A path towards average-case silicon via asynchronous resilient bundled-data design

Abstract

The periodic nature of the global clock in traditional synchronous designs forces circuits to be margined for the worst possible case of process, voltage, temperature, and data conditions. This constrains the silicon to operate at worst-case frequencies and at conservative supply voltages. Resilient architectures promise to remove these margins, by detecting and correcting timing errors when they occur, thereby creating the potential to achieve real average-case operation. However, synchronous resilient schemes previously proposed can suffer from multiple issues, including being susceptible to metastability and requiring often complex changes to the architecture to support replay-based recovery from timing errors. These problems respectively lead to circuit failures and/or incur high timing penalties when errors occur. This paper reviews a recently proposed asynchronous bundled-data resilient template called Blade that is robust to metastability issues, requires no replay-based logic, and has low timing error penalties. It also describes some open issues and new research opportunities this template presents, including automation problems to target average-case operation, specific circuit optimizations to minimize resiliency overhead, and the need for new test procedures to tune delay lines and screen out bad chips.