DEEPX带来25TOPS NPU改为Raspberry Pi 5 边AI发展

全文内容 Full Text

Home » Embedded Systems » DEEPX Brings 25-TOPS NPU to Raspberry Pi 5 Edge AI Development DEEPX and Sixfab enable local vision and language AI on Raspberry Pi 5 using a 3-watt NPU accelerator. DEEPX is bringing vision and language AI inference to the Raspberry Pi 5 through the Sixfab AI HAT+, which uses the company’s DX-M1M neural processing unit (NPU). The accelerator delivers up to 25 trillion operations per second (TOPS), with typical sustained NPU power consumption of approximately 3 watts. In an article published in Raspberry Pi’s official newsroom on Oct. 5, DEEPX and Sixfab demonstrated how the accelerator can run convolutional neural networks (CNNs), compact vision-language models (VLMs) and small language models (SLMs) on a single device. CNNs handle tasks such as object detection, image classification and anomaly detection. VLMs add scene interpretation and natural-language descriptions, while SLMs can process commands and generate brief explanations. One example involves monitoring a 3D printer. A CNN detects a printing failure, a VLM interprets the scene, and an SLM generates an explanation for the user. Similar configurations could support visual inspection, robotic perception and safety monitoring in industrial environments. The approximately 3-watt figure applies to the NPU alone, not the complete Raspberry Pi 5 system. Local inference can reduce dependence on cloud connectivity and external data transfers, an advantage for applications requiring low latency or offline operation. DEEPX provides software tools to support model deployment. DX-COM converts supported ONNX models into the company’s .dxnn format, while DX-RT runs models through C++ and Python APIs. DX-Stream supports GStreamer-based video processing pipelines, and the DEEPX ModelZoo offers preoptimized models for common vision tasks. The company says its runtime and model format are shared across development boards and industrial products, including the Sixfab AI HAT+ and ALPON X5 AI gateway. This allows developers to move validated models from prototyping to industrial deployments within the same software environment. Sixfab and DEEPX launched the AI HAT+ for Raspberry Pi 5 in June. The 25-TOPS version is available, while a 13-TOPS variant is expected toward the end of 2026. “For physical AI to scale, developers around the world need to be able to build AI applications easily and bring them into real products,” said Dr. Lokwon Kim, CEO of DEEPX. “Our goal is to provide the semiconductor technology, software tools and development environment that allow an experiment on Raspberry Pi to become a product deployed in the global market.”