The Role of

Hans Moravec May 12, 1976 CONTENTS 1 Introduction 2 Acknowledgement 3 Section 1: The Natural History of Intelligence 3 Product lines 5 Unifying principles 6 Harangue 7 References 10 Section 2: Measuring Processing Power 10 Low level vision 11 Entropy measurement 13 A representative computer 14 A typical nervous system 15 Thermodynamic efficiency 17 References 18 Section 3: The Growth of Processing Power 20 References 21 Section 4: Interconnecting Processors 21 Log^2 sorting net construction 23 Communication scheme organization 27 Package counts 29 Speed calculations 30 Possible refinements 31 References 32 Section 5: Programming Interconnected Processors 33 A little Lisp 35 A little Algol 37 A little operating systems 38 Disclaimer 39 Section 6: Bombast


This essay is an argument that an essential ingredient is absent in many current conceptions of the road to Artificial Intelligence. The first section discusses natural intelligence, and notes two major branches of the animal kingdom in which it evolved independently, and several offshoots. The suggestion is that intelligence need not be so difficult to construct as is sometimes assumed. The second part compares the information processing ability of present computers with intelligent nervous systems, and finds a factor of one million difference. This abyss is interpreted as a major distorting influence in current work, and a reason for disappointing progress. Section three examines the development of electronics, and concludes the state of the art can provide more power than is now available, and that the gap could be closed in a decade. Parts four and five introduce hardware and software aspects of a system which is able to make use of the advancing technology.


The following entities provided inspiration, encouragement, suggestions, data, slave labor, proof reading services, etc.: (in carefully randomized order) PDP-KA10, Scientific American, Marc Le Brun, Andy Moorer, Ed Mcguire, Electronics magazine, Don Gennery, Bill Gosper, John McCarthy, Macsyma, Mike Farmwald, Russ Taylor, Cart Project, Les Earnest, Pierre van Nypelseer, Robert Maas, Jeff Rubin, Bruce Baumgart, HAL-9000, Tom Binford, Clem Smith, Tom Gafford, Brian Harvey, ...

Section 1: The Natural History of Intelligence

Product lines: Natural evolution has produced a continuum of complexities of behavior, from the mechanical simplicity of viruses to the magic of mammals. In the higher animals most of the complexity resides in the nervous system. Evolution of the brain began in early multi-celled animals a billion years ago with the development of cells capable of transmitting electrochemical signals. Because neurons are more localized than hormones they allow a greater variety of signals in a given volume. They also provide evolution with a more uniform medium for experiments in complexity. The advantages of implementing behavioral complexity in neural nets seem to have been overwhelming, since all modern animals more than a few cells in size have them [animal]. Two major branches in the animal kingdom, vertebrates and mollusks, contain species which can be considered intelligent. Both stem from one of the earliest multi-celled organisms, an animal something like a hydra made of a double layer of cells and possessing a primitive nerve net. Most mollusks are intellectually unimpressive sessile shellfish, but one branch, the cephalopods, possesses high mobility, large brains and imaging eyes, evolved independently of the corresponding vertebrate structures. There are fascinating differences. The optic nerve connects to the back of the retina, so there is no blind spot. The brain is annular, forming a ring encircling the esophagus. The circulatory system, also independently evolved, has three blood pumps, a systemic heart pumping oxygenated blood to the tissues and two gill hearts, each pumping venous blood to one gill. The oxygen carrier is a green copper compound called hemocyanin, evolved from an earlier protein that also became hemoglobin. These animals have some unique abilities. Shallow water octopus and squid are covered by a million individually controlled color changing effectors called chromatophores, whose functions are camouflage and communication. The capabilities of this arrangement have been demonstrated by a cuttlefish accurately imitating a checkerboard it was placed upon, and an octopus in flight which produced a pattern like the seaweed it was traversing, coruscating backward along the length of its body, diverting the eye from the true motion. Deep sea squid have photophores capable of generating large quantities of multicolored light. Some are as complex as eyes, containing irises and lenses [squid]. The light show is modulated by emotions in major and subtle ways. There has been little study of these matters, but this must provide means of social interaction. Since they also have good vision, there is the potential for high bandwidth communication. Cephalopod intelligence has not been extensively investigated, but a few controlled experiments indicate rapid learning in small octopus [Boycott]. The Cousteau film shows an octopus' response to a "monkey and bananas" problem. A fishbowl containing a lobster is sealed with a cork and dropped into the water near it. The octopus is attracted, and spends a long while alternately probing the container in various ways and returning to its lair in iridescent frustration. On the final iteration it exits its little hole in the ground and unhesitatingly wraps three tentacles around the bowl, and one about the cork, and pulls. The cork shoots to the surface and the octopus eats. The Time-Life film contains a similar sequence, with a screw top instead of a cork! If small octopus have almost mammalian behavior, what might giant deep sea squid be capable of? Birds are more closely related to humans than are cephalopods, their common ancestor with us being a 300 million year old early reptile. Size-limited by the dynamics of flying, some birds have reached an intellectual level comparable to the highest mammals. Crows and ravens are notable for frequently outwitting man. Their intuitive number sense (ability to perceive the cardinality of a set without counting) extends to seven, as opposed to three or four in man. Such a sense is useful for keeping track of the number of eggs in a nest. Experiments have shown [Stettner] that most birds are more capable of high order "reversal" and "learning set" learning than all mammals except the higher primates. In mammals these abilities increase with increasing cerebral cortex size. In birds the same functions depend on areas not present in mammalian brains, forebrain regions called the "Wulst" and the hyperstriatum. The cortex is small and relatively unimportant. Clearly this is another case of independent evolution of similar mental functions. It is interesting to speculate whether penguins, now similar to seals in behavior and habitat, will ever become fully aquatic, and evolve analogously to the great whales. The cetaceans are related to man through a small 30 million year old primitive mammal. Some species of dolphin have body and brain masses identical to ours, and archaeology reveals they have been this way several times as long. They are as good as man at many kinds of problem solving, and perhaps at language. The references contain many anecdotes, and describe a few controlled experiments, showing that dolphins can grasp and communicate complex ideas. Killer whales have brains seven times human size, and their ability to formulate plans is better than dolphins', occasionally being used to feed on them. Sperm whales, not the largest animals, have the world's largest brains. There may be intelligence mediated conflict with large squid, their main food supply. Elephants have brains about six times human size, matriarchal tribal societies, and complex behavior. Indian domestic elephants learn 500 commands, limited by the range of tractor-like tasks their owners need done, and form voluntary mutual benefit relationships with their trainers, exchanging labor for baths. They can solve problems such as how to sneak into a plantation at night to steal bananas, after having been belled (answer: stuff mud into the bells). And they remember for decades. Inconvenience and cost has prevented modern science from doing much work with them. The apes are man's cousins. Chimps and gorillas can learn to use tools and to communicate with human sign languages at a retarded level. As chimps have one third, and gorillas one half, human brainsize, similar results should be achievable with the larger brained, but less man-like animals. Though no other species has managed to develop cultures comparable to modern man's, it may be that some of them can be made partners in ours, accelerating its evolution by their unique capabilities. Unifying principles: A feature shared by all living organisms whose behavior is complex enough to place them near human in intelligence is a hundred billion neuron nervous system. Imaging vision requires a billion neurons, while a million is associated with fast and interesting, but stereotyped, behavior as in a bee. A thousand is adequate for slow moving animals with minimal sensory input, such as slugs and worms. A hundred runs most sessile animals. The portions of nervous systems for which tentative wiring diagrams have been obtained, including much of the brain of the large neuroned sea slug, Aplysia, the flight controller of the locust and the early stages of some vertebrate visual systems, reveal that the neurons are configured into efficient, clever, assemblies. This should not be too surprising, as unnecessary redundancy means unnecessary metabolic load, a distinct selective disadvantage. Evolution has stumbled on many ways of speeding up its own progress, since species that adapt more quickly have a selective advantage. Most of these, such as sex and individual death, have been refinements of one of the oldest, the encoding of genetic information in the easily mutated and modular DNA molecule. In the last few million years the genetically evolved ability of animals, especially mammals, to learn a significant fraction of their behavior after birth has provided a new medium for growth of complexity. Modern man, though probably not the most individually intelligent animal on the planet, is the species in which this cultural evolution seems to have had the greatest effect, making human culture the most potent force on the earth's surface. Cultural and technological evolution proceeds by massive interchange of ideas and information, trial and error guided by the ability to predict the outcome of simple situations, and other techniques mediated by the intelligence of the participants. The process is self reinforcing because its consequences, such as improved communication methods and increased wealth and population, allow more experiments and faster cross fertilization among different lines of inquiry. Many of its techniques have not been available to biological evolution. The effect is that present day global civilization is developing capabilities orders of magnitude faster. Of course biological evolution has had a massive head start. Although cultural evolution has developed methods beyond those of its genetic counterpart, the overall process is essentially the same. It involves trying large numbers of possibilities, selecting the best ones, and combining successes from different lines of investigation. This requires time and other finite resources. Finding the optimum assembly of particular type of component which achieves a desired function usually requires examination of a number of possibilities exponential in the number of components in the solution. With fixed resources this implies a design time rising exponentially with complexity. Alternatively the resources can be used in stages, to design subassemblies, which are then combined into larger units, and so on, until the desired result is achieved. This can be much faster since the effort rises exponentially with the incremental size of each stage and linearly with the number of stages, with an additional small term, for overall planning, exponential in the number of stages. The resulting construct will probably use more of the basic component and be less efficient than an optimal design. Biological evolution is affected by these considerations as much as our technology. Presumably there is a way of using the physics of the universe to construct entities functionally equivalent to human beings, but vastly smaller and more efficient. Terrestrial evolution has not had the time or space to develop such things. But by building in the sequence atoms, amino acids, proteins, cells, organs, animal (often concurrently), it produced a technological civilization out of inanimate matter in only two billion years. Harangue: The existence of several examples of intelligence designed under these constraints should give us great confidence that we can achieve the same in short order. The situation is analogous to the history of heavier than air flight, where birds, bats and insects clearly demonstrated the possibility before our culture mastered it. Flight without adequate power to weight ratio is heartbreakingly difficult (vis. Langley's steam powered aircraft or current attempts at man powered flight), whereas with enough power (on good authority!) a shingle will fly. Refinement of the aerodynamics of lift and turbulence is most effectively tackled after some experience with suboptimal aeroplanes. After the initial successes our culture was able to far surpass biological flight in a few decades. Although there are known brute force solutions to most AI problems, current machinery makes their implementation impractical. Instead we are forced to expend our human resources trying to find computationally less intensive answers, even where there is no evidence that they exist. This is honorable scientific endeavor, but, like trying to design optimal airplanes from first principles, a slow way to get the job done. With more processing power, competing presently impractical schemes could be compared by experiment, with the outcomes often suggesting incremental or revolutionary improvements. Computationally expensive highly optimizing compilers would permit efficient code generation at less human cost. The expanded abilities of existing systems such as MACSYMA, along with new experimental results, would accelerate theoretical development. Gains made this way would improve the very systems being used, causing more speedup. The intermediate results would be inefficient kludges busily contributing to their own improvement. The end result is systems as efficient and clever as any designed by more theoretical approaches, but sooner, because more of the labor has been done by machines. With enough power anything will fly. The next section tries to decide how much is needed.
References: [animal] JERISON, Harry J., "Paleoneurology and the Evolution of Mind", Scientific American, Vol. 234, No. 1, January 1976, 90-101. BITTERMAN, M. E., "The Evolution of Intelligence", Scientific American, Vol. 212, No. 1, January 1965, 92-100. GRIFFIN, Donald R., ed. "Animal Engineering", W.H. Freeman and Company, San Francisco, June 1974. BURIAN, Z. and Spinar, Z.V., "Life Before Man", American Heritage Press, 1972. RIOPELLE, A.J., ed. "Animal Problem Solving", Penguin Books, 1967. GOODRICH, Edwin S., "Studies on the Structure and Development of Vertebrates", Dover Publications Inc., New York, 1958. BUCHSBAUM, Ralph, "Animals without Backbones", The University of Chicago Press, 1948. FARAGO, Peter, and Lagnado, John, "Life in Action" Alfred A. Knopf, New York, 1972. BONNER, John Tyler, "Cells and Societies", Princeton University Press, Princeton, 1955. [squid] COUSTEAU, Jacques-Yves and Diole, Philippe, "Octopus and Squid", Doubleday & Company, Inc., Garden City, New York, 1973. (also a televised film of the same name) "The Octopus", a televised film, Time-Life films. BOYCOTT, Brian B., "Learning in the Octopus", Scientific American, Vol. 212, No. 3, March 1965, 42-50. LANE, Frank W., "The Kingdom of the Octopus", Worlds of Science Book, Pyramid Publications Inc. October, 1962. [bird] BAKKER, Robert T., "Dinosaur Renaissance", Scientific American, Vol. 232, No. 4, April 1975. STETTNER, Laurence Jay and Matyniak, Kenneth A. "The Brain of Birds", Scientific American, Vol. 218, No. 6, June 1968, 64-76. [whale] LILLY, John. C., "The Mind of the Dolphin" & "Man and Dolphin", Doubleday and Company, New York, 1967. COUSTEAU, Jacques-Yves and Diole, Philippe, "The Whale", Doubleday & Company, Inc., Garden City, New York, 1972. FICHTELIUS, Karl-Erik and Sjolander, Sverre, "Smarter than Man?", Ballantine Books, New York, 1974. STENUIT, Robert, "The Dolphin, Cousin to Man", Bantam Books, New York, 1972. "Whales and Dolphins", A BBC produced film shown in the NOVA television series McINTYRE, Joan, ed. "Mind in the Waters", Charles Scribner's Sons, San Francisco, 1974. [elephant] RENSCH, Bernhard, "The Intelligence of Elephants", Scientific American, February 1957, 44. [primate] "The First Words of Washoe", Televised film shown in the NOVA series LeGROS CLARK, W.E., "History of the Primates", The University of Chicago Press, Chicago 1966. PFEIFFER, John, "The Human Brain", Worlds of Science Books, Pyramid Publications Inc., New York, 1962.

Section 2: Measuring Processing Power

Low level vision: The visual system of a few animals has been studied in some detail, especially the layers of the optic nerve near the retina. The neurons comprising these structures are used efficiently to compute local operations like high pass filtering and edge, curvature, orientation and motion detection. Assuming the visual cortex (and possibly the optic nerve itself) is as computationally intensive as the retina, successive layers producing increasingly abstracted representations, we can estimate the total capability. There are a million seperate fibers in a cross section of the human optic nerve. At the retina each is connected to several of 10 million light sensitive rods and cones. The thickness of the optical cortex is a thousand times the depth occupied by the neurons which apply a single simple operation. The eye is capable of processing images at the rate of ten per second (flicker at higher frequencies is detected by special operators). This means that the human visual system evaluates 10,000 million pixel simple operators each second. A tightly hand coded simple operator, like high pass filtering by subtraction of a local average, applied to a 256x256 (1/16 million) pixel picture takes at least 10 seconds when executed on a PDP-10 (KA), not counting timesharing. A million pixel image would require 160 seconds, if storage constraints permitted it. Since the computer can evaluate only one at a time, the effective rate is 1/160 million pixel simple operators per second. Thus a hand coded PDP-10 falls short of being the equal of the human optical system by a speed factor of 1.6 million. It may not be necessary to apply every operator to every portion of every picture, and a general purpose computer, being more versatile than the optic nerve, can take advantage of this. I grant an order of magnitude for this effect, reducing the optic nerve to a mere 100,000 PDP-10 equivalents. The size of this factor is related to having chosen to implement our algorithms in machine language. If we had opted to disassemble a number of PDP-10's and reconfigure the components to do the computation, far fewer would have been required. On the other hand if we had run our algorithms in interpreted Lisp, 10 to 100 times as many would be needed. The tradeoff is that the design time varies inversely with the execution efficiency. A good Lisp program to compute a given function is easier to produce than an efficient machine language program, or an equivalent piece of hardware. Compilers and automatic design programs blur the issue a little by passing some of the effort of implementation to a machine, but the essential character remains. Entropy measurement: Is there a quantitative way in which the processing power of a system, independent of its detailed nature, can be measured? A feature of things which compute massively is that they change state in complicated and unexpected ways. The reason for believing that, say, a stationary rock does little computing is its high predictability. By this criterion the amount of computing done by a device is in the mind of the beholder. A machine with a digital display which flashed 1, 2, 3, 4 etc., at fixed intervals would seem highly predictable to an adult of our culture, but might be justifiably considered to be doing an interesting, nontrivial and informative computation by a young child. Information theory provides a measure for this idea. If a system is in a given state and can change to one of a number of next states, the information in the transition, which I will call the Compute Energy (CE), is given by __N_ \ > - p log(p ) / i i /---- i=1 where N is the number of next states, and p[i] is the probability that the i'th state will be occur next. If the base of the logarithm is two, the measure is in binary digits, bits. If we consider the system in the long run, considering all the states it might ever eventually be in, then this measure expresses the total potential variety of the system. A machine which can accomplish a given thing faster is more powerful than a slower one. A measure for Compute Power is obtained by dividing each term in the above sum by the expected time of the corresponding transition. This is __N_ \ - p log(p ) CP = > ___i______i_ / t /---- i i=1 and the units are bits/second. These measures are highly analogous to the energy and power capacities of a battery. Some properties follow: They are linear, i.e. the compute power and energy of a system of two or more independent machines is the sum of the individual power and energies; Speeding up a machine by a factor of n increases the CP by the same factor; A completely predictable system has a CP and CE of zero; A machine with a high short term CP, which can reach a moderate number of states in a short time, can yet have a low CE, if the total number of states attainable in the long run is not high; If the probabilities and times of all the transitions are equal the measures simplify to CE = log2(N) CP = log2(N)/t For N distinct outcomes and equal transition times the maximum CE and CP is obtained if all the probabilities of occurrence are 1/N, so the above simplifications represent an upper bound in cases where the probabilities are variable or cannot be determined. A representative computer: For the KA-PDP10, considering one instruction time, we have (roughly) that in one microsecond this machine is able to execute one of 2^5 different instructions, involving one of 2^4 accumulators and one of 2^18 memory locations, most of these combinations resulting in distinct next sates. This corresponds to a CP of 5 4 18 log(2 x 2 x 2 ) bits 6 --------------------- = 27 x 10 bits/sec -6 10 sec This is an extreme upper bound, and represents the most efficient code possible. It cannot be maintained for very long, because many different sequences of instructions have the same outcome. Very often, for instance, the order in which a number of instructions is executed does not matter. Assuming for the moment that this is always the case, we can calculate the effect on the CP, measured over a second. The raw power says there are 6 27 x 10 2 distinct possible states after a second, or one million instruction times. If all permutations result in the same outcome, this must be reduced by a factor of 1000000!. The log of a quotient is the difference of the logs, so the adjusted CP is 6 6 6 6 6 27 x 10 - log(10 !) = 27 x 10 - 18.5 x 10 = 8.5 x 10 bits/sec. The CP is also limited by the total compute energy. If we ignore external devices, this is simply the total amount of memory, about 36x2^18 = 9.4x10^6 bits. The PDP 10 could execute at its maximum effectiveness for 9.4/8.5 = 1.1 seconds before reaching a state which could have been arrived at more quickly another way. Large external storage devices such as disks and tapes can increase the compute energy indefinitely, because they are a channel through which independently obtained energy may be injected. On the other hand, they have only a moderate effect on the power. A disk channel connected to our KA-10 has a data rate of 6.4x10^6 bits/sec. If run at full speed, constantly stuffing new information into memory, it would add slightly less than that to the power, because it uses memory cycles the processor might want. If, as is usually the case, the disk is used for both reading and writing, the improvement is reduced by a factor of two. Further reductions occur if the use is intermittent. The overall impact is less than 10^6 bits/sec, about 10% of the power of the raw processor. Combining the above results, we conclude that the processing power of a typical major AI center computer is at most 10^7 bits/sec. Time sharing reduces this to about 10^6 b/s per user. Programming in a moderately efficient high level language (vs. optimal machine code), costs another factor of 10, and running under an interpreter may result in a per user power of a mere 10,000 bits/sec, if the source code is efficient. These reductions can be explained in the light of the CP measure by noting that the a compiler or interpreter causes the executed code to be far more predictable than optimal code, eg. by producing stereotyped sequences of instructions for primitive high level constructs. A typical nervous system: We now consider the processing ability of animal nervous systems, using humans as an example. Since the data is even more scanty than what we assumed about the PDP-10, some not unassailable assumptions need to be made. The first is that the processing power resides in the neurons and their interconnections, and not in more compact nucleic acid or other chemical encodings. There is no currently widely accepted evidence for the latter, while neural mechanisms for memory and learning are being slowly revealed. A second is that the neurons are used reasonably efficiently, as detailed analysis of small nervous systems and small parts of large ones reveals (and common sense applied to evolution suggests). Thirdly, that neurons are fairly simple, and their state can be represented by a binary variable, "firing" or "not firing", which can change about once per millisecond. Finally we assume that human nervous systems contain about 40 billion neurons. Considering the space of all possible interconnections of these 40 billion (treating this as the search space available to natural evolution in its unwitting attempt to produce intelligence, in the same sense that the space of all possible programs is available to someone trying to create intelligence in a computer), we note that there is no particular reason why every neuron should not be able to change state every millisecond. The number of combinations thus reachable from a given state is 2^(40x10^9) the binary log of which is 40x10^9. This leads to a compute power of 9 40 x 10 bits 12 --------------------- = 40 x 10 bits/sec -3 10 sec which is about a million times the maximum power of the KA-10. Keep in mind that much of this difference is due to the high level of interpretation in the 10, compared to what we assumed for the nervous system. Rewiring its gates or transistors for each new task would greatly increase the CP, but also the programming time. If the processor is made of 100,000 devices which can change state in 100 ns, the potential CP available through reconfiguration is 10^5 bits/10^-7 sec = 10^12 b/s. The CE would be unaffected. If automatic design and fabrication methods result in small quantity integrated circuit manufacture becoming less expensive and more widely practiced, my calculations may prove overly pessimistic. Thermodynamic efficiency: Thermodynamics and information theory provide us with a minimum amount of energy per bit of information generated at a given background temperature (the energy required to out shout the thermal noise). This is approximately the Boltzmann constant, -16 erg 1.38 x 10 ------------ deg molecule for our purposes the units will be rephrased as erg/(deg bit), since for normal matter molecules represent the limit of organization, due to the quantum mechanical fuzziness of things. This measure allows us to estimate the overall energy efficiency of computing engines. For instance, we determined the computing rate of the brain, which operates at 300 degrees K, to be 40x10^12 bits/sec. This corresponds to a physical power of 12 bit -16 erg erg 40 x 10 --- x 300 deg x 1.38 x 10 ------- = 1.656 --- sec deg bit sec -7 = 1.656 x 10 watt The brain runs on approximately 40 watts, so we conclude that it is 10^-8 times as efficient as the physical limits allow. Doing the same calculation for the KA10, again at 300 deg, we see that a CP of 8.5x10^6 bit/sec is worth 3.519x10^-14 watts. Since this machine needs 10 kilowatts the efficiency is only 10^-18. Conceivably a ten watt, but otherwise equivalent, KA10 could be designed today, if care were taken to use the best logic for the required speed in every assembly. The efficiency would then still be only 10^-15. As noted previously, there is a large cost inherent in the organization of a general purpose computer. We might investigate the computing efficiency of the logic gates of which it is constructed (as was, in fact, done with the brain measure). A standard TTL gate can change state in about 10ns, and consumes 10^-3 watt. The switching speed corresponds to a CP of 10^8 bit/sec, or a physical power of 4.14x10^-13 watt. So the efficiency is 10^-10, only one hundred times worse than a vertebrate neuron. The newer semiconductor logic families are even better. C-MOS is twice as efficient as TTL, and Integrated Injection Logic is 100 times better, putting it on a par with neurons. Experimental superconducting Josephson junction logic operates at 4 deg K, switches in 10^-11 sec, and uses 10^-7 watts per gate. This implies a physical compute power of 5x10^-12 watt, and an efficiency of 5x10^-5, 1000 times better than neurons. At room temperature it requires a refrigerator that consumes 100 times as much energy as the logic, to pump the waste heat uphill from 4 degrees to 300. Since the background temperature of the universe is about 4 degreees, this can probably eventually be done away with. It is thus likely that there exist ways of interconnecting gates made with known techniques which would result in behavior effectively equivalent to that of human nervous systems. Using a million I2L gates, or 10 thousand Josephson junction gates, and a trillion bits of slower bulk storage, all running at full speed, such assemblies would consume as little as, or less than, the power needed to operate a brain of the conventional type. Past performance indicates that the amount of human and electronic compute power available is inadequate to design such an assembly within the next few years. The problem is much reduced if the components used are suitable large subassemblies. Statements of good high level computer languages are the most effective such modularizations yet discovered, and are probably the quickest route to human equivalence, if the necessary raw processing power can be accessed through them. This section has indicated that a million times the power of typical existing machines is required. The next suggests this should be available at reasonable cost in about ten years. References: WILLOWS, A.O.D., "Giant Brain Cells in Mollusks", Scientific American, Vol. 224, No. 2, February 1971, 69-75. KANDEL, Eric R., "Nerve Cells and Behavior", Scientific American, Vol. 223, No. 1, July 1970, 57-70. AGRANOFF, Bernard W., "Memory and Protein Synthesis", Scientific American, Vol. 216, No. 6, June 1967, 115-122. KENNEDY, Donald, "Small Systems of Nerve Cells", Scientific American, Vol. 216, No. 5, May 1967, 44-52. BAKER, Peter F., "The Nerve Axon", Scientific American, Vol. 214, No. 3, March 1966, 74-82. HUBEL, David H., "The Visual Cortex of the Brain", Scientific American, November 1963, 54-62. WILLIAMS, Peter L., Warwick, Roger "Functional Neuroanatomy of Man", W.B. Saunders Company, Philadelphia, 1975. "PDP-10 Reference Handbook", Digital Equipment Corporation, Maynard Mass., 1971. TRIBUS, Myron and McIrvine, Edward C., "Energy and Information", Scientific American, Vol. 224, No. 3, Setember 1971, 179-188. GLASSTONE, Samuel, Lewis, David, "Elements of Physical Chemistry", D. Van Nostrand Co. Inc., New York, 1960. MILLER, Richard T., "Super Switch", 284, in Science Year annual 1975, Field Enterprises Educational Corp., 1975. LANDAUER, Rolf, "Fundamental Limitations in the Computational Process", IBM Research Report, Yorktown Heights N.Y., June 1976.

Section 3: The Growth of Processing Power

The references below present, among other things, the following data points on a price curve: Transistor price .0001c .01 1c $1 $100 +---+---+---+---+---+---+---+---+---+---+ Year | | +- O -+ 1950 | # | 1951 $100 transistor | # | 1952 transistor hearing aid | # | 1953 | # | 1954 +- # -+ 1955 transistor radios | # | 1956 | O | 1957 $10 transistor | # | 1958 | # | 1959 +- O -+ 1960 $1 transistor | # | 1961 | # | 1962 $100,000 small computer (IBM 1620) | # | 1963 | O | 1964 +- # -+ 1965 $0.08 transistor (IC) | # | 1966 $1000 4 func calculator | # | 1967 $6000 scientific calc. | # | 1968 $10,000 small computer (PDP 8) | # | 1969 +- O -+ 1970 $200 4 func calculator | # | 1971 | # | 1972 1K RAMS (1 c/bit) | # | 1973 | # | 1974 $1000 small computer (PDP 11) +- O -+ 1975 4K RAMS (.1 c/bit) | # | 1976 $5 4 func calc (.05 c/trans) | | 1977 | | 1978 | | 1979 +---+---+---+---+---+---+---+---+---+---+ The numbers indicate a remarkably stable evolution. The price per electronic function has declined by a steady factor of ten every five years, if speed and reliability gains are included. Occasionally there is a more precipitous drop, when a price threshold which opens a mass market is reached. This makes for high incentives, stiff price competition and mass production economies. It happened in the early sixties with transistor radios, and is going on now for pocket calculators and digital wristwatches. It is begining for microcomputers, as these are incorporated into consumer products such as stoves, washing machines, televisions and sewing machines, and soon cars. During such periods the price can plummet by a factor of 100 in a five year period. Since the range of application for cheap processors is larger than for radios and calculators, the explosion will be more pronounced. The pace of these gains is in no danger of slackening in the forseeable future. In the next decade the current period may seem to be merely the flat portion of an exponential rise. On the immediate horizon are the new semiconductor techniques, I2L, and super fast D-MOS, CCD for large sensors and fast bulk memory, and magnetic bubbles for mass storage. The new 16K RAM designs use a folded (thicker) cell structure to reduce the area required per bit, which can be interpreted as the first step towards 3 dimensional integration, which could vastly increase the density of circuitry. The use of V-MOS, an IC technique that vertically stacks the elements of a MOS transistor is expanding. In the same direction, electron beam and X-ray lithography will permit smaller circuit elements. In the longer run we have ultra fast and efficient Josephson junction logic, of which small IC's exist in an IBM lab, optical communication techniques, currently being incorporated into intermediate distance telephone links, and other things now just gleams in the eye of some fledgling physicist or engineer. My favorite fantasies include the "electronics" of super-dense matter, either made of muonic atoms, where the electrons are replaced by more massive negative particles or of atoms constructed of magnetic monopoles which (if they exist) are very massive and affect each other more strongly than electric charges. The electronics and chemistry of such matter, where the "electron" orbitals are extremely close to the nucleus, would be more energetic, and circuitry built of it should be astronomically faster and smaller, and probably hotter. Mechanically it should exhibit higher strength to weight ratios. The critical superconducting transition field strengths and temperatures would be higher. For monopoles there is the possibility of combination magnetic electric circuitry which can contain, among many other goodies, DC transformers, where an electric current induces a monopole current at right angles to it, which in turn induces another electric current. One might also imagine quantum DC transformers, matter composed of a chainlike mesh of alternating orbiting electric and magnetic charges. I interpret these things to mean that the cost of computing will fall by a factor of 100 during the next 5 years, as a consequence of the processor explosion, and by the usual factor of 10 in the 5 years after that. As an approximation to what is available today, note that in large quantities an LSI-11 sells for under $500. This provides a moderately fast 16 bit processor with 4K of memory. Another $500 could buy an additional 32K of memory, if we bought in quantity. The result would be a respectable machine, somewhat less powerful than the KA-10, for $1000. At the crude level of approximation employed in the previous section, a million machines of this type should permit human equivalence. A million dollars would provide a thousand of them today (a much better buy, in terms of raw processing power, than a million dollar large processor). In ten years a million dollars should provide the equivalent of a million such machines, in the form of a smaller number of faster processors, putting human equivalence within reach. We now have the problem of what to do with this roomful of isolated computers. The next section describes what seems to be the best known solution to the problem of interconnecting a large number of processors with maximum generality, at reasonable cost. References: McWHORTER, Eugene W. "The Small Electronic Calculator", Scientific American, Vol. 234, No. 3, March 1976, 88-98. HODGES, David A., "Trends in Computer Hardware Technology", Computer Design, Vol. 15, No. 2, February 1976, 77-85. SCRUPSKI, Stephen E., et al., "Technology Update", Electronics, Vol. 48, No. 21, McGraw-Hill, October 16, 1975, 74-127. VACROUX, Andre G. "Microcomputers", Scientific American, Vol. 232, No. 5, May 1975, 32-40. TURN, Rien "Computers in the 1980's", Columbia University Press, Rand Corporation, 1974. TIEN, P. K. "Integrated Optics", Scientific American, Vol. 230, No. 4, April 1974, 28-35. HITTINGER, William C. "Metal-Oxide-Semiconductor Technology", Scientific American, Vol. 229, No. 2, August 1973, 48-57. BOBECK, Andrew H. and Scovil, H. E. D. "Magnetic Bubbles", Scientific American, Vol. 224, No. 6, June 1971, 78-90. HEATH, F. G. "Large-Scale Integration in Electronics", Scientific American, Vol. 222, No. 3, February 1970, 22-31. RAJCHMAN, Jan A. "Integrated Computer Memories", Scientific American, Vol. 217, No. 1, July 1967, 18-31. HITTINGER, William C. and Sparks, Morgan "Microelectronics", Scientific American, Vol. 213, No. 5, November 1965, 56-70.

Section 4: Interconnecting Processors

The highest bandwidth and most flexible way for a number of computers to interact is by shared memory. Systems of the size considered here would require a large, but not unreasonable, address space of 100 billion words (40 bits of address). They would also demand memories with a thousand to a million ports. Although a variant of the method below could be used to construct such monsters, they would cost much more than the processors they served. Alternatively, we might consider how a mass of the usual kind of machine, each with a moderate amount of local memory, could communicate through their bandwidth limited IO busses. Define a message period as the minimum interval in which a processor may transmit and receive one message. Assuming that messages are addressed to particular destination machines, an ideal interconnection system would allow every processor to issue a message during each message period, and deliver the highest priority message for each processor to it, providing notices of success or failure to the originators. The delay between transmission and reception of a message would be uniform and short, and the cost of the network would be reasonable (on the order of the cost of the processors served). It so happens that this specification can be met. The following construction is a design due to K.E. Batcher [Batcher], extended considerably to permit pipelining of messages and higher speed. Log^2 sorting net construction: Batcher's major contribution is a method for constructing nets that can sort N numbers in (log(N))^2 time, using only Nxlog^2 N primitive two element sorters. Two nets plus a little additional circuitry can accomplish the message routing task. The construction consists of a method for making a merger for two ordered lists of 2N numbers (call it a 2N-merger) from two N-mergers and 2N additional primitive elements. A primitive sorting element is a 1-merger. An N-merger can thus be constructed from Nxlog2(N) primitive elements. A sorter is made of a cascade of ever larger mergers, forming length two ordered lists from individual inputs with 1-mergers, combining these pairwise into length 4 lists with 2-mergers, and so on, until the entire sorted list of N numbers comes out of a final N/2-merger. The number of primitive elements involved is Nxlog2(N)x(log2(N)+1)/4. sorting __________ elements A1 ______________| |_________________________ _____| | _____ A2 ______ | ____| 3 |_______________| |_____ | || | | ____________| |_____ A3 ______|_|| |__ M |__|__________ |___| | | | E | | | A4 _____ | | ___| R |__|_________ | || | ___| G | | || _____ A5 _____||__|| __| E |__|________ ||_| |_____ I || || _| R | | _______||__| |_____ O A6 ____ || ||| | |__||______ || |___| N ||| ||| |________| || ||| U ||| ||| || ||| P ||| ||| || ||| _____ T ||| ||| || |||__| |_____ U ||| ||| || _____||___| |_____ P ||| ||| ||| || |___| T B1_____|||___||| __________ ||| || U |||____||_| |__||| || S ||_____||_| | || || _____ T B2 ___ |______||_| 3 |___|| ||___| |_____ | || | | | ____|____| |_____ B3 ___|_______|| |__ M |____|| | |___| | | | E | | | B4 __ | | ___| R |_____| | _____ ||________|_| G | |____| |_____ B5 __|_________| | E |_______________| |_____ |___________| R | |___| B6 ______________| |_________________________ |________| FIG 1: An illustration of Batcher's odd-even merger construction, producing a 6-merger from two 3-mergers and five primitive elements. The first and last outputs undergo less delay than the intermediate ones, which is undesirable. The delays can be equalized by passing the first and last outputs through an extra primitive element. Batcher provides an alternate construction, which he calls a "bitonic sorter" which merges, has constant delay, and uses the same number of elements as the above construction with the extra element added. Communication scheme organization: ___________ _________ _____ ____________ 1 ->| |-->| |-->| |-->| |-> 1 | | | | | | | | 2 ->| |-->| |-->| |-->| |-> 2 | N | | | | | | | Processor Processor 3 ->| |-->| |-->| |-->| |-> 3 | Input | | | | | | | Acknowledgement Outgoing | | | | | | | | . | Sorter | . | | . | | . | | . Input Message . | | . | | . | | . | | . . | (Desti- | . | | . | | . | 2N | . Ports Ports | nation | | | | | | | | field) | | 2 | | E | | Input | | | | N | | x | | | N ->| |-->| |-->| c |-->| Sorter |-> N |_________| |__ M | | h | | | | e | | a | | (Source | ---------------- __| r | | n | | field) | 1 -->| g |-->| g |-->| |-> 1 | e | | e | | | 2 -->| r |-->| r |-->| |-> 2 | | | | | | Processor Dummy 3 -->| |-->| |-->| |-> 3 | | | | | | Incoming Message | | | | | | . | | . | | . | | . Message Injection . | | . | | . | | . . | | . | | . | | . Ports | | | | | | | | | | | | N -->| |-->| |-->| |-> N |_______| |___| |__________| FIG 2: Block diagram of the multiprocessor interconnection scheme. The cost per processor, and the delay in the net, grows as the square of the log of N, the number of processors. _____________________________________________________________________________ | | | | | | | Data | Source Address | 0 | Priority | Destination Ad | |________________________|_________________|___|__________|_________________| FIG 3: Processor message format. High order bit is on the right. _____________________________________________________________________________ | | | | | | | Unused | Position Number | 1 | Infinity | Position Number | |________________________|_________________|___|__________|_________________| FIG 4: Dummy message format. The interconnection scheme is diagrammed in Figs 2, 3 and 4. Each processor is assigned a number, its "address", as indicated. In the sorters and the merger the smaller numbers come out towards the top of the diagram. Messages are serial bit streams, and consist of a destination processor address, a priority number (invented by the originating computer), a one bit "dummy" flag field (set to 0 for actual messages), the address of the source processor (i.e. a return address), and the data to be communicated. A low priority number implies high priority. Zero is the highest priority. The net is assumed to run at 100% duty cycle, with the processors emitting successive synchronized waves of messages. Every processor emits a message every message interval. The following discussion examines a single message wave. The first sorting net orders the messages by destination address, and within a given destination by priority number. Thus the upper inputs of the merger receive a list of messages, grouped by destination, with the highest priority message to each processor heading its group. The lower inputs of the merger receive N dummy messages, exactly one for each destination processor. The priority field is the highest possible (i.e. zero), the dummy flag is 1, the source address is the same as the destination, and the data portion is unused. Merging these dummies with the sorted list of real messages results in a list still grouped by destination, with each group headed by a dummy, by virtue of its high priority, followed immediately by the highest priority real message, if any, for the destination. This list is fed to the exchange network, which examines adjacent pairs of messages (considering overlapping pairs), and exchanges the data portions of a pair if the first member happens to be a dummy to a given address, and the second is a real message to the same address (i.e. it is the highest priority real message to that destination). The sorting network following the exchanger sorts the messages by the field begining with the dummy flag, which acts as the high order bit, followed by the source address. Since there were N real messages, one from each processor, and N dummies, also nominally one from each processor, and since real messages are sorted ahead of dummy messages due to the high order bit, (i.e. the dummy flag) being 0, the second sorter restores the messages into the same order in which they were introduced. Each processor has two input ports, one labeled "acknowledgement", and the other "incoming message". The acknowledgement input of processor i is connected to output i of the second sorter. The incoming message input is connected to output N+i (i.e. the i'th output of the lower half). In the absence of the exchange network, the i'th processor would receive its own message back on its acknowledgement input, and the i'th dummy message on the incoming message input. Because of the exchanger, however, if the message that processor i had sent happened to be the highest priority message to the requested destination, then the data portion of the message on the acknowledgement input would be that of the dummy it had been swapped with (signaling success). Also, if any messages had been addressed to processor i, the data portion of the highest priority one would arrive on the incoming message port, in place of the dummy message. Thus a processor receives the highest priority message addressed to it on its incoming message port, or a dummy if nobody wanted to talk to it. It receives a dummy on its acknowledgement port if its message has gotten through, or the message back if it hasn't, due to the existence of a higher priority message to the same destination. Actually the serial nature of the sorter causes the destination and priority field to be lost in the source address sorter (it tails after the previous wave of messages). In the case of messages that fail to get delivered, this means that the originating processor must remember to whom it sent the message (about four message times ago in a typical design, due to the latency of the net), if it wants to try again. This is probably undesirable. Also, delivered messages contain no indication of who sent them, having had their source address field exchanged with that of a dummy, unless the source address is included in the data field. These shortcomings can be overcome if the exchanger shuffles the destination address, source address and priority fields in the manner suggested by Figs 5, 6 and 7. Such shuffling can be accomplished with an amount of storage at each exchanger position equal to the number of bits in the destination and priority fields. Before: _____________________________________________________________________________ | | | | | | | Unused | Destination Ad | 1 | Infinity | Destination Ad | |________________________|_________________|___|__________|_________________| _____________________________________________________________________________ | | | | | | | Data | Source Address | 0 | Priority | Destination Ad | |________________________|_________________|___|__________|_________________| After: _____________________________________________________________________________ | | | | | | | Priority | Source Address | Data | Destination Ad | 1 | |__________|_________________|________________________|_________________|___| _____________________________________________________________________________ | | | | | | | Infinity | Destination Ad | Unused | Source Address | 0 | |__________|_________________|________________________|_________________|___| FIG 5: Rearrangements effected by the exchanger in an exchanged pair. Before: _____________________________________________________________________________ | | | | | | | Data | Source Address | 0 | Priority | Destination Ad | |________________________|_________________|___|__________|_________________| After: _____________________________________________________________________________ | | | | | | | Priority | Destination Ad | Data | Source Address | 0 | |__________|_________________|________________________|_________________|___| FIG 6: Rearrangements in an unsuccessful message. Before: _____________________________________________________________________________ | | | | | | | Unused | Destination Ad | 1 | Infinity | Destination Ad | |________________________|_________________|___|__________|_________________| After: _____________________________________________________________________________ | | | | | | | Infinity | Destination Ad | Unused | Destination Ad | 1 | |__________|_________________|________________________|_________________|___| FIG 7: Rearrangements in an isolated dummy message. Package counts: If the numbers to be sorted are sent into such a net as serial bit streams, high order bit first, then a primitive sorting element has two output wires labelled "smaller" and "larger", two input wires and a reset and a clock line, and works as follows: Just before the first bit time the element is reset. Bits then stream into the input terminals, and simply stream out of the output terminals until the first bit position in which the two inputs differ comes along. At that instant, the input with the 0 is connected to the "smaller" output, and the one with the 1 is connected to "larger". This interconnection is latched by the element and all subsequent bits stream from the inputs to the outputs on the basis of it, until the next reset. Such a unit can be built with approximately 20 gates, and introduces one bit time of delay. Careful design should permit an ECL version with a 100 or 200 MHz bit rate. These could be packed into 48 (say) pin LSI packages, 8 independent elements per package (the clock and reset lines are common), 16 per package, configured into 4 2-mergers, 24 per package, arranged into two 4-mergers, and 32 per package containing a single 8-merger. Larger mergers can be constructed out of these using an extension of the bitonic sorter strategies given in [Batcher], resulting in total package counts (and partial eight merger, four, two and single element package counts) shown in Fig. 8 (to re-iterate, a merger size of N refers to one which takes two lists of size N and produces a list of size 2N). Merger size Package count Total Eights Fours Twos Ones 1 1/8 1/8 2 1/4 1/4 4 1/2 1/2 8 1 1 16 4 2 2 32 8 4 4 64 16 8 8 128 32 32 256 96 64 32 512 192 128 64 1,024 384 256 128 2,048 768 768 4,096 2,048 1536 512 8,192 4,096 3072 1024 16,384 8,192 6144 2048 32,768 16,384 16384 65,536 40,960 32768 8192 131,072 81,920 65536 16384 262,144 163,840 131072 32768 524,288 327,680 327680 1,048,576 786,432 655360 131072 FIG 8: Package counts for mergers These counts can now be used to calculate the number of packages required to build sorters of various sizes: Sorter size Package count Total Eights Fours Twos Ones 2 1/8 1/8 4 1/2 1/4 1/4 8 3/2 1/2 1/2 1/2 16 4 1 1 1 1 32 12 4 2 2 4 64 32 12 4 8 8 128 80 32 16 16 16 256 192 96 32 32 32 512 480 256 64 64 96 1,024 1,152 640 128 192 192 2,048 2,688 1536 384 384 384 4,096 6,144 3840 768 768 768 8,192 14,336 9216 1536 1536 2048 16,384 32,768 21504 3072 4096 4096 32,768 73,728 49152 8192 8192 8192 65,536 163,840 114688 16384 16384 16384 131,072 368,640 262144 32768 32768 40960 262,144 819,200 589824 65536 81920 81920 524,288 1,802,240 1310720 163840 163840 163840 1,048,576 3,932,160 2949120 327680 327680 327680 2,097,152 8,650,752 6553600 655360 655360 786432 FIG 9: Package counts for sorters An interconnection net for N processors involves an N sorter, an N merger and a 2N sorter. Thus a 128 processor system would require 304 48 pin sorter packages, 2.3 for each processor. A 1024 processor needs 4224, roughly four per processor. A size 16,384 system needs 7 for each computer. A million processor would have 12.75 per machine. It is likely that the biggest versions of this system will require denser packaging. Remember, though, that a thousand processor system is sufficient for human equivalence if each machine is big and fast enough. It will probably not be necessary to build a megaprocessor in the decade envisioned here. Speed calculations: As outlined, a message consists of a destination address, a priority, a bit, a source address and a data portion. The two addresses must be at least large enough to uniquely specify each machine. Considering the case of a thousand and a million machine system, we note that the address lengths are 10 and 20 bits. Let's make the priority field the same length, leaving room for considerable flexibility in priority assignment schemes. The message portion should be fairly long, to permit messages like memory write requests, which require both a memory address and the data. Four address lengths, say. This gives us a message length 7 addresses long, 70 bits in a 1000 processor, 140 bits in a megaprocessor. The full time taken by a message from start of transmission to completion of reception, in bit times, is the message length, plus the depth of the net (in primitive elements), plus an address and a priority time due to the buffering at the exchanger. The depth of a 1024 processor net is 110 elements. This combines with the message and exchanger delays to result in a transit time of 110+70+20 = 200 bit times. If the bit rate is 100 MHz then messages are delivered in two microseconds. If the bit rate is 200 MHz, the time is 1 microsecond. The net contains about two full message waves at any instant, and a message time is 700 nanoseconds for 100 MHz, or 350 nanoseconds for 200 MHz. The same calculation for a megaprocessor reveals a depth of 420 and a total transit time of 420+140+40 = 600 bit times. This corresponds to 6 and 3 microsecond transits for 100 and 200 MHz data rates. Corresponding message times are 1.4 microseconds and 700 nanoseconds. The net contains slightly more than 3 message waves at a time. Possible refinements: The transit time of the net can be decreased, at the cost of increasing the message times a little, by running some of the primitive sorter elements asynchronously, clocking (and introducing bit time delays) at larger intervals. For instance, an 8-merger might accept a bit time worth of inputs, which would then trickle through four stages of primitive sorters built without shift register delays between them. When everything had settled down, the entire merger would be clocked, and those elements which had decided to latch at this bit time would do so. The delay of the unit is one bit time rather than four, which reduces the 110 or the 420 term in the transit calculations. The settling time is longer, however, so the bit rate would be slower. Perhaps a factor of two in transit time can be gained in this way, at a cost of 1.5 in data rate. At an increase in gate count, several levels of asynchronous primitive sorters can be replaced by an equivalent circuit with fewer gate delays. This might enable a given data rate to be maintained while the transit time was reduced. [Van Voorhis] offers some slight reductions in primitive sorter count, essentially by substituting special case solutions better than the systematic construction wherever possible in a large net. Unfortunately these invariably have an uneven amount of delay along the various paths, making them almost worthless as synchronous nets. They may be useful as designs for asynchronous subnets. I have generalizations of Batcher's constructions using primitive elements that are M sorters (as opposed to 2 sorters). These allow building a merger which combines M sorted lists of size NxM into a single sorted MxMxN length list, out of M mergers each capable of merging M size N lists, and some layers of extra M sorter primitive elements to combine the output of the mergers (these are analogous to the single layer of 2-sorters following the mergers in the odd-even merger of Fig 1. The number of such layers grows (empirically) roughly as log(M)). Although the number of gates needed for a given size sorter grows slowly with M, the number packages used shrinks. This is because each package, being a sorter rather than a merger, contains more logic. My constructions are complete for M<=8, and partially complete for general M. Well, now we have a room full of not only processors, but a massive switching system as well. Can it be made to pay for its keep? The next section examines some programming implications and opportunities. References: BATCHER, K.E. "Sorting Networks and their Applications", 1968 Spring Joint Computer Conference Proceedings April 1968, 307-314. VAN VOORHIS, David C. "An Economical Construction for Sorting Networks", 1974 National Computer Conference Proceedings April 1974, 921-927. KNUTH, D.E. "Sorting and Searching", The Art of Computer Programming, Vol. 3 Addison-Wesley, 1973.

Section 5: Programming Interconnected Processors

A major feature of the scheme outlined is its flexibility. It can function as any of the fixed interconnection patterns of the current lackluster multiprocessors, like Illiac IV, or as a hexagonal mesh, or a 7 dimensional cubic lattice, should that be desired, or the tree organization being considered in a Stanford proposal. It can simulate programmed pipeline machines, such as CDC Star, by using processors as arithmetic units. What is more, it can do all of these things simultaneously, since messages within one isolated subset of the processors have no effect on messages in a disjoint subset. This permits a very convenient kind of "time" sharing, where individual users get and return processors as their demands change. Such mimicry fails to take advantage of the ability to reconfigure the interconnection totally every message wave. It is easy to find particular applications, such as tree searching, where the net could be used effectively by special purpose programs. It is clearly desirable to have systems which minimize the user effort needed to implement these algorithms. As an example of a high level language well suited to the architecture, consider the following parallel variant of Lisp: A little Lisp: Take an existing Lisp and purify it to something closer to its lambda calculus foundations. Flush RPLACA and RPLACD and even SETQ (a pseudo SETQ can be introduced later), and discourage PROGs, which are inherently sequential and do things that could have been stated more clearly as recursive functions. In this system recursive programs are also more efficient. An evaluation is handled as follows. A free processor (one not currently doing an evaluation) receives a message demanding the value of Expr with variable bindings given in ALIST. Call this the task [Expr,ALIST]. If the top level of Expr is F(exprA,exprB,exprC), it generates the tasks [exprA,ALIST], [exprB,ALIST] and [exprC,ALIST], and passes them to three other free processors. These evaluate them and sooner or later return valA (the value of exprA), valB and valC to the original processor. This processor then refers to the definition of F, and in particular to the names of the dummy arguments in the definition. Suppose these were A, B and C. It combines the ALIST it was given with the (name, value) pairs (A, valA), (B, valB), (C, valC) to form a new ALIST'. Then it generates the task [bodyF,ALIST'], where bodyF is the body of the definition of F, which is passed to another free processor. On receiving the value of this expression, it passes it back to whoever had given it the original task, and then declares itself free. If Expr had happened to be an atom, the processor would simply have looked it up in ALIST, and returned its value. If F had been a predefined system function (such as CAR or CONS) the sequence of actions would have been whatever the machine code for those functions (of which each processor would have a copy) specified. The parallelism comes from the fact that the arguments in a multi-argument function can be evaluated simultaneously. This causes moderate parallelism when functions are nested, and can cause explosive parallelism when a function which at some level uses a multi-argument function, invokes itself recursively (as in a tree search). Most non-trivial programs stated as recursive functions do this. The description above implies that the processor waits for the results of tasks it has farmed out to other machines. These waits can be arbitrarily long. Also, the number of tasks spawned can easily become more than the number of processors. In that case, presumably, a processor with a task to farm out would have to wait until a machine becomes free. Keeping a processor idle under these conditions is clearly undesirable. If each processor were time shared, pretending to be many machines, then when the job being run becomes temporarily idle there may be another to switch to. When a message pertaining to a given waiting job arrives, the processor deals with it, and if it provides the information necessary to allow that job to resume, it is resumed. This scheme makes the number of processors available seem larger, perhaps enough to make it possible to acquire a processor for a new task simply by picking a machine at random and asking if it has a free job slot. This works well if the answer is usually yes, (i.e. if there reasonably more slots than tasks), and replaces a more complicated free processor pool method that must be able to deal with many requests simultaneously. Alternatively instead of each processor having its own little pool of running jobs, the whole Lisp system can maintain a communal pool, which processors refer to as they become free. In this organization a task that must pause is placed into the pool (freeing the processor that was running it), to be taken up again by a free processor when its requirements are met. Moderate processing power is required to manage the pool. The list structure is spread out among all the processors in the system. Pointers consist of a processor number and address within processor field. A machine evaluating an expression whose top level function is CONS creates the new cell in its own memory. A CAR or CDR involves sending a message to the processor which owns the cell involved, and waiting for a reply (thus a CONS is usually cheaper than a CAR!). Since RPLACA and RPLACD are eliminated, circular lists cannot be created. This permits garbage collection to be mediated by reference counts. A small fixed reference count field (three bits wide, say) is part of each cell. A processor doing a CONS sends messages to the owners of the cells that the new cell will be pointing to, indicating that their reference count should be incremented. If a cell getting a message of this type has a reference count that has already reached maximum (7 in our example), it sends back an indignant message to the CONS'ing processor saying, effectively, "I'm full, you can't point to me. But here are my CAR and my CDR, roll your own and point to that". This not only makes the reference count field size manageable, but reduces message traffic jams to processors containing very popular cells. It does make it mandatory to use EQUAL rather than EQ when comparing lists. When a processor no longer requires a cell to which it had a pointer it sends a message to the owner saying so. The reference count of the cell is decreased by one, and if it reaches zero, it is freed, and similar messages are sent to the cells pointed to by its CAR and CDR. This garbage collection process goes on continuously and independently of the main computations. How is the A-list of bindings to be handled? The canonical representation, using a simple list of dotted pairs, is very inefficient, since it forces a sequential search. A hash table, as in most current Lisp implementations is also undesirable, partly because it requires an incompatible data type (a relatively large contiguous block of memory), but primarily because in this parallel system there are many different contexts (in different branches of the computation) active at one time. The entire hash table would need to be duplicated whenever new bindings are made (i.e. at every level of evaluation), since the older context is in use in another branch. A nice alternative is to use essentially an A-list, but to structure it as a balanced tree, with each node containing a binding, a left subtree for those variables with names lexicographically (say) less than it, and a right subtree for those greater. An entry in such a structure can be found in time proportional to the logarithm of the number of nodes, and a version of the list with an entry added, deleted or modified can also be constructed in log time without affecting the original in any way, by building a new path down to the element (and sometimes a few side paths, to keep the tree reasonably balanced), which points to many of the subtrees of the original. Since this structure can be built of standard list elements, it can be managed via the normal allocation and garbage collection mechanisms. Call this structure an A-tree. In general applications a balanced tree is capable of being used in this system more effectively than a simple list, because a parallel process can get to all the nodes in log time, whereas the last node takes linear time in a linear list. This suggests that programmers wanting maximum efficiency will be coerced into using them much more frequently than is now the case. This is similar to the pressure on present day Lisp programmers to use PROGS rather than recursive functions, because compiled PROGS run faster. The [expr,ALIST] task description is conveyed simply as a pair of pointers, expr to an S expression version of expr, and ALIST to the head of the balanced A-tree. The result returned in an evaluation is likewise a pointer. Since the list structure is only built upon, and never altered, it is possible to speed up the operation of the system by having individual processors maintain software managed caches of frequently referred to cells that reside in remote machines. This cuts down on the message traffic, and generally speeds things up. A little Algol: Although recursive functions provide an excellent way of exploiting a very parallel architecture, there are other ways. An algorithmic (iterative) language can be made to serve, if a few features are present. Existing Algol unfortunately has few of these. Consider the following pair of statements excerpted from an imaginary program: A := 3 ; B := 4 ; These say that first A is set to 3, then B is assigned the value 4. There is an implied sequentialness. Presumably the programmer knew that the order of the assignments was unimportant, and that they could be done simultaneously. The syntax of the language provided no way for him to indicate this. A simple construct which permits information of this type to be conveyed is the "parallel semicolon", which we will indicate by a vertical bar "|". Statements seperated by parallel semicolons may be executed simultaneously. The following is an example of its use: A := 3 | B := 4 | C := 6 ; D := A+B+C ; The first three statements may be executed together, the fourth must wait for their completion. It is implied that compound statements, bracketed by BEGIN END pairs can be similarly separated by parallel as well as sequential semicolons. If compound statements are permitted to return as a value the value of their last component statement, then the structure T := BEGIN INTEGER A,B,C ; A := expressionA | B := expressionB | C := expressionC ; expression involving A, B and C END; is equivalent to a lambda expression construct in Lisp. A, B and C are the dummy arguments, evaluated in parallel, and the fourth expression, which must wait for A, B and C, is the body of the lambda. The constructs involved can be used in many other ways, unlike a real lambda expression. If the Algol also has recursion, then it is possible to obtain massive parallelism in the same way it was achieved in Lisp, namely by writing functions which recurse deeply, and invoke themselves a few times in parallel at each level. A minor form of parallelism already automatically available resides in arithmetic expressions. Different subexpressions of larger formulas can be evaluated simultaneously, and then combined. This is analogous to the evaluation of Lisp expressions. More massive concurrency can be obtained if the data types of Algol are expanded to include a complete set of array operators and genuine dynamic allocation of arrays. It is this type of data and operator set that makes APL an extremely powerful language in spite of an execrable control structure. The idea is that operators such as addition and multiplication work not only on simple variables representing single numbers, but on whole multi-dimensional arrays. Since arrays are more complex objects a much larger range of operators is required. Included are such things as array restructuring, subscript permutation (generalized transpose), element shuffling, subarray extraction, generalized cross and dot products, etc.. In general an operator combines one or more arrays and produces a value which is a new array, often of a different size and shape. Compounding of such operators substitutes for a large amount of explicit loop control, and results in substantially more compact source code. On conventional computers such code also runs much faster because most the run time is spent in carefully hand coded implementations of the basic operators, which do a great deal at each invokation if the arrays are large. Typical equivalent Algol programs execute second rate compiler produced code almost exclusively. An operator set of this kind provides the same capabilities as a hypothetical parallel FOR loop construct, with greater clarity. Conditionals can be handled by first selecting out subarrays according to the condition, and then applying the appropriate operations. It is desirable in our system for large arrays to be spread out among many machines, so that array operations can be carried out in parallel where possible. An often occurring process is the application of an arithmetic or other operator to one or more arrays of a given size, producing a result of the same size. This could be made maximally efficient if corresponding elements of the arrays resided in the same processor. Assigning a processor to an array element by means of a hash function of both the array dimensions and the indices of the element will cause arrays of different size to be stored in different places, but will put corresponding elements of arrays of the same size in the same place. Now comes the question of how a gaggle of processors managing a given array gets word that an operation concerning the array is to be executed. The initiator of the operation is the single processor running the code requesting it. The fastest way to propagate such information is to initiate a "chain letter". The originating processor can consider the array as two equal (or almost equal) pieces, and send a message to a member of each piece. Each recipient then divides the piece of which it is a member into two, and sends a message to a representative of each of those smaller subsets, etc. When the subset size becomes one, the operation is performed. It may be more efficient to store larger pieces of arrays in individual processors. This adds a slight serial component to the run times, but saves message handling time. How are programs communicated from processor to processor, as the number of executing instruction streams grows? It would be an extravagant waste of memory to duplicate the entire source program in every processor (besides, it might not fit). A software caching scheme is called for. When a processor initiates a subtask, it obtains a free processor and passes to it a moderate amount of the code needed for the task. When the new processor runs out, it sends a message back to the first machine for more. This machine trys to obtain it from its own cache, and if it too is out, sends a message further up in the hierarchy, to the machine which had initiated the task which it is running, and so on until the requested code is obtained. At top level the entire runnable program is assumed to be available from some sort of file, maintained by the combined storage of a number of processors. A little operating systems: Assume that most input/output devices are connected to individual processors, and that, to maximize the bandwidth, each of these processors has only one, or a most a very small number, of devices associated with it. Included among these are communications lines to user terminals, so that each console talks directly to a dedicated processor with sorting net access to the entire system. If we expect to support more than one user on such a system, it will be necessary to have some type of protection scheme, to prevent one user's processes from accidentally interfering with another's. If each processor contains a monitor, then message sending can handled by system calls. The monitor can then check for validity, testing if the requested destination is within the set of processors assigned to the user. This monitor, which every machine assigned to a user must run, can be flexible enough to time share the machine it runs on, to provide multiple simulated processors. Controlling the state of the individual machines' monitors is the task of a global system monitor, operated by several machines, which maintains a pool of free processors, and parcels them out on request, and which also handles file system requests (bulk storage would be connected to a handful of the processors), and allocation of other devices. Processes belonging to a single user will be initiated by a particular master machine, probably the one connected to his console. This master can create a tree of subprocesses, possibly intercommunicating, running on different machines. It should be possible, for example, to have one subset configured as an array processor for efficient implementation of low level vision operations, while another is running an Algol/APL for the less structured analytic geometry needed to interpret the image, and yet a third is operating a Lisp system doing abstract reasoning about the scene. Many existing systems permit this kind of organization, but they are hampered by having an absurdly small amount of computing power. How is a system of this kind initialized, and how does one abort an out of control process taking place in part of it without affecting the rest? A possibility is to have an "executive" class of messages (perhaps signalled by a particular bit in the data portion), which user jobs are not permitted to emit. Reception of such messages might cause resetting of the processor, loading of memory locations within it, and starting execution at a requested locations. A single externally controllable machine can be used to get things going, fairly quickly if emits a self replicating "chain letter". Now consider reliability. The system can obviously tolerate any reasonable number of inoperable processors, by simply declaring them unavailable for use. Failures in the communication net are much more serious, and under most situations will require the system to stop operating normally. It is possible to write diagnostic programs which can track down defective comparator elements or broken data wires. If something should happen to the clock signals to a given level it would be necessary to wheel out an oscilloscope. If reliability were a critical issue it would be possible to include a duplicate net, to run things the while other was being debugged. Disclaimer: The software outlines are obviously only partly baked. This is mostly due to the limited amount of thought and work that has gone into them. On the other hand, it is my belief that even well thought out designs at this point will look naive in the light of experience with a working version of such a machine. Many of the fundamental decisions depend on things difficult to estimate, including the number of processors, communication speed compared with processor speed, memory size, and most importantly the kinds, sizes and mix of operations that people will tend to run (these will surely differ from what is being done now, the limitations being so different). This is not to say that more thought isn't called for. The nature of the memory protection, interrupt scheme, word size, instruction set, I/O structure, etc. of the individual machines should be tailored to permit convenient implementation of the individual processor operating systems, and typical message types. What these are can best be determined by trying to write some monitors, interpreters, compilers and array crunchers. Simulation of the system on a conventional machine is next to useless. Since the total amount of memory envisaged is larger than conventional machines carry, and since a simulation will have to jump around from simulated core image to simulated core image, in order to keep a realistic message synchrony, there would be an enormous amount of disk accessing. The slowness of this, combined with the inherent speed difference, would allow only the most trivial (and therefore unrepresentative) things to be tried, and not many of these. Construction of a moderate size version is a better use of manpower. Undoubtedly there will be mistakes made in the first (or first few) models, which will become apparent after a bit of experience. The flexibility offered should make this design much more attractive to a large class of programmers than current essentially special purpose architectures such as ILLIAC IV. Making it out of existing processors with proven machine languages will help too. I, for one, can hardly wait to start programming even a flawed version of a machine that can process and generate real time television with programs written in Algol, and simultaneously jump over tall game and proof trees in a single bound. Bombast: The enormous shortage of ability to compute is distorting our work, creating problems where there are none, making others impossibly difficult, and generally causing effort to be misdirected. Shouldn't this view be more widespread, if it is as obvious as I claim? In the early days of AI the thought that existing machines might be much too small was widespread, but there was hope that clever mathematics and advancing computer technology could soon make up the difference. Since then computers have improved by a factor of ten every five years, but, in spite of reasonably diligent work by a reasonable number of people, the results have been embarrassingly sparse. The realization that available compute power might still be vastly inadequate has since been swept under the rug, due to wishful thinking and a feeling that there was nothing to be done about it anyway and that voicing such an opinion could cause AI to be considered impractical, resulting in reduced funding. There is also an element of scientific snobbery. Many of the most influential names in the field seem to feel that AI should be like the theoretical side of physics, the essential problem being to find the laws of universe relating to intelligence. Once these are known, the thinking goes, construction of efficient intelligent machines will be trivial. Suggestions that the problems are essentially engineering ones of scale and complexity, and can be solved by incremental improvements and occasional insights into sub-problems, are treated with disdain. This attitude is a variant of the philosophical notion that all truth can be arrived at by pure thought, and is unfounded and harmful. One wonders what state space travel would be in if the Goddards and von Brauns had spent their time trying to find the universal laws of rocket construction before trying to build space ships. AI needs a stronger experimental base. Like other branches of endeavor (notably physics, aeronautics and meteorology), we should realize our desperate need for more computing, and do things about it.