A case for code-representative microbenchmarks

Abstract

Microbenchmarks are fundamental in the design of a microarchitecture. They allow rapid evaluation of the system, while incurring little exploration overhead. One key design aspect is the thermal design point (TDP), the maximum sustained power that a system will experience in typical conditions. Designers tend to use hand-coded microbenchmarks to provide an estimation for TDP. In this work we make the case for a systematic methodology to automatically generate code-representative microbenchmarks that can be used to drive the TDP estimation.