Building Silicon Brain Cube for Green and Trustworthy AI

This project aims to pioneer the Silicon Brain Cube, a groundbreaking hardware architecture with a three-dimensional implementation capable of enhancing energy efficiency of demanding AI workloads while supporting trustworthy processing. We will carry out joint research based on the expertise in energy-efficient deep learning models and in three-dimensional hardware architecture of the Japanese team, and in trustworthy AI design and in multi-level static and dynamic optimization and tools of the UK team. We will innovate an AI model that enables implementations to best achieve user-defined trade-offs between performance, resources required, energy efficiency, predictive accuracy, and level of uncertainty for trustworthy AI. More than an order of magnitude improvement in energy efficiency would be obtained by novel strategies for reducing memory accesses and irregular/sparse processing on a wired-logic and in-memory reconfigurable computing fabric. To promote research and practice of the Silicon Brain Cube beyond the end of the project, an open-source repository will be developed containing documented designs and tools, as well as online tutorials and application studies.