Back to the Future: The Case for Reversible Computing

There is one, and only one way, consistent with fundamental physics, that the efficiency of general digital computation can continue increasing indefinitely, and that is to apply the principles of reversible computing. We need to begin intensive development work on this technology soon if we want to maintain advances in computing and the attendant economic growth NOTE: This paper is an extended author’s preprint of the feature article titled “Throwing Computing Into Reverse” (print) or “The Future of Computing Depends on Making it Reversible” (online), published by IEEE Spectrum in Aug.-Sep. 2017. This preprint is based on the original draft manuscript that the author submitted to Spectrum, prior to IEEE edits and feedback from external readers. Since the dawn of the transistor, technologists, and the world at large, have grown accustomed to a steady trend of exponentially-improving performance for information technologies at any given cost level. This performance growth has been enabled by the underlying trend, described by Moore’s Law, of the exponentially-increasing number of electronic devices (such as transistors) that can be fabricated on an integrated circuit. According to the classic rules of semiconductor scaling, as transistors were made smaller, they became simultaneously cheaper, faster, and more energy-efficient, a massive win-win-win scenario, which resulted in concordantly massive investments in the ongoing push to advance semiconductor fabrication technology to ever-smaller length scales. Unfortunately, there is today a growing consensus within industry, academia, and government labs that semiconductor scaling has not very much life left; maybe 10 years or so, at best. Multiple issues that come into play as we dive deeper into the nanoscale mean that the classic scaling trends are losing steam. Already, the decreasing logic voltages required due to various short-channel effects resulted in the plateauing of clock speeds more than a decade ago, driving the shift towards today’s multi-core architectures. But now, even multi-core architectures face the looming threat of increasing amounts of “dark silicon,” as heat dissipation constraints prevent us from being able to cram any more operations per second into each unit of chip area, due to the energy that is converted to heat in each operation. Fundamentally, achieving higher performance within a system of any given size, cost, and power budget requires that individual * This work was supported by the Laboratory Directed Research and Development program at Sandia National Laboratories and by the Advanced Simulation and Computing program under the U.S. Department of Energy’s National Nuclear Security Administration (NNSA). Sandia National Laboratories is a multimission laboratory managed and operated by National Technology and Engineering Solutions of Sandia, LLC., a wholly owned subsidiary of Honeywell International, Inc., for NNSA under contract DE-NA0003525. Approved for public release SAND2017-9424 O. M.P. Frank Back to the Future: Extended preprint Sandia National Labs The Case for Reversible Computing for IEEE Spectrum Version 5.7 arXiv:1803.02789 [cs.ET] Page 2 of 19 3/7/2018 7:48 PM operations have to become more energy-efficient, and the energy efficiency of conventional digital semiconductor technology is beginning to plateau for a variety of reasons, all of which can ultimately be traced back to fundamental physical issues. Looking forward, as transistors become smaller, their per-area leakage current and standby power increases; meanwhile, as signal energies are decreased, thermal fluctuations become more significant, eventually preventing any further progress within the traditional computing paradigm. Heroic efforts are being made within the semiconductor industry to try to allay and forestall these problems, but the solutions are becoming ever more expensive to deploy, with new leading-edge chip fabrication plants (“fabs”) now costing on the order of $10 billion each. But, it’s worth pointing out that no level of spending can ever defeat the laws of physics. Beyond some point that is, now, not very far away, a new conventionally-designed computer that simply has smaller transistors would no longer be any cheaper, faster, or more energy-efficient than its predecessors, and at that point, the progress of conventional semiconductor technology will stop, being no longer economically justifiable. The writing is on the wall. Obviously, however, we would prefer if the progress in the cost-efficiency of information technology were not to stop, since a large portion of our potential future economic progress would be empowered by the continuing advancement of this technology. So then the question arises, can we perhaps keep progress in computing going by transitioning over to some new technology base that is not “conventional semiconductor technology?” Unfortunately, some of the most crucial fundamental physical barriers that will prevent conventional complementary metal-oxide-semiconductor (CMOS) technology from advancing very much further will also still apply, in a more or less comparable way, to any alternative technology as well, as long as we insist on maintaining the present-day computing paradigm, namely irreversible computing. No other irreversible “beyond CMOS” technology can ever be very much better than end-of-the-line CMOS—at most, it will be better only by some relatively modest, limited factor. However, for several decades now, we have known that there exists a theoretically possible alternative computing paradigm, called reversible computing. Developing reversible computing (and then continuing to improve it) is in fact the only possible way, within the laws of physics, that we might be able to keep computer energy-efficiency and cost-efficiency for general applications increasing indefinitely, far into the future. So far, the concept of reversible computing has not received very much attention, which has perhaps made sense up until now, since it is indeed highly challenging to implement effectively, and the alternative of advancing conventional technology was much easier. Nevertheless, significant conceptual progress on reversible computing has been made over the decades by the small number of researchers pursuing it. Still, many difficult problems remain to be solved, and it is going to require a much larger effort, looking forwards, to address them. But, this effort will be highly worthwhile, because the potential upside that reversible computing offers is many orders of magnitude of information technology efficiency improvements, with associated economic advancements, compared to all possible irreversible computing technologies. With the end of conventional technology now in sight, it’s now time that the world’s best physics and engineering minds turn committed attention towards reversible computing, and begin an all-out effort to tackle its remaining engineering challenges, so as to bring this idea to practical fruition. The first person to describe the energy-efficiency implications of the conventional irreversible computing paradigm was Rolf Landauer of IBM, who wrote a paper in 1961 called “Irreversibility and Heat Generation in the Computing Process.” This paper has generated controversy in some circles, but Landauer’s key insight in this paper really does just follow directly as an immediate logical consequence of our most thorough, battle-tested understanding of fundamental physics. All of our most fundamental laws of low-level physical dynamics are reversible, meaning that if you were to have complete knowledge of the state of any given closed system at some time, and of the values of all of the relevant physical constants, M.P. Frank Back to the Future: Extended preprint Sandia National Labs The Case for Reversible Computing for IEEE Spectrum Version 5.7 arXiv:1803.02789 [cs.ET] Page 3 of 19 3/7/2018 7:48 PM you could always, conceptually, run the laws of physics backwards, and determine the system’s past state at any previous time exactly. (This is even true in quantum mechanics, if you knew the exact quantum state of the system.) As a consequence, it is impossible to have a situation wherein two different possible detailed states at some earlier time, could both evolve to become the exact same detailed state as each other at some later time, since this would mean that the earlier state couldn’t be uniquely determined from the later one. In other words, at the lowest level in physics, information cannot be destroyed. It’s important to realize how absolutely essential to our most basic understanding of physics this principle is. If it wasn’t true, then the Second Law of Thermodynamics (which says that entropy cannot decrease) could not be true, since entropy is just unknown information. If physics was not reversible, then entropy could simply vanish, and the Second Law would not hold. How does the indestructibility of information relate to the energy efficiency of irreversible computing? The point is that, since physics is reversible, whenever we think that we are destroying some information in a computer, we actually are not. Putatively “irreversible” operations (such as erasing a bit of information, or destructively overwriting it with a newly-computed value) are, in some sense, really just a convenient fiction. What’s actually happening, at the most fundamental level, is that the physical information that is embodied within the systems whose state we think we are “erasing” or “overwriting” (e.g., a circuit node charged to a particular voltage) is simply getting pushed out into the machine’s thermal environment, where it effectively becomes entropy (in essence, randomized information), and is manifested as heat. To increase the entropy of a thermal environment at temperature T by an increment ∆S requires adding an increment of heat ∆Q = T∆S to that environment; that is simply the thermodynamic definition of temper