ECE Seminar Wednesday, September 28

Electrical & Computer Engineering

Graduate Student Seminar

Lin Jiang

Will present a talk entitled:

“Fast and accurate thermal simulation of multi-core CPUs based on a data-driven learning method”

SC160

Wednesday, September 28 2022

4pm

Abstract: As the technology node aggressively downscaled and the power density in semiconductor chip continues increasing, the temperature-related issues are becoming one of the major challenges for high performance computing processors. Many dynamic thermal management techniques have been proposed in the past to reduce the temperature and to suppress the hot spots in modern processors. Effective thermal management however requires the accurate temperature profile of the processor. Therefore, the demand for efficient architecture-level dynamic thermal simulations for the real-time thermal management of multi-core processors has been rapidly growing in recent years. In addition to simulation efficiency, a high resolution is desirable in order to accurately predict crucial hot spots and high thermal gradients in the chip. This work investigates a data-driven learning-based simulation technique derived from the proper orthogonal decomposition (POD) for the architecture-level dynamic thermal simulation of multi-core processors.

The POD projects the heat transfer problem onto a functional space constituted by a finite set of basis functions (or POD modes) that are generated (or trained) by the thermal solution data collected from direct numerical simulations (DNSs). The accuracy, efficiency and robustness of the POD simulation approach influenced by the quality of thermal data, especially in the areas with high thermal gradients, are examined thoroughly. Our investigation revealed that, if the POD modes are trained by the good-quality data, the POD simulation is able to offer an accurate prediction with a resolution as fine as the DNS within and beyond the training range, using an extremely small degree of freedom (DoF). As a consequence of the significant reduction in DoF, the POD approach can achieve a computing speedup by several orders of magnitude, compared to the FEM simulation. Since the accuracy of the POD model is strongly determined by the quality of training data collected from thermal simulation tools, the POD approach can also be used to rigorously determine the accuracy of numerical solutions offered by the simulation tools.

Bio: Lin Jiang received his B.S. degree from Wuhan University, China, in 2016 and the M. S. degree from University of Science and Technology of China in 2019. He is currently pursuing his Ph.D. degree in Electrical & Computer Engineering department at Clarkson University, where he is co-advised by Prof. Yu Liu and Prof. Ming-Cheng Cheng. His current research interests include thermal modeling of multi-core CPUs/GPUs using a data-learning approach, reduced-order model and finite element method. He was a recipient of the Prof. Avram Bar-Cohen Best Paper Award from the 21 st Intersociety Conference on Thermal and Thermomechanical Phenomena in Electronic Systems in June 2022.

Also on zoom: https://clarkson.zoom.us/j/97971978951?pwd=dHFJZlAvZndvNnRZTzBoYTU0bk03dz09

Meeting ID: 979 7197 8951
Passcode: 145872
One tap mobile
+16469313860,,97971978951# US
+16468769923,,97971978951# US (New York)

*Co-Sponsored by IEEE student branch and HKN

You Might Also Like