SHANGHAI -- VeriSilicon (688521.SH) announced the joint launch of the Coral NPU IP with Google, targeting always-on, ultra-low-energy edge Large Language Model (LLM) applications. The IP is based on Google’s foundational research in open machine learning compilers and enhanced with AI security features, providing developers with a unified open-source platform to build a robust edge AI ecosystem.

The Coral NPU is built on the open RISC-V instruction set architecture, featuring native tensor processing capabilities, supporting mainstream machine learning frameworks such as JAX, PyTorch, and TensorFlow Lite (TFLite), and utilizing open-standard tools like Multi-Level Intermediate Representation (MLIR) from the Low Level Virtual Machine (LLVM) project for compiler infrastructure. It is designed for always-on, ultra-low-energy edge AI applications, particularly for wearable devices and ambient sensing systems.

The Coral NPU IP is now open-sourced on the Google Developers website and available to developers worldwide. VeriSilicon will offer a commercial-ready, enterprise-grade version of the IP and leverage its rich IP portfolio and chip design and verification capabilities to offer one-stop custom silicon services. Currently, VeriSilicon is working on a validation chip based on the Coral NPU IP, targeting AI/AR glasses and smart home applications, accelerating LLM deployment at the edge.

“This builds on our prior experience on the Open Se Cura open-source project and represents the continued deep integration of Google’s open-source technology with VeriSilicon’s chip design and commercialization capabilities,” said Wiseway Wang, Executive Vice President and General Manager of Custom Silicon Platform Division at VeriSilicon. “We will continue to leverage our strengths in chip design, verification, and system-level optimization to advance the edge AI ecosystem, support the deployment of open-source technologies in real-world products, and enable edge LLM applications, providing strong support for industry innovation.”