CEA-LETI, 17 rue des Martyrs, 38054 Grenoble, France
Digital information processing has become a key driver for the growth of the world economy. It has been fueled by the progress of microelectronics that is usually described by an exponential growth of some performance metric, often named Moore’s Law. Actually, the early papers of Gordon Moore1-3 only stressed the continuing search for a higher integration density of circuits, mostly through feature size reduction, while their electrical behavior was not even mentioned. It was not until the mid-1970’s that Robert Dennard formalized the benefits of downscaling device dimension:4 the present paradigm that miniaturization makes integrated circuits denser, faster, less power-hungry, cheaper and more reliable was born. The question then is how long will this trend last?
For more than two decades, technical papers announced the imminent dismissal of Si CMOS technology, stressing first the 1 μm barrier, then the 100 nm brick walls, and recently the 10 nm limit: to date they all proved to be wrong. After looking at some similar claims, we will focus on the so-called physical limits of the processing unit, trying to outline the underlying assumptions of such assertions and their possible shortcomings.
Assuming that classical Si CMOS will encounter some practical limits in the future, the latest version of the ITRS roadmap gives a thorough analysis of possible candidates for the “beyond CMOS” era. We will review the necessary criteria for a successful replacement of the Si CMOS gate for information processing and make a critical assessment of some of the proposed approaches.
However the real question is: are we looking at the right problem? The focus on logic gates may just be the tree hiding the forest of issues to be addressed in information processing. We should also consider major pending problems, like interconnecting those gates or manufacturing complex circuits useful for specific applications.
As a word of caution, this paper will restrict its analysis to digital processing units, leaving other semiconductor technologies, such as memories and analog devices, outside of the scope of this discussion.
From the early days of microelectronics, an abundant literature explored the potential limits of the silicon technology: this chapter will mention a few milestones in this quest.
In the 1960’s and 197O’s, the focus was on the integration density – as exemplified by Moore’s paper1 – and the associated metric was the minimum achievable feature size. Among others, J. T. Wallmark published a detailed analysis of potential limits in shrinking dimensions5-7 and concluded that “the end of the road to smaller size has already been reached”8 – the smaller size being typically 5–10 microns in his early work. Pinpointing the flaws in his reasoning is not straightforward, but one can make a few remarks that may also apply to more recent papers on Si limits:
In the 1980’s and 1990’s, the focus shifted towards speed and power consumption. Following these metrics, J. D. Meindl established a hierarchy of constraints in designing an integrated circuit:9
However, considering the later levels of this hierarchy of constraints, as stressed by R. W. Keyes, “the potential tradeoffs [and the underlying assumptions] are too numerous and complex […] and they obscure the quantitative significance of [deriving] performance limits from purely physical reasoning”.10
The most recent works concentrate on the so-called “fundamental physical limits”. The general analysis is along the following lines11 (see Fig.1):
However, each step of this logical chain can and perhaps should be challenged. First of all, implicit assumptions related to the thermodynamic limit are numerous and potentially questionable, casting some doubt about the unavoidability of this barrier. Some of these questionable assumptions are made explicit below:
Further, it should be noted that the first Heisenberg uncertainty relationship,
is used with the underlying assumption that the information carrier is localized. This imposes a much stronger constraint on the physical implementation of the computing mechanism than just assuming that the system should evolve among distinguishable (i.e. mutually orthogonal) quantum states. Furthermore many device models (see, for example, Ref. 11) assume a free quasi-particle as an information carrier in a quasi-static system, using a semi-classical approximation and without quantum confinement of the carrier: how far the resulting conclusions apply to a more realistic physical device needs to be explored. Finally, in these models Δx is loosely defined as the physical width of the energy barrier in the device and at the same time as the linear size of the transistor. A more rigorous discussion of the applicability of this limit in the most general instance of a physical computing device is still lacking.
The relationship between energy and time in a quantum system, using for example the so-called second Heisenberg uncertainty relationship,
is still more doubtful. Actually, two models may apply to the discussion:
Further, it should be noted that the first Heisenberg uncertainty relationship,
is used with the underlying assumption that the information carrier is localized. This imposes a much stronger constraint on the physical implementation of the computing mechanism than just assuming that the system should evolve among distinguishable (i.e. mutually orthogonal) quantum states. Furthermore many device models (see, for example, Ref. 11) assume a free quasi-particle as an information carrier in a quasi-static system, using a semi-classical approximation and without quantum confinement of the carrier: how far the resulting conclusions apply to a more realistic physical device needs to be explored. Finally, in these models Δx is loosely defined as the physical width of the energy barrier in the device and at the same time as the linear size of the transistor. A more rigorous discussion of the applicability of this limit in the most general instance of a physical computing device is still lacking.
The relationship between energy and time in a quantum system, using for example the so-called second Heisenberg uncertainty relationship,
is still more doubtful. Actually, two models may apply to the discussion:
Beside these explicit criteria the device is expected to have a number of implicit qualities. For example, the device should:
Unfortunately, according to the ITRS, no emerging logic devices21 passed the exam so far. In fact, many new devices have no chance of matching, even in the distant future, the integration density and/or speed of the CMOS technologies presently in production! On the other hand, one cannot exclude the discovery of a new structure or a revolutionary concept in the future.
One common pitfall of many disruptive device proposals is that, starting from an interesting physical effect for a unit information processing mechanism, it fails to show the capability to integrate complex functions in a way competitive with the present or future CMOS systems. Quantum-dot cellular automata (QCA, for short) provide an instructive example.
The concept of QCA was initially proposed by Tougaw, Lent and Porod22 in 1993. The basic idea relies on a bistable cell (e.g. quantum dots positioned at the 4 corners of a square and close enough to allow excess electrons to tunnel between neighboring dots) that appears to promise an extremely fast and low-power device. The equilibrium state (or polarization) of this cell depends on the states of the nearest-neighbor cells that are arranged in a way to map logic functions (e.g. majority gate) and allow signal propagation along a line. The computation is performed by forcing the states of the input cells at the periphery of the device and by letting the system relax to the ground state. The computation result is read out looking at the final state of the output cells at the periphery of the device. However, it became rapidly clear that this relaxation mechanism led to unpredictable computation time and possibly wrong results if the system got stuck into metastable states. This led to the need to implement a complex multiphase “adiabatic” (ie. slow) clocking scheme.23 A comparison of QCA with CMOS for logic blocks of small complexity concluded24 in 2001 that an ultimately scaled molecular QCA would barely compete with advanced CMOS in terms of operational frequency. Analyzing the achievable integration density would lead to the same conclusion. The application of QCA to complex circuits would also need more in-depth analysis of other potential issues:
In short, after more than 12 years of research, QCA failed to provide a significant competitive advantage with respect to CMOS in terms of integration density, speed, power consumption, and implementation of complex systems. This interim conclusion has led to QCA using excess electrons in quantum dots and electrostatic interaction (so-called e-QCA) to be dropped from the list of potential emerging devices in the latest version of the ITRS. Surprisingly enough, magnetic and molecular variants of QCA did not share that fate despite facing the same basic issues.
Another, more subtle pitfall of many emerging devices and architectures lies in the difficulty of integrating the constraints of the whole chain from system architecture to complex circuits design, to manufacturing, test and assembly. In the recent years, there were many a ticles exploring the FPGA concept applied to nanoscale devices characterized by a very high fault rate. The basic idea25 resides in fabricating regular arrays of extremely miniaturized active devices, then in mapping the defects of the resulting structure, and finally in compiling and implementing the intended algorithm into the working elements. This potentially very attractive approach raises many questions, however.
In conclusion, up to now all the emerging logic devices failed to show the capacity to be integrated into complex processing units. Part of the present situation may be attributed to the fact that the researchers who propose disruptive approaches are often marginally aware of the complexity of designing and manufacturing complex functions into an integrated system. Conversely, CMOS circuit developers are usually unable to explicitly enumerate all the criteria for a “good” device implemented into a “good” circuit architecture: convergence of both worlds is a challenge in itself.
The main focus of the major microelectronics conferences is the elementary processing unit (i.e. transistor or switch), the memories and their association into complex circuits.
More specifically, in the quest for higher performances, many publications look at different ways to increase the on-current ION of the transistor (or reduce the associated metrics CV/ION). Using the approximate formula,
where μ is the carrier mobility of the carrier, CG the gate capacitance and LG the physical gate length, one may infer three main directions in the progress of MOS transistors, enumerated below.
Scaling of the critical LG dimension of the transistor
This is the conventional way initiated by Dennard and others. However, in moving to smaller LG and in packing the different elements more closely together, unwanted interactions and parasitic effects become increasingly detrimental to the behavior of the classical planar MOS transistor. Alternative structures, like SOI, 3D fin-FETs, multiple-gate or multiple-channel devices, have been proposed for many years and in some limited cases are already implemented in production – e.g. SOI in some microprocessors. The rather slow technological insertion of these alternative architectures is not only related to the added processing complexity and its potential impact on yield. One should not forget that most of these “unconventional” MOS devices require a rethinking of the whole design flow of complex circuits, which is a major effort most companies cannot afford (even assuming the new technology significantly outperforms the traditional approach).
Increasing the capacitive coupling CG of the gate and the conduction channel
The better electrostatic control of the conducting channel by the gate electrode was conventionally achieved by scaling the gate oxide thickness. Unfortunately, we reached dimensions where direct tunneling through the gate dielectrics leads to increasingly higher gate currents. To move away from this dead end, dielectrics with a high permittivity (the so-called high-κ materials) were proposed. However one cannot expect that the quasi-perfect Si-SiO2 system will be easily replaced: mobility degradation, Fermi level pinning and instabilities, among other drawbacks, are the present keywords of those high-κ dielectrics. As a result, these new materials are not expected to be introduced in the next CMOS generation.
As increasing transistor speed via traditional scaling becomes more and more difficult, enhanced transport of the carriers has been explored.
In the short term, the use of strain applied to the piezoresistive silicon and/or of other crystal orientations is a hot topic at device conferences. The combination of different techniques (see Fig. 2) allows balancing the benefits and drawbacks of each specific approach without adding too much complexity to the manufacturing process.
In the longer term, some researchers have proposed new channel materials to replace Si, such as Ge, III-V compound semiconductors, or carbon nanotubes. As in the case of high-κ dielectrics, one may expect significant processing difficulties and a diminishing return in the introduction of these materials into very complex integrated circuits.
The excitement about these “enhanced” MOS devices diverts attention from other equally important issues. In fact, one observes a widening gap between the research community and the real concerns of the microelectronics industry. The development engineer does not ask for breakthroughs in transistors and has more mundane questions, like:
In fact, the present bulk planar MOS transistor provides legacy solutions even for the 45 nm technologies. As long as new transistor architectures do not answer all these questions related to their integration into complex circuits, they will remain “exotic” solutions for the future.
Interconnecting the processing units and memories in an efficient way is probably a more critical issue than building faster transistors: at this point, the wiring technology lags transistor performance. Optical interconnections, carbon nanotubes or even ultra-low-κ dielectrics are not expected to bring orders of magnitude improvement in performance requested by the transistor roadmap. Instead, further progress will likely require a wise combination of design techniques and advanced manufacturing.
The microelectronics is clearly at a turning point in its history.
While the future of microelectronics is fuzzier than ever, it is too early to conclude that we are (almost) at the end of the road. Too many bright scientists predicted insurmountable limits and declared some physical implementations impossible, from airplanes to nuclear power, for us to be sure of the progress in Si logic is over. I would suggest applying a creative skepticism to any claim related to approaching limits, by clarifying all the underlying assumptions and trying to surmount or sidestep the obstacles.
Looking back to the past decades there is no reason to be pessimistic and it is safe to say that one cannot imagine what technology will be possible a few decades from now. On the other hand, it is also wise not to be too optimistic: the recent hype about new switches or new computational approaches may prove to be just hype. One should never underestimate the real complexity of building a working information processing technology and of replacing the accumulated knowledge and know-how of the present microelectronic industry by something totally new.
I would like to specifically thank Daniel Bois, Sorin Cristoloveanu, Hervé Fanet, Paolo Gargini, James Hutchby, Hiroshi Iwai, Thomas Skotnicki, Paul Solomon and Claude Weisbuch, who raised my interest in exploring the limits of Si CMOS and helped me (some of them unconsciously) to put this work together.