A Scalable Black-Box Optimization System for Auto-Tuning VLSI Synthesis Programs

Modern logic and physical synthesis tools provide numerous options and parameters that can drastically impact design quality; however the large number of options leads to a complex design space difficult for human circuit designers to navigate. We tackle this parameter tuning problem with a novel system employing intelligent search strategies and parallel computing, thus automating one of the key design tasks conventionally performed by a human designer. We provide an overview of this system, called SynTunSys, as well as results from employing it during the design of the IBM z13 22nm highperformance server chip, currently in production. During this major processor design, SynTunSys provided significant savings in human design effort and achieved a quality of results beyond what human designers alone could achieve, yielding on average a 36% improvement in total negative slack and a 7% power reduction.