⚡ Edge AI: the Sixfab × DEEPX HAT brings 25 TOPS at 3 watts to the Pi 5

Raspberry Pi's official blog hosts a guest post by Sixfab and DEEPX on the AI HAT+ for Raspberry Pi 5, a third-party accelerator board built around the DEEPX DX-M1M NPU: 25 TOPS of dedicated acceleration at roughly 3 W of sustained NPU power, leaving the Pi's CPU free for camera handling, application logic and connectivity.

The technical message is counter-intuitive: what matters at the edge is not the peak TOPS on the box but the power and thermal budget. An accelerator that holds ~3 W under sustained load enables smaller enclosures, passive cooling and AI that runs continuously rather than in bursts — exactly what smart cameras, robots and industrial sensors need, where cloud latency and data privacy become the limiting constraints. The module even carries 2 GB of dedicated LPDDR4x, enough for compact language models alongside vision workloads.

The post walks through three classes of workload on one chip: classic CNNs (YOLO-family detection, segmentation, pose via the DEEPX ModelZoo and DX-Stream GStreamer pipelines), compact VLMs that describe a scene in natural language, and SLMs that turn the device into a local conversational system. That is the bigger 2026 trend: edge AI no longer just detects — it understands and responds without the cloud.

The authors also flag a golden rule: the 3 W figure is the NPU alone, not the whole system — always measure your real configuration. Worth noting that Raspberry Pi deliberately leaves the accelerator market to its ecosystem (its own AI HAT+, Hailo, and now DEEPX). Source: Raspberry Pi.