Holger Fröning from Ruprecht-Karls University of Heidelberg in Germany will give a guest lecture about

“Towards efficient communication methods and models for scalable GPU-centric computing systems “

on Wednesday, March 26th, 13:00, in our Seminarroom IDEG134, Inffeldgasse 16c, ground floor.

Abstract

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. While this helped to overcome the end of Dennard scaling and significantly increased the energy efficiency of computations, communication is left unaddressed. However, current analyses show that a large fraction of the power consumption originates to communication. Actually this fraction will increase for future technologies, making communication more expensive than computation in terms of energy and time. Being amplified by the advent of Big Data, we are observing a fundamental transition to communication-centric systems composed of heterogeneous computing units. This talk will review some of our explorations in the area of inter-GPU communication, discuss current options for data movements, and summarize our key findings. We will see that besides specialization in computation it is mandatory to also design specialized communication models, as existing models poorly match the execution model of such thread-collaborative processors.