The realization of exascale computing will require overcoming semiconductors' mounting energy consumption issues, and Sandia National Labs' Erik DeBenedictis outlines three scenarios for evolving supercomputers to the 1- to 50-exaflop range.
The first case he cites involves finding a lower-voltage transistor or millivolt switch to enable continued reduction in energy per operation so existing software performance can be raised with minimal rewriting. However, DeBenedictis notes the electronics industry has little control over the time frame for the switch's discovery and commercialization.
The second scenario he offers involves more three-dimensional (3D) integration of processors and memory, although it entails enough architectural alteration to require retuning existing code for performance. DeBenedictis sees a likely industry development of 3D technology for storage and mobile devices separate from supercomputers, but he cautions "industry will spend [research and development] funds on developing the technology and is likely to demand premium prices for new product."
The third scenario calls for an application-specific architecture, which may be a challenge for software and algorithms and carries the pitfall of the supercomputer only being proficient at the function manufactured into the system. "This scenario is very likely to present the programmer with idiosyncratic architectures that require extensive experience by the programmer, and may lead to code that is not easily repurposed to other applications," DeBenedictis says.
From Scientific Computing
View Full Article
Abstracts Copyright © 2015 Information Inc., Bethesda, Maryland, USA