Metastability Tolerant Computing

Synchronization using flip-flop chains imposes a latency of a few clock cycles when transferring data and control signals between clock domains. We propose a design scheme that avoids this latency by performing synchronization as part of state/data computations while guaranteeing that metastability is contained and its effects tolerated (with an acceptable failure probability). We present a theoretical framework for modeling synchronous state machines in the presence of metastability and use it to prove properties that guarantee some form of reliability. Specifically, we show that the inevitable state/data corruption resulting from propagating metastable states can be confined to a subset of computations. Applications that can tolerate certain failures can exploit this property to leverage low-latency and quasi-reliable operation simultaneously. We demonstrate the approach by designing a Network-on-Chip router with zero-latency asynchronous ports and show via simulation that it outperforms a variant with two flip-flop synchronizers at a negligible cost in packet transfer reliability.