• Technical Conference:  30 March – 03 April 2025
  • Exhibition: 01 – 03 April 2025
  • Moscone Center, San Francisco, California, USA

Betting on Exascale

By Casimer DeCusatis Ph.D.

What will the world be like in 2020?  Will it be a science fiction paradise with robot moon bases and synthetic human brains? Although the future is notoriously difficult to predict, it can sometimes be done with surprising accuracy; for example, take a look at Isaac Asimov’s view of what 2014 would be like, written at the 1964 world’s fair.

One hot topic for the futurists at OFC 2014 is bound to be the outlook for exascale computing, and the role of optical interconnects in making it a reality.  As with any discussion about the future, this can be a controversial topic. Some predictions say we’re about a decade away from realizing optically enabled exascale systems. High performance computing (HPC) consortiums worldwide have been drafting strategies and roadmaps for the coming revolution.  It’s hard to recall the last time such intense effort has been applied on a machine nearly a decade away from reality.  On the other hand, the Deputy Director at the Lawrence Berkely National Laboratory’s National Energy Research Scientific Computing Center just wagered $2000 of his own money that exascale won’t happen by the end of this decade.

This view is understandable, since most of the world is arguably still running their HPC applications at terascale, and just starting to deploy petascale systems. We seem to be getting ahead of ourselves talking about systems a million times more powerful than what most people are running today.  So why do we really need optically enabled exascale computing systems?

Other than assuming that we’ll keep inventing new applications for ever increasing compute power (a corollary to Parkinson’s Law), there are some potential applications for exascale on the horizon.   New applications like big data and predictive analytics hold tremendous potential (including their impact on visualization, another hot topic at OFC).  These applications are one reason why IBM recently made a billion dollar investment in Watson, the analytics platform that rose to fame by defeating the world champion at the game show Jeopardy! While Watson currently relies on more conventional networking technology, future HPC applications for public cloud, social media, and other systems are looking towards optical interconnect.  Late last year, Intel teased their interest in 1.6 Tbit/s optical interconnects for next generation servers. Given this interest in optical system interconnect and big data, it’s not hard to imagine a future where exascale systems become widely available at low cost, opening up a host of new possibilities.

On the other hand, the path to exascale isn’t a straightforward extension of our current technology; it can require fundamentally re-engineering the computational process.  One of the biggest obstacles is power consumption, particularly the energy required to move bits across a chip, compute node, or multi-rack server cluster.  While power efficiency per processing core has improved, interconnect power consumption hasn’t scaled nearly as well.  At current rates, by 2020 it will cost far too much energy to move data around inside an HPC cluster using current approaches. One possible answer is optical interconnects; even relatively small incremental gains in power efficiency can make a huge difference in a system that uses over 10 quintillion FLOPS per second.  So don’t be surprised if the discussion at OFC turns to building lasers that run a few picojoules more efficiently than before; it could make the difference on whether or not we ever approach the ultimate model of computational efficiency, the human brain, which runs on a scant 20 MW.

How are we going to address the role optics might play in future exascale compute systems?  Perhaps the answer lies in education, another area which OFC has consistently emphasized over the years. While we don’t have a good idea about what skills might be needed in the computing workforce a decade from now, if we prepare ourselves with a fundamental understanding of the issues and technology, we’ll be better able to adapt to our future environment.  Our future exascale workforce is currently in high school, at best; how can we prepare them for the challenges ahead?

These are only a few of the tantalizing questions for us to ponder as we look to the future during OFC 2014.  We may not have a clear view from our crystal ball, but I’m sure that won’t stop us from having some lively discussions about 2020.  Personally, I’ll just be glad if I still look as good as Brad Pitt at 56. To find out for yourself, look me up during OFC 2014, or drop me a line on Twitter (@Dr_Casimer).

Posted: 5 February 2014 by Casimer DeCusatis Ph.D. | with 0 comments

Blog post currently doesn't have any comments.
 Security code

The views expressed in this blog are those of the authors and do not necessarily reflect the views or policies of The Optical Fiber Communication Conference and Exposition (OFC)  or its sponsors.