© 1999 Robert A. Freitas Jr. All Rights Reserved.

Robert A. Freitas Jr., Nanomedicine, Volume I: Basic Capabilities, Landes Bioscience, Georgetown, TX, 1999

10.2.1 Nanomechanical Computers

Electronic computers were evolutionarily preceded by purely mechanical computational devices, starting with the venerable abacus (hand-operated movable beads on rods) in ~3000 BC,^1726,1727 the Pascaline (the first hands-on algorithm-executing machine) in 1642,¹⁷²⁸ and the Difference Engine (the first hands-off algorithm-executing machine) designed in 1821 by Charles Babbage.^{1728-1731,1744} A 2000-part working subsection of the brass-geared Engine was demonstrated in 1832; an entire working Difference Engine was reconstructed by historians in 1991, proving that Babbage's design was sound.¹⁷³² In the 1840s, Thomas Fowler built and exhibited a calculating device using sliding rods made of wood instead of metal.¹⁷³³ Whereas Babbage's engines used the familiar 0-9 decimal system with each number represented by a discrete position of a rotating gear wheel, Fowler's machine was more fully digital, using as its active element not rotating wheels but sliding "trinary" rods which could occupy only one of three positions at any time, the first known example of "rod logic." (By reducing the number of distinct physical states, parts could be made less precisely.)

By 1834, Babbage had also conceived detailed plans for his Analytical Engine, intended as a general-purpose programmable computing machine but based entirely on 19^th century mechanical technology. The Analytical Engine was to have a random-access memory consisting of 1000 words of 50 decimal digits each (~175,000 bits), with separate memory and central processing unit (CPU), stored program control, data entry via punched metal cards, and even an output printer.^{1728,1730,1732} This ambitious device, though well specified, was never built.

The mechanical computing tradition was not entirely abandoned. Vannevar Bush built his analog mechanical computer, the Differential Analyzer, in 1930 at MIT.¹⁷³⁶ In 1954, M. Minsky and R. Silver²⁸⁹ used hydraulic logic elements to build a mechanical "hydroflip computer" which was operated at ~30 Hz and powered by a 3-inch high column of water. (Section 9.2.7.6 describes the basis for a mechanical fluidic computer operating at ~5000 Hz.) In 1975, D. Hillis and B. Silverman¹⁷³⁸ built a special-purpose all-mechanical computer ~2 meters in size, entirely out of Tinkertoys, and powered by a hand crank, that was able to play tic-tac-toe. In 1990, University of Minnesota engineers¹⁷³⁴ fabricated a complete family of micromechanical digital logic devices, including electrostatically-actuated linear-sliding ~30 micron mechanical logic elements confined to a one-dimensional track, forming NAND and NOR gates suitable for low-speed radiation-hard digital functions "in environments hostile to electronic devices." Sandia's pin-in-maze microlocks²³⁵⁶ could also be used to make a mechanical computer, albeit at the above-micron scale. In 1996, J. Gimzewski and colleagues¹⁷³⁵ at IBM Zurich used a scanning tunneling microscope (STM) probe to repeatedly reposition spherical C₆₀ fullerene molecules ~1 nm in diameter along a terraced copper substrate that constrained the buckyballs to move only in a straight line, operating the mechanical array like beads on an abacus. In 1997, Stoddart's group²⁵⁴⁰ described a mechanical XOR gate based on their "molecular shuttles."

Perhaps the best-characterized (though not yet built) mechanical nanocomputer is Drexler's rod logic design.^10,2282 In this design, one sliding rod with a knob intersects a second knobbed sliding rod at right angles to the first. Depending upon the position of the first rod, the second may be free to move, or unable to move. This simple blocking interaction serves as the basis for logical operations. Figure 10.1 shows a nanomechanical implementation of a Boolean NAND "interlock" gate, using clock-driven input and output logic rods 1 nm wide which interact via knobs that prevent or enable motion, all encased in a housing, allowing ~16 nm³/interlock. (Any logic function, no matter how complicated, can be built from NAND or NOR gates alone.¹⁷³⁶) Figure 10.2 is an exploded diagram of the moving parts of a thermodynamically efficient class of register capable of mechanical data storage, using rods ~1 nm in width with 0.1-nanosec switching speeds, allowing ~40 nm³/register. The activity sequence depicts a simplified version of Drexler's register, omitting the reading mechanism.¹⁷⁴³ Initially the register shows a 0 (black ball position in A) or a 1 (B). Then the barrier is lowered and the ball wanders freely (C); entropy increases. The register is then reset to 0 by spring-rod compression, converting ~kT ln(2) joules of work into heat (D). To write a 0, the barrier is raised (E). To write a 1, the input rod at right is first extended and then the barrier is raised (F); the input rod does work compressing the ball into the spring, but this energy can be retrieved when the spring rod is retracted. Finally, the spring rod is retracted, returning the device to state (A) or (B).

Figure 10.3 illustrates a programmable logic array (PLA) using knobbed rods (omitting several drive and spring systems for clarity) to implement nanomechanical logic, which, in combination with rod-based registers, can be used to build much of the control circuitry for a CPU in a >~1 GHz clocked computer system. For details of operation, see Drexler.¹⁰ While a PLA system requires three successive rod displacement cycles to compute a set of Boolean functions, the functions AND, OR, and NOT can also be computed in a single displacement cycle by employing a linkage in which any input rod can displace the output rod without displacing any other input rod; Figure 10.4 illustrates this alternative approach for an asynchronous-input OR gate.

Drexler's benchmark mechanical nanocomputer design has 10⁶ interlock gates, 10⁵ logic rods, 10⁴ registers, an energy-buffering flywheel and other components with total cubic volume (~400 nm)³, mass ~10^-16 kg, and total power ~ 60 nW, giving a power density of ~10¹² watts/m³.¹⁰ Power dissipation per logic operation is ~0.013 zJ per gate per cycle, giving (including register dissipation) ~2 x 10⁴ operations/sec-pW.¹⁰ Processing speed is ~10⁹ operations/sec (~1 gigaflop) or ~10²⁸ operations/sec-m³; assuming one bit per register, processing speed is ~10¹³ bits/sec or ~10³² bits/sec-m³. In 1998, the typical desktop PC operated at ~10⁸ operations/sec. Cooling is provided by a refrigerant fluid flowing through an integral fractal plumbing system at near the speed of sound.¹⁰

The most primitive 4-bit Intel 4004 microprocessor, introduced in 1971 for early-generation pocket calculators, had 2300 transistors. A comparably simple early mechanical nanocomputer with 2000 interlock gates, 100 registers, and a 1 KHz clock speed could have a volume as small as 36,000 nm³, a (~33 nm)³ cube, assuming 16 nm³/gate and 40 nm³/register, and could process ~10⁵ bits/sec.

One major disadvantage of mechanical nanocomputers is that they will almost certainly be slower than nanoelectronic systems, because a device that depends on moving heavy nuclei (~10^-27 kg) will necessarily be slower than a device that depends on moving less-massive electrons (~10^-30 kg).²⁸⁰ Mechanical signals travel near the speed of sound (~10⁴ m/sec in diamond), whereas electronic signals may travel near the speed of light (~10⁸ m/sec). Thus while nanomechanical computers may be limited to switching speeds of ~50 picoseconds, electronic devices may be ~10³-10⁴ times faster. On the other hand, mechanical computers are more EMP (electromagnetic pulse) resistant, conceptually easier to understand and to model, and mechanical designs scale readily from the macroscale to the nanoscale, unlike most nanoelectronic designs. In 1996, a ~50 nm reversible mechanical latch was fabricated and physically cycled using an AFM; the similarity to Drexler's mechanical OR gate (Figure 10.4) was explicitly noted in the paper.¹⁷³⁷

What about mechanical data storage? In the MEMS world, Halg¹⁷⁴² fabricated a nonvolatile microelectromechanical memory cell consisting of a longitudinally-stressed surface-micromachined ~10 micron bridge that mechanically buckles into one of two states, making a bistable storage device. Moving to the nanoscale, data could be mechanically stored using compact three-dimensional arrays of diamondoid register rods (~10⁷ bits/micron³, ~10¹⁰ bits/sec access speed; Figure 10.2, Section 7.2.6, and Drexler¹⁰). Another theoretical mechanical storage medium is spooled hydrofluorocarbon memory tape (~10¹⁰ bits/micron³, ~10⁹ bits/sec access speed; Sections 7.2.1.1 and 7.2.6 and Drexler¹⁰), or polymer or diamond surface-bound patterns of H and F atoms that could be read using an optimized (CH₃)₃PO scanning probe tip with a maximum raw error rate of 6 x 10^-8 per read.¹²⁰⁰ (Linear DNA molecules achieve ~10⁹ bits/micron³ data storage, by comparison.) Direct AFM reading and recording of 10-nm pits on a polycarbonate surface (~10⁶ bits/micron², up to ~10⁷ bits/sec access speed)¹⁷³⁹ and other means of directly reading and writing data at the atomic level^1749,2182 have been studied, and the fullerene abacus¹⁷³⁵ described earlier is one of the first experimental efforts to mechanically store numerical information using individual molecules at room temperature.

It has been speculated that individual atoms or atomic vacancies could serve as information units.^{278,1200,1739,1745,2711} For example, Y. Mo¹⁷⁴¹ used an STM to store information in the reversible rotational states of individual antimony dimers deposited on a silicon substrate. In theory, a block of diamond with data encoded as single-atom lattice vacancies could store up to ~176 bits/nm³, or nearly ~2 x 10¹¹ bits/micron³, although cryogenic temperatures might be required to reduce diffusion effects and structural lability to acceptable levels consistent with long-term information storage. Phonon probes or other nondestructive means of lattice interrogation could be employed for readout.

Last updated on 23 February 2003