Title:

DataFlow SuperComputing for BigData DeepAnalytics

 

Abstract:

This presentation, possibly followed by an on-site or an on-line mini-course or a full-blown course on DataFlow Programming, analyses the essence of DataFlow SuperComputing, defines its advantages and sheds light on the related programming model. The stress is on issues of interest for Applications of Mathematics, Physics, Geo Sciences, and Civil Engineering. The DataFlow paradigm (which itself represents an application of graph theory), compared to the ControlFlow paradigm, offers: (a) Speedups of at least 10x to 100x and sometimes much more (depends on the algorithmic characteristics of the most essential loops and the spatial/temporal characteristics of the Big Data Streem, etc.), (b) Potentials for a better precision (depends on the characteristics of the optimizing compiler and the operating system, etc.), (c) Power reduction of at least 10x (depends on the clock speed and the internal architecture, etc.), and (d) Size reduction of well over 10x (depends on the chip implementation and the packiging/cooling technology, etc.). The bigger the data, and the higher the reusability of individual data items (which is typical of ML), the higher the benefits of the dataflow paradigm over the control flow paradigm. However, the programming paradigm is different, and has to be mastered. The ongoing research of the speaker has been highly influenced by four different Nobel Laureates: (a) from Richard Feynman it has been learned that future computing paradigms will be successful only if the amount of data communications is minimized; (b) from Ilya Prigogine it has been learned that the entropy of a computing system could be minimized if spatial and  temporal data get decoupled; (c) from Daniel Kahneman it has been learned that the system software should offer options related to approximate computing; and (d) from Tim Hunt it has been learned that the system software should be able to trade latency for precision. Special attention is given to two dataflow systems co-architected by the speaker, one tuned to physics of Silicon (Maxeler Groq for Higlhy Iterative Code) and the other tuned to physics of GaAs (Systolyc Array for Gram Schmidt Orthogonalization), both for applications in complex math, and math for computational physics and geo/atmo/astro/space-physics.

 

Bio:

About the Speaker:

Veljko Milutinovic received PhD from the University of Belgrade in Serbia, has been on faculty positions in the USA (more recently at Purdue University and Indiana University), in Europe (more recently at Technical University of Vienna and Technical University of Graz), and is credited for the DARPAs first GaAs (Gallium Arsendie) microprocessor at 200MHz (about a decade before mainstream and the DARPAs first GaAs Systolic Array with 4096 CPUs, plus for various innovations in the domain of the dataflow paradigm (arithmetic operations and graph mappings). His current academic research and industrial development interests are in dataflow acceleration of complex mathematical algorithms needing low power and high speed, for math-intensive applications in computational physics and geo sciences. This presentation has been delivered before, as a one-hour talk or a full-semester course, at a relatively large number of leading universities of the USA (MIT, NEU, UMass, Harvard, Michigan, Ohio, Illinois, Wisconsin, Purdue, Indiana, NYU, Columbia, FIU, FAU, etc...).