Home         Authors   Papers   Year of conference   Themes   Organizations        To MES conference

A Machine Learning-Based Switching Power Prediction at Floorplan Stage of IC Physical Design
Authors
	Janpoladov V.A.
Date of publication
	2022
DOI
	10.31114/2078-7707-2022-3-45-52
Abstract
	This paper demonstrates the effectiveness of using machine learning algorithms to predict the switching power at the floorplan stage of the physical design for a specific block architecture. Since the design process of integrated circuits (ICs) includes iterative stages for circuit optimization and time-to-market is critical for the industry, it is crucial to develop effective methods to evaluate the parameters of the IC at the early stages of the design process. The effectiveness of the proposed method was demonstrated for a block containing a 64-bit Arithmetic Logic Unit (ALU) coupled with General-Purpose Registers (GPRs). This block has been designed for 486 different configurations and scenarios. The study was carried out for the following cell groups: registers, sequential cells, combinational cells, and cells in clock network. To solve the regression problem of switching power prediction five algorithms were selected and studied: Polynomial Regression, Decision Tree Regression, Random Forest Regression, Supported Vector Regression and Multi-layer Perceptron (MLP). To evaluate prediction accuracy Root Mean Squared Error, R2 and Mean Absolute Percentage Error (MAPE) metrics were used. The main goal was to identify the algorithm with a minimum MAPE value for each cell group. The best performance with lowest MAPE values was obtained with MLP algorithms. The MAPE value for registers is ~8.35%, for serial cells ~18.13%, for the clock network ~2.36%, and for combinational cells ~3.77%. The proposed method does not use technology-dependent data, which makes it universal for any technology nodes used to design different blocks. The disadvantage of the method is the need for implementation of the whole design flow for a selected design with the selected range of parameters for collecting necessary training data, which requires additional time and machine resources.
Keywords
	dynamic power, prediction, machine learning, integrated circuits, regression, neural networks, arithmetic logic unit
Library reference
	Janpoladov V.A. A Machine Learning-Based Switching Power Prediction at Floorplan Stage of IC Physical Design // Problems of Perspective Micro- and Nanoelectronic Systems Development - 2022. Issue 3. P. 45-52. doi:10.31114/2078-7707-2022-3-45-52
URL of paper
	http://www.mes-conference.ru/data/year2022/pdf/D053.pdf