Featured Community Response

Print Friendly, PDF & Email
Do you believe a roadmap exists that will get us to Exascale?

Thomas Sterling

Thomas Sterling

A roadmap may incorporate the design and establish the sequence of actions required to manufacture such a design based on expectation of the enabling technologies at the time to completion. The cost, MTBF, and power unspecified, this roadmap need only satisfy the criterion of a workload class that can deliver 1018 operations per second. Such roadmaps have been suggested by colleagues on both sides of the Pacific Rim. Such a system if too narrow in application, too difficult to program, too limited in uninterrupted execution time, consuming too much power to be deployable at but a few sites, and too expensive to acquire and operate may be considered a “stunt machine” whose principal value is a form of shallow and brief status. Within this perspective, I think the answer is “yes”, there is a roadmap (perhaps several) to realize an exascale stunt machine sometime around the end of this decade.

A roadmap may incorporate a set of established concepts and understood methods that when appropriately instituted in a future design that intersects a viable technology operating point at a given time to deliver practical systems for general applicability. Within this perspective, I think the answer is “no”. There is no roadmap that defines a set of development steps from known ideas in a proven synthesis that will deliver a general purpose, reliable, programmable, and cost effective (including power of operation) exascale computer.

A roadmap may be a plan to conduct a research program directed at key, specifiable, quantifiable challenges the results of which will inform a new design and operational methodology to ultimately deliver practical exascale systems of general applicability. Within this perspective, I think the answer is “yes”. There is a research strategy that serves as a high level roadmap, although only partly instituted, which if carried out would result in general purpose, reliable, programmable and cost effective exascale computers.

A roadmap should be a research and development plan to answer the critical questions about efficiency, scalability, programmability, generality, reliability, and energy that is devised and accepted by national leadership and technical contributors, funded by government and industry, and led by responsible representatives with vision of mission-driven stakeholders for sufficient duration to realize this achievement. Within this perspective, I think the answer is “no”. Due to a failure of vision and courage shared across the necessary constituent institutions, both political and technical, there is no consensus and supported roadmap, in the US or elsewhere that will deliver the necessary innovation and persevere through the daunting challenges to the final realizable conclusion. But there could be; there should be.

Thomas Sterling
Professor, School of Informatics and Computing
Chief Scientist and Associate Director, CREST
Indiana University

For related stories, visit The Exascale Report Archives.