Computer Science Department

Computer Science Department

The Democratization Of Parallel Computing
David Luebke
NVIDIA
Feb. 8, 3:00 pm, McGl 020

Parallel computing had a "Golden Age" in the 1980s and early 1990s, as researchers and companies made tremendous strides in architecture, programming models, languages, and algorithms for parallel computing. The field of high-performance computing revolved around supercomputers like the MasPar and the Connection Machine that demonstrated the breathtaking potential of data-parallel computing. Unfortunately, their impact was limited by their equally breathtaking cost: parallel computing was an exotic, expensive technology practiced by a small group of acolytes attending a few supercomputers at major universities, corporations, and national laboratories. Eventually data-parallel supercomputers became economically infeasible, steamrollered by the incredible pace of commodity microprocessor technology. Exotic supercomputers were replaced by relatively mundane clusters of commodity PCs, and the field of data-parallel computing grew dormant. GPU computing has changed all that. NVIDIA's Tesla architecture platform provides a massively multithreaded architecture with up to 128 processor cores and thousands of threads in flight, programmable in C and capable of hundreds of billions of floating-point operations each second. Researchers throughout the scientific and engineering disciplines are using the accompanying CUDA software platform to speed up their code tenfold, a hundredfold, or even more.

This new wave of parallel architecture rides the wave of commodity graphics hardware - about forty million CUDA-capable processors have already shipped - thus benefiting from rather than battling against the commodity technology curve. GPU computing and CUDA represent the democratization of parallel computing. By bringing high-performance data-parallel processing to the masses, along with a programming model suitable for the manycore, hetereogeneous processors of the future, CUDA is bringing about a Renaissance of data-parallel programming models and algorithms. Researchers are rediscovering, rethinking, and improving upon classic algorithms and approaches from the Golden Age. In this talk I will describe the architecture and design of CUDA, give some examples that illustrate this revival of data-parallel algorithms, and close with some promising directions for future research.

Biography
David Luebke is a Research Scientist at NVIDIA Corporation, which he joined in 2006 after eight years on the faculty of the University of Virginia. He has a Ph.D. in Computer Science from the University of North Carolina and a Bachelors degree in Chemistry from the Colorado College. Luebke's principal research interests are general-purpose GPU computing and realistic real-time computer graphics. Specific recent projects include fast multi-layer subsurface scattering for realistic skin rendering, temperature-aware graphics architecture, scientific computation on graphics hardware, advanced reflectance and illumination models for real-time rendering, and image-based acquisition of real-world environments. Past projects include the book "Level of Detail for 3D Graphics", for which Luebke was the lead author, and the Virtual Monticello museum exhibit, which ran for over 3 months and helped attract over 110,000 visitors as a centerpiece of the major exhibition "Jefferson's America and Napoleon's France" at the New Orleans Museum of Art.