Google Research has open-sourced a new architecture, enabling AI to run directly on smartwatches and headphones without relying on the cloud.

Google Research recently announced the open - sourcing of the Coral NPU platform, which is a full - stack open - source solution for hardware engineers and AI developers. It helps them overcome the current bottlenecks that hinder the implementation of artificial intelligence in wearable and edge devices, including limited performance, fragmented ecosystems, and lack of user trust.

The core goal of the Coral NPU is to enable all - day AI applications to run efficiently on battery - powered devices. Meanwhile, it provides flexible configuration options for high - performance scenarios to achieve a balance between energy consumption and computing power.

For AI to truly play the role of an "assistant", such as actively helping users plan schedules, translating conversations in real - time, or understanding the physical environment, it must be able to run locally on devices worn or carried by users. This requirement brings fundamental challenges: how to embed environment - aware AI in edge devices with limited power, enabling the devices to operate independently of the cloud, thus achieving a private, secure, and truly all - day intelligent experience.

Google researchers pointed out that hardware devices built with the Coral NPU can support a variety of AI application scenarios, including user activity and environment perception, audio - video processing (such as voice recognition, real - time translation, and face recognition), and gesture recognition.

• During the process of integrating AI into wearable and edge devices, the Coral NPU platform focuses on solving three key problems:
• Bridging the gap between the limited computing power of edge devices and the high requirements of cutting - edge large models;
• Alleviating the development difficulties caused by the fragmentation of the device ecosystem and unifying the proprietary chips and hardware environments of different manufacturers;

Ensuring that user data is protected from unauthorized access and strengthening privacy protection.

At the architectural level, the Coral NPU redefines the chip logic through a reverse - design concept: instead of prioritizing traditional scalar computing, it focuses on the machine - learning matrix engine, optimizing for AI from the underlying silicon to achieve more efficient local inference performance.

In terms of privacy protection, the Coral NPU uses technologies such as CHERI. Through a refined memory - level security mechanism and a scalable software isolation system, it builds a security sandbox enforced by hardware to ensure the security of user data.

This platform is built based on a set of IP modules that comply with the RISC - V Instruction Set Architecture (ISA). Its basic design can achieve a performance of 512 GOPS (giga - operations per second) while consuming only a few milliwatts of power. In contrast, the early non - open - source version of Google Coral could achieve 4 TOPS (tera - operations per second), but its power consumption was about 1 watt.

The Coral NPU platform consists of three core components:

• A scalar core that manages data flow;
• A vector execution unit compatible with the RISC - V vector instruction set;
• A matrix execution unit for accelerating neural network computing.

At the programming level, the Coral NPU architecture is deeply integrated with modern C compilers (such as IREE and TFLM) and supports multiple machine - learning frameworks, including TensorFlow, JAX, and PyTorch.

To further improve performance, Google's research team has developed a comprehensive and complex toolchain: Machine - learning models built using TensorFlow, JAX, or PyTorch are first converted into a common intermediate representation (MLIR), then undergo multi - level progressive lowering to gradually approach the low - level hardware language, and finally are compiled into binary files for efficient deployment.

It is worth mentioning that Google Research has also collaborated with Synaptics to jointly develop the first IoT processor based on this new architecture, providing the first example for the implementation of the Coral NPU.

Currently, the Coral NPU platform has been open - sourced on GitHub.

Original link: https://www.infoq.com/news/2025/10/google-coral-npu-platform/

This article is from the WeChat official account "InfoQ". Author: Sergio De Simone. Republished by 36Kr with permission.