Articles

The article introduces Google's Coral NPU, a comprehensive, open-source platform designed to overcome critical challenges in deploying powerful, always-on AI on low-power edge devices and wearables. It highlights three main issues: the performance gap between complex ML models and device limitations, the fragmentation tax from diverse proprietary processors, and the user trust deficit concerning data privacy. Coral NPU addresses these with an AI-first hardware architecture, prioritizing ML compute from the silicon up, based on RISC-V ISA. It offers ultra-low power consumption (512 GOPS at milliwatts) and a unified developer experience through integration with compilers like IREE and TFLM, supporting major ML frameworks. The platform is co-designed for current vision/audio models and future small transformer models, aiming to bring LLMs to wearables. Emphasizing hardware-enforced privacy (e.g., CHERI support), Coral NPU seeks to build user trust. Google is also fostering an ecosystem, partnering with companies like Synaptics, whose new Torq™ NPU is the first production implementation of the Coral NPU architecture. The platform aims to accelerate innovation in ambient sensing, wearables, mobile, and IoT devices.