In this video, Thomas Sterling from Indiana University presents: A Roadmap to Exascale. Recorded at the 2012 National HPCC Conference in Newport.
After years of talk, the task of achieving Exaflops sustained performance on practical and usable systems by the end of this decade has been initiated both in the US and internationally. Unlike the era leading up to Petaflops, it is accepted that government investment by many nations will be required to realize this goal and provide such resource capabilities. What is not clear is whether like the achievement of Petaflops, Exascale is simply a scaled up version of prior computing techniques or if an entirely new model is required. Now, real work has been undertaken in the US to begin to ferret out the answer to this question. Under DARPA sponsorship of the UHPC program and more recent DOE support the hardware and software technology base requirements and strategies is being explored in-depth. At risk is either disruption or worse failure. And of course, there is the possibility of getting it wrong. If conventional techniques cannot scale the factor of four orders of magnitude in concurrency probably required, across a broad range of applications, then Exascale will be unachievable through incremental techniques. But then, this requires a high-risk revolutionary strategy that may prove disruptive to legacy codes and skill-sets. This presentation will discuss the progress made in the US in the field of Exascale over the previous year, describe DARPA and DOE projects in this area, present recent results, and suggest their implications for the future.