Application Specific Electronic Design Synthesis: Neural Network Silicon Compiler

Future weapon systems require significant advances in computing systems scale and functionality. Neural networks hold considerable promise to deliver very high performance and powerful new capabilities, as well as compact physical size, low power and high reliability. Such electronic neural networks are very attractive to system developers. To date actual systems incorporating electronic neural networks have been limited (and expensive), due to a lack of flexibility in constructing custom network architectures with the available hardware modules. We propose to develop electronic design synthesis tools for constructing custom integrated circuits that implement application-specific neural networks. These tools will assist the application engineer in selecting between broad classes of analog and digital implementation technologies, architectures, interconnect topologies, interfaces, and permit direct circuit synthesis of manufacturable integrated circuits. The tools would also interface to simulators and development systems, to assist the designer in validating the neural network using experimental and training data. In Phase I, we will prove the tools by FABRICATING demonstration circuits built on existing Tanner Research technology. Our approach will utilize commercial integrated circuit technology; products arising from this R&D can be fabricated reliably and economically by a number of vendors.