JIE Seminar: Towards Scalable Spectral Sparsification ofGraph Laplacians and Integrated Circuits

Updated : September 1, 2016
Press release by Yingneng Gu

Time:      10:00am~11:00am, Friday, Sept.2

Venue:     Lecture Hall l02, JIE Building, East Campus of SYSU

Speaker:  Prof.Zhuo Feng, Michigan Technological

Host: Dr.Kai Wang from JIE iSense Laboratory


Present-day nanoscale integrated circuits(lCs) are integrating billions of transistors into a single chip, while even key subsystems, such as clockdistribution networks, power delivery networks(PDNs), embedded memoryarrays, as well as analog and mixed-signal systems, may reach an unprecedented complexity of hundreds of millions of circuit components. As a result, it becomes extremely difficult and even intractable to model, simulate, optimize and verify future nanoscale ICs at large scale with existing design methodologies.0n the other hand, emerging spectral graph sparsification techniques allow to construct ultra-sparse subgraphs(a.k.a. graph sparsifiers) that can well approximate the spectra of the original graph, leading to the development of much faster numerical and graph-based algorithms. For instances, spectrallysparsified transportation networks allowto develop much faster navigation(routing)algorithms in large transportation systems, spectrally sparsified social networks allow to more efficiently understand and predict information propagation phenomenon in large social networks, spectrally sparsified circuit networks allow to more efficientlysimulate, optimize and verify large circuit systems, etc.

ln this talk, l will first describe our recent spectral perturbation based approach for scalable spectral sparsification of large graphs and integrated circuits, and its role in designing nearly-lineartime numerical algorithms forsolving large symmetric, diagonally dominant (SDD) matrices, as well as scalable design automation algorithms that are critical for designing future nanoscale microprocessors, 3D- lCs, analog/mixed-signal circuits, as well as RF and microwave ICs. ln the last, l will discuss howto leverage graph sparsification techniques for optimallysolving large sparse matrices on emerging heterogeneous parallel CPU-GPU computing platforms.


Dr. Zhuo Feng received the Ph.D. degree in ComputerEngineering from TexasA&M University, College Station, TX in2009. He is currently an associate professor in the Department of Electrical and Computer Engineering of Michigan Technological University. His current research interests includeVLSl design and computer-aided design(CAD), heterogeneous parallel computing as well as graph-theoretic algorithms forsolving sparse matrices. He received a Faculty Early Career Development(CAREER)Award from the National Science Foundation(NSF) in2014, a Best PaperAward from ACM/lEEE Design Automation Conference(DAC) in2013, and two Best PaperAward Nominations from IEEE/ACM lnternational Conference on Computer-Aided Design(lCCAD) in2006and2008. He is the principle investigator of the CUDA Research Center at Michigan Technological University named by Nvidia Corporation. He has served on the technical program committees of major international conferences in the areas of electronic design automation(EDA) and VLSl design, including DAC, ICCAD,ASP-DAC, ISQED, etc. He has been a technical reviewerfor many leading IEEE/ACM journals as well as a panelist for NSF and DoE proposals.

JIE Seminar-20160902