Post-Dennard Performance Scaling in it's Teenage Decade: picoJoule replaces nanosecond
Talk by Associate Professor Holger Fröning of Heidelberg University
Date/time: Wednesday, January 3rd, 2018, 15:00-16:00
Location: Universitetsparken 5, room 01-0-029 (APL meeting room)
The concurrency galore is currently defining computing at all levels, leading to a vast amount of parallelism even for small computing systems. Technology constraints prohibit a reversal of this trend, and the still unsatisfied need for more computing power has led to a pervasive use of accelerators to speed up computations. The inherent reason for these two fundamental transitions is a hard power budget, which is the result of the end of Dennard Scaling. As both concurrency and specialization can improve energy efficiency, they allow for more operations per Joule and thereby lead to performance improvements. While this has helped to overcome the first decade after the end of Dennard Scaling, it only improves the energy efficiency of computations, while data movements are left unaddressed. However, current analyses show that a large fraction of overall power consumption originates to communication. Even worse, this fraction will increase for future technologies, and additionally the inevitable use of heterogeneity and asymmetry will lead to more data movements.
This talk will review the implications of the resulting fundamental transition to communication-centric systems, which are dominated in performance and power consumption by data movements. While exploiting heterogeneity, asymmetry and concurrency for improved energy efficiency continues, the associated and fast-growing complexity increase will dramatically reduce productivity and usability. We therefore urgently need new techniques and methods to address this semantic gap between programmer and architecture, by hiding system complexity without any significant degrade of performance or energy efficiency.
Holger Fröning is an associate professor at the Department of Mathematics and Computer Science at the Ruprecht-Karls University of Heidelberg (Germany), leads the Computer Engineering Group at the Institute of Computer Engineering, and is deputy director of the same institute. His recent research interests focus on closing the semantic gap in between computer architecture and programmer, addressing asymmetry, heterogeneity, concurrency and related CMOS scaling effects. As a computer engineer, he is most interested in application-driven heterogeneous computing ranging from high-performance data centers to resource-constrained embedded devices, focusing on data movement optimizations, domain-specific languages and compilers, and associated power and energy aspects.
In 2016, he spent 5 months at NVIDIA Research (Santa Clara) as visiting scientist, sponsored by Bill Dally. Early 2015 he was visiting professor at the Technical University of Graz (Austria), sponsored by Gernot Kubin. From 2008 to 2011 he was senior researcher at the Parallel Architectures Group at the Technical University of Valencia (Spain), led by Jose Duato. He has received his PhD and MSc degrees 2007 respectively 2001 from the University of Mannheim, Germany. His publications have been accepted at top-tier conferences, workshops and journals; in addition, he contributed to several books. Parts of his research results have been commercialized. He chaired tracks for EuroPar 2015 and International Supercomputer Conference 2017, and recently served as program committee member for IPDPS2018, CCGRID2018, SC2017, ICPP2017/16, CLUSTER2016 and HiPC2015.
He leads the NVIDIA GPU Research and Education Center at Heidelberg University. Four of his publications have received a best paper award. His recent sponsors include BMBF, DFG, German Excellence Initiative, Google, SAP, NVIDIA, Xilinx, Micron, and the European Union. For more information, please visit his website: http://www.ziti.uni-heidelberg.de/compeng
All are welcome. No registration required.
Host: Fritz Henglein