Fine-Grained Power Modeling of Multicore Processors Using FFNNs

Abstract To minimize power consumption while maximizing performance, today’s multicore processors rely on fine-grained run-time dynamic information—both in the time domain, e.g. $$\mu $$ μ s to ms, and space core-level. The state-of-the-art for deriving such information is mainly based predetermined models which use linear modeling techniques determine core-performance/core-power relationship. However, with becoming ever more complex, cannot capture all possible core-performance related states anymore. Although artificial neural networks (ANN) have been proposed coarse-grained of servers resolutions range seconds, few works yet investigated ANN-based modeling. In this paper, we explore feed-forward (FFNNs) core-level estimation rates 10 kHz. achieve a high accuracy minimizing overhead, propose multi-objective-optimization architecture using NSGA-II FFNNs being trained performance counter data from complex-out-of-order processor architecture. We show that relative error highest FFNN decreases average by 7.5% compared approach 5.5% multivariate polynomial regression model. For optimized both between 4.1% 6.7% offering significantly lower overhead FFNN. Furthermore, micro-controller-based an accelerator-based implementation inference area negligible.