Skip to Main Content
Architectural simulators are traditionally used to study the design trade-offs for processor systems. The simulators are implemented in a high-level programming language or a hardware descriptive language, and are used to estimate the system performance prior to the hardware implementation. The simulations, however, may need to run for long periods of time for even a small set of design variations. In this paper, we propose a machine learnt (neural network/NN) model for estimating the execution performance of a superscalar processor. Multiple runs for the model are finished in less than a few milliseconds as compared to days or weeks required for simulation-based methods. The model is able to predict the execution throughput of a processor system with over 85% accuracy when tested with six SPEC2000 CPU integer benchmarks. The proposed model has possible applications in computer architecture research and teaching.
Date of Conference: 5-8 Aug. 2007