What does a three-watt NPU add to a Raspberry Pi 5?

Why three watts matters more than TOPS

Peak TOPS figures make good marketing. Most edge products live and die by their power and thermal budget. The real system still needs headroom for the Raspberry Pi itself, a camera, storage, networking and the device it actually controls. An accelerator that stays around three watts under sustained load means smaller enclosures, simpler cooling, and AI that can run continuously rather than in short bursts. That efficiency is what makes local AI practical in the places where it matters most: smart cameras, robots, industrial equipment, distributed sensors. Cloud latency, connectivity or data privacy would otherwise become the limiting constraint.

The three-watt figure refers to typical sustained NPU power for the 25 TOPS DX-M1M option. It is not complete system wall power. Always measure your final Raspberry Pi configuration with its actual camera, storage, workload and cooling.

One NPU handling three model classes

Most edge products begin with perception. A CNN detects an object, segments a work area, estimates a pose. It can also flag an abnormal event. The next generation of edge systems also needs to understand context and communicate what it sees. The DX-M1M is designed to support that whole journey on a single piece of silicon and one toolchain.

CNN workloads cover object detection, classification, segmentation, pose estimation. They also handle depth and image enhancement. They turn camera pixels into structured events. The DEEPX ModelZoo ships pre-optimised examples including YOLO-family detectors, segmentation models. It also includes pose models. DX-Stream helps assemble them into GStreamer-based camera pipelines.

Compact vision-language models add contextual understanding on top of visual results. They describe a scene, answer a focused question about an image. They can also interpret an event in natural language locally with nothing leaving the device. Small language models can interpret a user command, summarise local events. They can also generate a concise explanation, creating a natural interface for cameras, robots, machines without routing every interaction through the cloud.

Memory and accuracy approach

The first question any developer asks about language models at the edge is how much memory is available. The DX-M1M module carries 2 GB of dedicated on-module LPDDR4x. That comfortably supports VLM and SLM workloads alongside vision models.

Accuracy matters more than a peak benchmark. Edge AI performance is only useful if the model stays accurate on the images and conditions your application actually faces. DEEPX treats INT8 optimisation as an accuracy-aware engineering process, not a simple conversion step. Start with a known FP32 accuracy baseline and a clearly defined application metric. Calibrate with representative data from the real deployment environment, not only clean sample images. Compile and optimise the model with DX-COM into a portable .dxnn deployment artifact. Compare accuracy, latency, power and thermal behaviour on the final Raspberry Pi system. This matters most in the difficult conditions edge cameras actually meet: low light, glare, motion blur, occlusion, unusual angles. A credible result shows both the efficiency gain and the retained application accuracy.

Measured performance numbers

Numbers beat adjectives. The figures below were measured on a Raspberry Pi 5 (8 GB) running the shipping Sixfab software release; this compares the DX-M1M and DX-M1 NPUs. Mobilenet_v2 at 240×240 runs at 2361 FPS on the DX-M1M and 3223 FPS on the DX-M1. Deeplabv3plus at 512×512 runs at 155 FPS on the DX-M1M and 231 FPS on the DX-M1. Qwen3-1.7B with 96 prefill tokens shows a time-to-first-token of 599.04 ms and 4.64 tokens per second on the DX-M1M. The same model on the DX-M1 shows 544.58 ms and 11.60 tokens per second. The DX-M1 uses LPDDR5 while the DX-M1M uses LPDDR4X. The performance difference comes from their memory bandwidths.

From first boot to first inference

The fastest way to understand the platform is to verify the hardware, run a packaged model and watch the NPU work. On Raspberry Pi OS run the apt update and install commands for the Sixfab repository and packages. Then run_hello_world to execute a packaged example. Then dxtop to observe the NPU in real time. From there pick an optimised ModelZoo workload or bring your own supported ONNX model through DX-COM. The resulting .dxnn file runs through DX-RT's C++ or Python APIs. Your Raspberry Pi application keeps handling cameras, business logic, user interfaces, networking, and control.

What you can build with it

Start at home. A camera on your workbench that recognises when your 3D print has failed and messages you a snapshot with a one-line description. A front-door camera that tells the difference between a courier, the neighbour's cat, and someone loitering. It also summarises the day's events each evening entirely locally. A garden or pet monitor that only alerts you about things worth seeing.

All of these follow the same loop. A CNN sees the event. A compact VLM adds context. An SLM explains it or interprets your next command. Your Raspberry Pi application decides and acts. The same loop scales to serious deployments: continuous local detection in smart cameras with reduced cloud dependence. Low-power visual quality and safety monitoring beside a production line. Perception and compact multimodal intelligence on robots while the Raspberry Pi runs ROS 2 and control logic. Privacy-aware occupancy and behaviour analytics in retail and smart spaces.

Where it stops being the right tool

Like any edge accelerator the AI HAT+ is not trying to compete with cloud inference. Applications that need broad world knowledge, very long conversational context or continuous learning will always run better where compute and memory are effectively unconstrained. The DX-M1M is at its best running tightly scoped always-on intelligence next to a camera or sensor. Privacy, latency, offline operation and a single-digit-watt power budget are the requirements that actually decide the design. Being honest about that boundary is exactly what makes the platform trustworthy inside it.

The 25 TOPS Sixfab AI HAT+ for Raspberry Pi 5 is available now from Sixfab. The 13 TOPS version is expected to launch toward the end of 2026. How many of the 25 TOPS units are already in the field the announcement does not say.