ANN-Based Performance Prediction in MoCs

Abstract

Due to high integration density and technology scaling, the manycore networks-on-chip (NoCs) often experience higher evaluation time by traditional simulations for a set of common performance characteristics. Artificial intelligence (AI) is being employed as an altered solution over the simulation-based performance evaluation. However, many AI techniques’ accuracy and estimation time are low and high. This paper proposes an artificial neural network (ANN) based framework to quickly and more accurately evaluate mesh-based NoCs (MoCs). Experiments show that the essential performance metrics latency, throughput, and energy consumption are 16.53–40.76 cycles, 7.63–76.46 × 10 - 3 flits/cycle/IP, and 1417–1625 μ J, respectively. The proposed ANN framework achieves an accuracy of up to 96%. It is around 50% more compared to many previous works. © 2022, The Author(s), under exclusive license to Springer Nature Switzerland AG.