COVER FEATURE GUEST EDITORS’ INTRODUCTION

Rebooting : New Strategies for Technology Scaling

Thomas M. Conte, Elie Track, nVizix Erik DeBenedictis, Sandia National Laboratories

Year-over-year exponential computer performance scaling has ended. Complicating this is the coming disruption of the “technology escalator” underlying the industry: Moore’s law. To fundamentally rethink and restart the performance-scaling trend requires bold new ways to compute and a commitment from all stakeholders.

10 COMPUTER PUBLISHED BY THE IEEE COMPUTER SOCIETY 0018-9162/15/$31.00 © 2015 IEEE ociety’s increasing reliance thread performance scaling uneco- on and demand for comput- nomical. Multicore processing was only ing power over the past half- a partial solution: parallel execution century has been ably met by accelerates some problems, and using Sexponential performance increases, yet parallelism requires rewriting software the heady march toward smaller, faster, instead of simply recompiling. Mean- cheaper, and more energy-e cient while, if Moore’s law is to resume, it computing technologies is slowing must be based on an architecture other as Moore’s law is being disrupted. In than multicore. fact, some argue that exponential per- What new technological escala- formance scaling ended a decade ago. tors are within reach, and how might But before society confronts this sober they impact computing performance? new era of limitations in which perfor- These were questions that inspired the mance supply no longer meets demand, launch of IEEE’s Rebooting Computing we must rethink—even reboot— Initiative (http://rebootingcomputing computing technology to revive his- .ieee.org) in . Thus far, the initia- toric exponential performance growth. tive’s work supports the notion that we This issue of Computer explores are nowhere near the end of computing some of the most promising new tech- performance growth; in fact, we are nologies, architectures, and engineer- entering an era of even more powerful ing strategies to fuel continued com- escalators. To get there, though, there puting improvements and what needs must be a real commitment from both to be done to make that happen. government and industry to maximize Moore’s law has had a really good and the remaining bene ts of Moore’s law; relatively long run as far as technology furthermore, research supporting a trends go. Reducing dimensions on the technological paradigm shift is neces- surface of a D chip enabled more capa- sary to reboot computing progress. ble devices, with higher energy e - ciency and higher clock rates. It charac- IN THIS ISSUE terized a golden era in which software As John M. Shalf and Robert Leland could be developed independent from argue in the  rst article, “Computing architecture, architecture independent beyond Moore’s Law,” ideas to reboot from implementation, and implemen- computing will require substantial tation independent from devices. Each funding and resources to reach the computing domain—software engi- marketplace. The authors were in u- neering, computer design, semicon- ential in developing the National Stra- ductor electronics, and so on—thrived tegic Computing Initiative (NSCI), with only an occasional need for inter- launched by the Obama administra- disciplinary collaboration. tion in late July  and tasked with In , physics intervened. The developing technology that could con- power densities of a CMOS-based micro- tinue scaling for the next decade. Shalf processor made continued single- and Leland describe the need for a new

DECEMBER 2015 11 GUEST EDITORS’ INTRODUCTION

and sustained R&D agenda to evalu- with traditional programming meth- as well as enormous supercomputing ate emerging semiconductor materials ods or legacy software. datacenters and the servers that han- and device physics. In “Energy-Efficient Abundant-Data dle smaller systems’ backend comput- How will we move from power- Computing: The N3XT 1,000×,” a ing. With trillions of dollars in annual ful numerical computing to the next cross-institutional team of authors revenue, the industry is large enough level—one at which we can not only introduce Nano-Engineered Com- to leverage room-temperature tech- analyze vast datasets but learn from puting Systems Technology (N3XT), nology for mobile devices and faster, them as well? The next five articles which reinterprets Moore’s law to be more power-efficient cryogenic tech- reflect our reasoned judgment of the more amenable to 3D manufacturing: nology for datacenters to exact the “top 5” new computing approaches instead of reducing dimensions in a right mix of benefits. In “Adapting to Thrive in a New Economy of Memory Abundance,” Kirk M. Bresniker, Sharad Singhal, and R. Stanley Williams describe how IF MOORE’S LAW IS TO RESUME, IT the availability of persistent mem- MUST BE BASED ON AN ARCHITECTURE ory, enabled by novel device phys- OTHER THAN MULTICORE. ics, could lead to replacement of the current processor-centric computing model based on Moore’s law with a memory-driven computing model. Until recently, the relative cost and that have the greatest likelihood of chip’s XY plane, their approach retains performance advantages of comput- driving performance improvement. fixed dimensions in the XY plane and ing throughput versus accessing data There is an important caveat: each increases the number of features in the storage has limited the development article is written by just one group of Z dimension. N3XT is based on carbon of data-intensive applications. How- researchers, and many other quali- nanotubes and memristors, but other ever, the emergence of 3D memory fied groups in the various fields have possibilities exist. The resulting com- makes such applications more attrac- different and very valuable perspec- puters could demonstrate benefits when tive and could serve as a key technol- tives. It is this community of insight- programmed with existing languages, ogy driver for data analytics. ful researchers—who both challenge but they could also be programmed In “Architecting for Causal Intel- and cooperate with one another—that with new neuromorphic methods. ligence at Nanoscale,” Santosh collectively empowers our faith in con- In “Superconducting Computing in Khasanvis and his colleagues out- tinued computing progress. Large-Scale Hybrid Systems,” D. Scott line the advantages of randomness in Three of the articles advocate the Holmes, Alan M. Kadin, and Mark W. conjunction with new neuromorphic reorganization of computing at the Johnson challenge the assumption programming methods. Randomness lower technology layers. They outline that all computers must use the same is contrary to current expectations research strategies aimed at improv- underlying technology. Highlighting that computers produce determinis- ing systems that look similar to those the speed and energy-efficiency advan- tic results and that any deviation is an in use today—utilizing existing pro- tages of superconducting electronics, error. Algorithms that make use of ran- gramming languages, software, and they propose Josephson junctions as dom sampling as a fundamental aspect manufacturing methods. The other two the new active elements for logic gates of their function can be executed articles present alternative approaches in the implementation of larger instal- on conventional computers using that are far more disruptive and reflect lations that could execute much of the software-based­ random-number­ gen- a fundamental change in the current existing software base. Of course, the erators, but nanoscale devices whose computing model. These methods offer computing industry includes intrin- behavior is random in well-​understood greater potential to achieve perfor- sically small systems such as Internet and controlled ways could be vastly mance gains to justify less compatibility of Things gadgets and smartphones more efficient. The authors propose a

12 COMPUTER WWW.COMPUTER.ORG/COMPUTER class of such devices that is effective in a probabalistic reasoning context calculation, rather than mere numeri- ABOUT THE AUTHORS cal analysis, to address a wide range of problems including . THOMAS M. CONTE is a professor with joint appointments in the Schools of Finally, in “Ohmic Weave: Computer Science and Electrical and Computer Engineering at Georgia Tech. Memristor-­Based Threshold Gate Net- His research interests include and compiler code gen- works,” David J. Mountain and his col- eration. Conte received a PhD in electrical engineering from the University of leagues describe how today’s Turing-­ Illinois at Urbana−Champaign. He is the 2015 IEEE Computer Society president, derived machines could incorporate co-chair of the IEEE Rebooting Computing Initiative, and a Fellow of IEEE. Con- neuromorphic capabilities. Neuro- tact him at [email protected]. morphic computing research often focuses on understanding the brain’s ELIE TRACK is CEO of nVizix LLC, a startup developing novel photovoltaic tech- learning and pattern-­recognition nology for solar power based in Stamford, Connecticut. His research interests capabilities and how best to reproduce include innovative high-efficiency solar cells as well as superconducting elec- them in computers, leading to super- tronics and its applications in high-performance communications and com- computer clusters that can learn from puting. He received a PhD in physics from Yale University. Track is co-chair of visual stimuli such as YouTube videos. the IEEE Rebooting Computing initiative, past president of the IEEE Council on Although these clusters model enor- Superconductivity, and a Fellow of IEEE. Contact him at [email protected]. mous numbers of neurons and syn- apses, they are organized into just a ERIK DEBENEDICTIS is a technical staff member in the Non-Conventional handful of layers. This contrasts with Computing Technologies Department at Sandia National Laboratories. His the multiple overlapping hierarchi- research interests include computing approaches across the entire technol- cal structures in software that are ogy stack, including further scaling of von Neumann architectures through much more amenable to engineering device improvements and packaging strategies, brain-inspired computing by groups of people. Integrating these approaches, and superconducting electronics. DeBenedictis received a PhD approaches could lead to a computer in computer science from Caltech. He is a member of the IEEE Rebooting with brain-like learning structures Computing Initiative, IEEE, the IEEE Computer Society, the IEEE Supercon- that are composed hierarchically like ductivity Council, ACM, and the American Physical Society. Contact him at programs, subroutines, operating sys- [email protected]. tems, and subroutine libraries.

he direction of US science and technology policy has become issue, hundreds of technical volunteers has played a seminal role by directing Tincreasingly clear during the have shaped our thinking through the federal resources and national atten- time in which we prepared this special Initiative. The various companies, aca- tion to important strategic goals. issue, and the articles that appear here demic institutions, and government closely align with the NSCI vision. We research labs that collaborate on the hope that they will stimulate your International Technology Roadmap own ideas regarding novel strategies for Semiconductors (ITRS 2.0; www to enable continued technology scal- .itrs2.net) have also contributed to ing, and we invite you to share your the technical discourse in fundamen- thinking with IEEE’s Rebooting Com- tal ways. In addition, the US Office of Selected CS articles and puting Initiative. Science and Technology Policy (www columns are also available for Although we alone selected the .whitehouse.gov/administration/eop free at http://ComputingNow .computer.org. final lineup of articles for this special /ostp), under the Office of the President,

DECEMBER 2015 13