Synchronization between independently clocked regions in a high performance system is often subject to latencies of more than one clock cycle. We show how the latency can be reduced significantly, typically to half the number of clock cycles required for high reliability, by speculating that a long synchronization time is not required. The small number of synchronization metastability times longer than normal are detected, and subsequent computations re-evaluated, thus maintaining the reliability of the system.
D. J. Kinniment, Alexandre Yakovlev