Aula103 - S. Niccolò Building - Via Roma, 56, Siena
11:00-12:00 Friday September 20 2019
* TITLE: It is difficult to make predictions, especially about the future, but for sure it is heterogeneous
* ABSTRACT: Dennard observed in 1974 that, as transistors become smaller, their power density stays roughly constant, so that the power consumption of integrated circuits stays roughly constant. This scaling law broke down around the year 2006, which motivated the multicore revolution. Furthermore, there are also strong signs that Moore’s law is close to ending. For these reasons it seems that the only way to provide even higher performance (besides, as of yet, unproven technologies such as quantum computers) is heterogeneous computer systems. In fact, such heterogeneous systems have dominated the TOP500 list of supercomputers in recent years, since they achieve higher performance per Watt than homogeneous designs.
In this talk, I will briefly describe some of my research in this area. In particular, some of the research results obtained in the EU projects Encore, LPGPU, and LPGPU2 will be presented. In addition, compute power and data availability have been key machine learning drivers in recent years. I will also concisely present some recent works my team members and I performed in this domain, such as performance counters based power modeling of mobile GPUs using deep learning, and autotuning stencil computations with structural ordinal regression learning.
Embedded Systems Architecture (AES)
Institute of Computer Engineering and Microelectronics (TIME)
Faculty IV - Electrical Engineering and Computer Science
Berlin University of Technology
EN12, Einsteinufer 17, 10587 Berlin, Germany
Tel: +49 (0)30 314-73130
b.juurlink@tu-berlin.de
www.aes.tu-berlin.de