February 2007
The Promise and Perils of the Coming Multicore Revolution and Its Impact
John L. Manferdelli, Microsoft Corporation

New applications

Can people user this much computing power? Yes.9 The ultimate application mix is hard to forecast (applications that need this level of computing don't, by definition, exist, and the application specialists will not invest the effort required until they see some hardware). Again, we can speculate.

It is uncontroversial that servers (including home servers) will also benefit from many-core computing, and this will also boost the need for powerful clients. With cheap and ubiquitous sensors and natural language processing, we can anticipate environment aware, multi-media (vision, speech, gesture, object recognition, etc.) input and output human computer interfaces that "learn" user behaviors and offer suggestions or possibly automatically manage some tasks for users.

Better data mining and modeling will provide business intelligence and targeted customer service. Automated medical imaging, diagnosis and well being monitoring will be commonplace. High-level tools like MATLAB or Excel, designed for parallelism, will take advantage of increased power and delegate processing across the network, provided the right workflow tools are integrated.

With terabyte disks, these systems will make superb media library, capture, edit, and playback systems. Film fans can purchase, download, and view protected feature films on opening day. Most printed material and other media can be replaced with electronic versions, accessed via a broadband connection, with vastly improved search and cross reference capabilities. These machines can make virtual reality and realistic games, well, real. Not only entertainment but education will benefit.

Today's corporate servers will shrink to a few racks and become highly resilient to failure. State check pointing and load balancing will improve performance and reliability. Damage from catastrophic failures is limited to a few seconds of downtime and rollback. Provisioning, deploying, and administering these servers and applications are simplified and automated.

Massively parallel computational grids built of commodity hardware already solve scientific problems like computational chemistry, protein folding and drug design. "Supercomputers" already analyze nuclear events and water tables and predict the climate and the economy. The power of these systems and the reach of these techniques will vastly improve with new hardware, and scientists will have supercomputers under their desks. By the way, scientific, financial and medical "supercomputing" are no longer "small" business opportunities. More than 10% of servers are used in scientific applications.

Classic computational techniques (known in the scientific community as the "seven dwarves"10 – including equation solvers, adaptive mesh modeling, etc.) will help explore regimes that will change our lives.11 Already, Microsoft researchers and world class scientists are using advanced computational techniques to explore potential cures for Aids and cancer, model Hydrologic activity in agriculturally sensitive regions, perform seismic modeling and run virtual laboratories for advanced physics. As in the past, use by scientists will help illuminate the path for the rest of us.

No time to waste

Programmable systems are playing an increasingly large part in our lives and, in many ways, provide a world-wide "paradigm shift" comparable to the appearance of cheap, mass market printing in scope and benefit. Many-core computers signal a shift in Computer Science, Computational Science, and classical Commercial Software that (as in all good technology shifts) marry the past advances of many "knowledge workers" as well as provide a new avenue for qualitatively new advances.

Acknowledgements The author is indebted to a number of colleagues at Microsoft for insight and review of this material including Craig Mundie, Burton Smith, David Callahan, Jim Larus, Jan Gray, Paul England, Brian LaMacchia and Tony Hey.
1 A related ILP technique breaks instructions into multiple cycles and attempts to execute different parts of successive instructions simultaneously, branch prediction is also needed here.
2 ILP also increases the "energy per useful computation" because of the discarded results and much larger controllers.
3 The cell processor does this.
4 In fairness, compilers, runtimes, frameworks and libraries were also improved to try to ameliorate this problem. See for example, Allen and Kennedy in the references.
5 Larus, J., Sutter, H. "Software and the Concurrency Revolution," ACM Queue, Vol. 3, No. 7, pp 54–62, September 2005.
6 Larus, J. R. , Rajwar , R. Transactional Memory. Morgan & Claypool, 2006.
7 Allen, R., Kennedy, K. Optimizing Compilers for Modern Architectures. Elsevier, 2002.
8 This made perfect sense when CPU's were expensive and memory subsystems were roughly comparable in speed.
9 We have a well documented historical answer if not a complete proof.
10 So named by Phil Collela of LBL, these include the important computational kernels for modeling and analysis.
11 Asanovic, K. et. al., "The Landscape of Parallel Computing Research: A View from Berkeley," UCB/EECS-2006-183.

Pages: 1 2 3 4 5 6

Reference this article
Manferdelli, J. "The Many-Core Inflection Point for Mass Market Computer Systems," CTWatch Quarterly, Volume 3, Number 1, February 2007. http://www.ctwatch.org/quarterly/articles/2007/02/the-many-core-inflection-point-for-mass-market-computer-systems/

Any opinions expressed on this site belong to their respective authors and are not necessarily shared by the sponsoring institutions or the National Science Foundation (NSF).

Any trademarks or trade names, registered or otherwise, that appear on this site are the property of their respective owners and, unless noted, do not represent endorsement by the editors, publishers, sponsoring institutions, the National Science Foundation, or any other member of the CTWatch team.

No guarantee is granted by CTWatch that information appearing in articles published by the Quarterly or appearing in the Blog is complete or accurate. Information on this site is not intended for commercial purposes.