Seminar: How to Build Dependable Computing Systems with Incomplete (or Wrong!) Data
Professor, Virginia Tech
Friday, October 25, 2019
11:15am - 12:15pm
2150 Torgersen Hall
This talk will present three in depth case studies of how future computing systems and system software can deliver high performance and meaningful results from data that is incomplete or even wrong. The first case study explores how to architect convolutional neural networks for deep learning with guaranteed accuracy without actually training the networks. We will look into a deep learning system called TAPAS, which automatically explores the architectural space of neural networks by using a minimal set of extracted data set features. The system can “train” more than 100 neural networks per second on a single inexpensive GPU. Our second case study explores a range of iterative algorithms for solving systems of partial differential equations by using dynamic precision tuning (“transprecision”) arithmetic. We will look into how an algorithm can characterize its input on-the-fly to avoid accuracy checks, identify opportunities for the cancellation of errors that may arise while the algorithm executes. Our third and final case study looks into how to build main memory (DRAM) systems with wider operating margins (meaning higher performance and better energy-efficiency) than conventional counterparts, at the risk of causing data corruption. We will look into how data corruption can be accurately captured and characterized in order to be mitigated with extremely high probability.
Dimitrios Nikolopoulos is the John W. Hancock Professor of Engineering and Professor of Computer Science at Virginia Tech. He does research in system software for large-scale computing and new computing paradigms. He is the recipient of a Royal Society Wolfson Research Merit Award, the NSF CAREER Award, the DOE CAREER Award, the IBM Faculty Award, the SFI-DEL Investigator Award and Best Paper Awards from premier IEEE and ACM conferences in his research area, including SC, PPoPP, and IPDPS. He has mentored 28 PhD students and 16 post-doctoral research fellows over his career. His research has been supported with extensive funding ($38.9m as Principal Investigator and $101.7m as a CoInvestigator) awarded competitively from NSF, DOE, EPSRC, SFI, NI DfE, Royal Academy of Engineering, Royal Society, the European Commission, IBM, SAP and Intel. Dimitrios has taught courses in computer organization, computer architecture, parallel programming, operating systems and embedded systems. He is a Distinguished Member of the ACM for outstanding scientific contributions to computing, Fellow of the Institute of Engineering Technology (FIET), Fellow of the British Computer Society (FBCS), and Senior Member of the IEEE. He earned PhD (2000), MSc (1997) and BEng (1996) degrees in Computer Engineering and Informatics from the University of Patras and he was trained as a postdoc at the University of Illinois at Urbana-Champaign.