. (VLM: A type of AI that can process images and generate descriptive text or answer questions about what it sees.)

The 3 W Advantage: Prioritizing Thermal Headroom over Peak TOPS

Sixfab and DEEPX claim the DX-M1M NPU delivers 25 TOPS of performance while maintaining a 3 W typical sustained power draw. This power efficiency is the primary design headline because most edge products are limited by their thermal budget rather than peak processing numbers. An accelerator that stays near 3 W allows for smaller enclosures and passive cooling, which is essential for systems that must run continuously rather than in short bursts.

Sustained power efficiency allows for a more reliable deployment in environments like smart

Sustained power efficiency allows for a more reliable deployment in environments like smart cameras or industrial equipment. By keeping the NPU’s power footprint low, the system retains enough thermal headroom for the Raspberry Pi 5 itself, a camera, and storage. This combination of 3 W sustained power and 25 TOPS performance allows for smaller enclosures and passive cooling while maintaining high throughput.

Three Classes of Intelligence on a Single Chip

The DX-M1M NPU supports three distinct levels of intelligence on one piece of silicon: CNNs for perception, compact VLMs for context, and SLMs for interaction. Convolutional Neural Networks (CNN) are used for visual tasks like object detection and pose estimation, turning camera pixels into structured data. (CNN: Neural networks used for visual tasks like object detection and pose estimation by turning camera pixels into structured data.) Compact Vision-Language Models (VLM) add a layer of contextual understanding, allowing the device to describe a scene or answer a focused question about an image without sending data to the cloud.

Small Language Models (SLM) provide a natural interface by interpreting user commands

Small Language Models (SLM) provide a natural interface by interpreting user commands or summarizing local events. Because these models run locally, the system avoids the latency and privacy risks associated with cloud-based LLMs. The DX-M1M is designed to support this entire journey—from perceiving an event to explaining it—within a single toolchain.

Dedicated Memory Enables Local Language Models

The DX-M1M module carries 2 GB of dedicated on-module LPDDR4x memory to handle the memory-intensive requirements of language models. Most edge AI developers worry about model fit; this dedicated memory helps ensure that VLM and SLM workloads do not compete for the Raspberry Pi 5’s main system RAM. Dedicated memory allows these larger models to run locally without competing for the Raspberry Pi’s main system RAM.

Having 2 GB of dedicated memory changes the feasibility of local interaction. It allows the device to maintain a larger model state in memory while the Raspberry Pi 5 CPU handles other tasks like connectivity and application logic. This hardware architecture specifically targets the transition from simple perception to complex, context-aware interaction.

The Accuracy-Aware Engineering Pipeline

DEEPX treats INT8 optimization—reducing a model’s size or precision so it fits on local hardware—as an accuracy-aware engineering process rather than a simple conversion. The pipeline begins with a known FP32 accuracy baseline and a defined application metric, then calibrates the model using representative data from the real deployment environment. This is critical for handling difficult conditions like low light, glare, or motion blur.

The final step involves compiling the model into a portable.dxnn deployment artifact

By calibrating with real-world conditions (glare, motion blur), the hardware maintains application accuracy during quantization. The final step involves compiling the model into a portable.dxnn deployment artifact using the DX-COM tool. (DX-COM: A tool used for compiling models into a portable deployment artifact for the hardware.) This helps ensure that the efficiency gains of the NPU do not come at the cost of the accuracy required for the specific application.

Software Workflow: From First Boot to Inference

The software stack includes DX-COM for model compilation, DX-RT for deployment via C++ or Python APIs, and the DEEPX ModelZoo for pre-optimized examples. Users can verify the hardware and run a packaged example in minutes by installing the Sixfab repository and running the provided hello world scripts. The system provides a clear path from an ONNX model to a functional.dxnn artifact.

Developers can observe the NPU in real time using the dxtop tool to monitor

Developers can observe the NPU in real time using the dxtop tool to monitor performance. Once the model is deployed, the Raspberry Pi 5 remains free to handle camera feeds, business logic, and user interfaces. This separation of concerns is what allows for the smooth integration of high-level intelligence into low-level hardware control.

Operational Boundaries: What the DX-M1M is Not For

The Sixfab AI HAT+ is not designed to compete with cloud-scale inference or large-scale LLMs. Because it is a low-power edge device, it cannot provide broad world knowledge, very long conversational contexts, or continuous learning capabilities. It is at its best running tightly scoped, always-on intelligence where privacy and a single-digit-watt power budget are the primary requirements.

If a project requires a model that can answer general-purpose questions or maintain a long-running conversation history, this hardware is the wrong choice. In those cases, developers should look toward high-end GPU clusters or cloud inference APIs. The DX-M1M is a specialized tool for localized, immediate interaction next to a sensor.

Who This Hardware Matters To

Robotics engineers can use the Sixfab AI HAT+ to offload vision tasks while the Raspberry Pi 5 handles ROS 2 and control logic. These engineers should compare ROS 2 execution overhead on the Raspberry Pi with and without the Sixfab AI HAT+ to see how much logic they can reclaim. The NPU frees the CPU to handle complex robot logic while providing 25 TOPS of vision/language power.

Smart camera developers

Smart camera developers can use the device to distinguish between couriers, pets, and loiterers without cloud latency. These developers should compare the response time of local Sixfab AI HAT+ inference against a standard Cloud API to quantify the privacy gains. The device can run continuous local detection and summarize events entirely on-device.

Industrial automation specialists

Industrial automation specialists can deploy the NPU for privacy-aware safety monitoring on production lines. These specialists should compare the detection capabilities of the Sixfab AI HAT+ against the logic limits of a standard PLC. The 3 W power budget allows for simple, passive cooling in industrial environments.

Who this is not for: Cloud-based LLM developers

Cloud-based LLM developers should avoid this hardware if they need to build applications that require broad world knowledge or long conversational context. For these users, documentation for high-end GPU clusters or cloud inference APIs remains the correct path. This hardware is optimized for localized “intelligence in a box” rather than “intelligence in the cloud.”

Warehouse Robot Scenario: Worked Example

Kwame runs a small firm that builds smart warehouse robots. He needs a robot that can identify damaged packages and explain the defect to a nearby operator in real-time. Kwame selects the Sixfab AI HAT+ for the Raspberry Pi 5 and uses DX-COM to compile a CNN for damage detection, a VLM for describing the damage type, and an SLM to generate the spoken explanation.

The robot operates with zero cloud latency and maintains privacy, as all processing happens on the 2 GB of dedicated LPDDR4x memory. If Kwame needed the robot to have broad world knowledge (like a general-purpose GPT-4), he would have to abandon the local NPU and use a cloud-connected GPU.

Practical Applications for Edge Intelligence

What this replaces: The Sixfab AI HAT+ allows developers to replace cloud-based inference

The Sixfab AI HAT+ allows developers to replace cloud-based inference with local NPU processing. By moving the workload to the edge, developers eliminate cloud latency, connectivity constraints, and data privacy concerns. This is particularly useful for applications where even a few hundred milliseconds of round-trip time is unacceptable.

The NPU also reduces

The NPU also reduces the need for CPU-bound AI models that currently struggle with high-resolution vision or language tasks. By offloading these workloads to the DX-M1M, the Raspberry Pi 5 CPU is freed to handle application logic, user interfaces, and networking. This separation improves the overall responsiveness of the system.

Unverified Capabilities and Next Steps

Sixfab has not officially confirmed the specific ROS 2 offloading overhead figures, though the device is designed to support such workflows. The 13 TOPS version of the hardware is scheduled for late 2026, which may change the performance-per-watt calculations for some users. Verification of the 3 W power figure depends on the final camera and storage configuration used by the end user.

Key Terms

NPU

Neural Processing Unit: A dedicated accelerator for AI workloads designed to handle the heavy math of neural networks efficiently.

VLM

Vision-Language Model: A type of AI that can process images and generate descriptive text or answer questions about what it sees.

SLM

Small Language Model: A compact version of a large language model designed to run on local hardware with limited memory.

TOPS

Tera Operations Per Second: A standard measure of the peak processing power of an NPU.

Original reporting: Seeing, understanding, and responding: low-power CNN, VLM and SLM workloads on Raspberry Pi 5 (Raspberry Pi), on 5 October 2026.