Inside the Biggest Man-made Brain (Apr, 1947)
This computer contains 13,000 relays, each rated to perform for at least 100 million operations. If the transistors in your CPU were this reliable it would last less than 100 milliseconds.
Inside the Biggest Man-made Brain
Navy’s new calculator has steel bones, silver nerves, paper impulses, and can make mistakes.
By Stephen L. Freeland
THE LARGEST brain in the world today is a mammoth electrical mathematician being built at Harvard’s Computation Laboratory for the U. S. Navy Proving Grounds at Dahlgren, Va. But its reign as king of the robots will be brief.
Work already has begun on faster, better calculators based on the lessons learned in creating this machine, known as the Dahlgren Calculator, or Mark II, just as this one was designed to be the big, tough brother of Mark I, which was built for Harvard during the war by the International Business Machines Corp. (PSM, Oct. ’44, p. 86). Mark II, however, will not be retired. Even Mark I has many years of useful labor ahead. There is plenty of work waiting for all the big calculators now in existence and on the drawing boards. Mark I is still churning out answers to abstruse mathematical problems 24 hours a day, and Mark II will be taken to Virginia next month to begin an equally strenuous career.
Mark II is simply larger, faster, more versatile and more flexible than Mark I. The difference between the two machines, however, starts right at the eight-foot-high front panel, which is 60 feet long and shaped like a shallow “U” in Mark II, 51 feet long and straight in Mark I. Adders, multipliers, interpolators and instruction mechanisms are all let into this front panel like drawers in a bureau in Mark II. Two men can handle them easily, and if one unit goes haywire, it can be removed and a spare one slid into place with no waste of time. In Mark I no such substitution is possible.
Answers Roll Out on Ribbons
There are other differences, too. Long strips of paper tape run on aluminum spools across four panels of Mark II, and one entire section is devoted to printing and storing these tapes. Punched in code, they pass through delicate fingering devices on the panel; depending on how they are punched and where they go, they tell the calculator what numbers to use and how to use them. Mark I also uses tapes, along with punch cards, but the tape is broad and heavy and forms a complicated and bulky maze behind the panel, where it is hard to get at. On Mark II a tape can be changed in a minute.
Mark II is hooked up to a battery of four teletypewriters, with a fifth as a standby, while Mark I has only two. Mark I displays a 30-foot stretch of glass window behind which the spinning gears of the counters may be observed, whereas Mark II has no visible moving parts except for the spools of tape.
Unlike as the two machines are in appearance, it is not until one goes behind the panels that their real difference is seen. While Mark I performs its functions almost noiselessly by a combination of mechanical gears and electrical relays, Mark II’s calculations are all done in a whirling clatter of relays. There are 13,000 of them in place now, and positions are available for 2,000 more if needed.
Actually, Mark II uses three kinds of relays: a single coil, a double coil, and a double coil with a latch. All were designed and developed by the staff of the Harvard Computation Laboratory in co-operation with the Autocall Company, of Shelby, Ohio. Their contacts are made of a special silver alloy that enables them to operate at least 100,000,000 times at a pulse rate of five milliseconds.
Each type of relay has a different job to perform. The single coil simply acts as an immediate transfer agent for a number, shooting it right into a computation; the double-coil relay can take a number and hold it for a certain predetermined number of pulsesâ€”carry the number in its head, in effect; the third type, the double coil with latch, will hold a number indefinitely until it receives a special signal to put it into the calculation.
Mark II can perform in less than one second multiplication that takes Mark I nearly six seconds. And since Mark II has four multiplier units, compared to only one in Mark I, it can do four such operations at the same time. All the other jobs are speeded up in like degree, and frequently the Mark II will get answers faster than the one-per-second rate at which the printer can record them.
Faster Recorders Are Needed
The business of speeding up recording, incidentally, is bothering the builders of big calculators right now. The Army’s ENIAC (Electronic Numerical Integrator and Computer) and EDVAC (Electronic Discrete Variable Computer) are perhaps 1,000 times faster in getting answers than Mark II (PSM, Apr. ’46, p. 83). They hit the same bottleneck of printer speed, however, when it comes to writing them down.
The not-yet-completed EDVAC, for example, uses a series of mercury tanks capable of “memorizing” up to 1,000 10-digit numbers and referring to any of them in an average time of 1/5,000 second. Also, its system of magnetic tapes, upon which problems are coded and their results obtained, is similar in purpose but superior in speed of operation to that of Mark IIs perforated paper tapes. Its actual over-all speed, however, is still dependent upon the rate at which its conclusions can be written down. A possible solution to the problem of printer speed has been suggested by the development at Harvard of the “numero-scope,” a cathode tube that would flash an actual number when energized. A battery of these tubes would, of course, permit the calculator to write any number it wanted as quickly as it got the number. At present, however, there is no film fast enough and of sufficiently fine grain to get clear photographs of numbers flashing so swiftly.
Simplicity Speeds Assembly
Mere speed, even when its full use is made possible, is only one of the advantages obtained by using relays in Mark II. Equally important is their greater accessibility for maintenance and repair. Ranged in banks along the exteriors of six 15-foot-long cubicles, the relays are easy to get at.
They are, moreover, interchangeable with others of the same type; to make replacement completely foolproof, each type of relay has its own color on its case and on the blank into which it should go. A man would have to be color-blind to try to jam a black double-coil latch relay, for instance, into the
place belonging to a red single-coil relay. The wiring of Mark II is also a masterpiece of accessibility and mobility. Every panel can be moved without touching any but external wires. These external wires, like the wires on the panel itself, are clearly identified. Panel wiring was completely laid out, laced on jigs and brought to the panel itself as a unit before a single connection was made.
When Mark II goes from the computing lab in Cambridge, Mass., to the Proving Grounds at Dahlgren sometime in June, only 15,000 out of some 300,000 connections for over 1,000,000 feet of wire will have to be broken and remade. The bolted angle-iron construction of the machine’s framework makes it simple to take apart and move.
Convoy Will Move It
Moving will be done by a 25-trailer-truck convoy, with one whole truckload devoted to testing equipment. Once the machine gets to Dahlgren, it will take two to three months to set it up and test it. The boss on the job will be a graduate engineer from Cornell, Mr. Frederick G. Miller, who has worked on Mark II from the blueprint stage.
Mr. Miller, with a staff of five mathematicians and a crew of 12 maintenance and repair men, all trained under the supervision of Dr. Howard H. Aiken, head of the Computation Laboratory at Harvard and a leading authority on large-scale calculators, will stay with Mark II as civilian employees of the Navy. They will keep the machine running 24 hours a day, every day in the year.
Testing a big calculator is like training a muleâ€”the man has to know more than the mule. There will be the arduous matter of checking all contacts, then every circuit, then groups of circuits (there is a special machine for this, built at Harvard, which acts like a baby calculator itself) and then of making trial runs of problems.
Because Mark II is built in two halves, like the two lobes of the human brain, it will be possible to check one half against the other. Finally, the entire machine will be given a problem for which the answers are knownâ€”perhaps something Mark I has done already. If it pulls through that, Mark II will be ready to roll for the Navy.
Even after it gets into operation, however, Mark II can make mistakes, and one of the biggest jobs Mr. Miller and his team will have will be guarding against them. The machine itself will help in this. It has 30 “check circuits” built into it, and the mathematician who sets up a problem punches instructions into the tape that brings these checks into operation.
If a result fails to pass the check circuit, bells ring, lights flash, and the whole computer stops like a tilted pinball machine. Then the trouble-shooters go to work, tracking down the source of error by the same process of progressive isolation a radio repairman uses in fixing a balky set.
Six Twelves vs. Twelve Sixes
A simple example of how a check circuit is used is in multiplication. Whether done with paper or pencil or by relays, multiplication is, after all, only a way of adding. Six times 12, for instance, is 72, whether one gets the answer by grade-school methods or by adding a column of 12 sixes. Mark II, like most calculators, does it the hard way, taking 12 separate sixes out of 12 separate relays and adding them.
To check that, the machine is also asked to take six twelves out of their relays and . add them. Then both results are passed through the check circuit: if they agree, fine; if not, the machine stops. In much the same way, checks are put in for using trigonometric functions to establish mathematical identities. These check circuits were developed by the Computation Laboratory and are found both on Mark I and II, but not on any other existing machines.
Sometimes these check answers are not absolutely the same, but they can still get past the check circuit at the mathematician’s discretion. The difference comes about because Mark II is correct only to 10 significant figures (Mark I works out to 23 figures and can be rigged to go to 46), and in “rounding off” two series of numbers from a problem done two different ways a discrepancy may pop up in the last figure.
There is a special “tolerance” control in the check circuit to take care of this. The mathematician may decide that a difference of one in the 10th significant figure is okay for his purposes, so he sets that in the circuit, and if the two answers differ by one or less, they go right on; if they differ by more, the machine stops.
Obviously, much depends on the mathematician. He has to know what the machine can do and just how it does it. That is why those in the laboratory, who even now are engaged in building Mark III and planning Mark IV, look forward to the day when they will turn away from constructing machines and get to work on the even more important task of training mathematicians and engineers to handle them.
Large-scale computers like Mark II are not put to work to find one answer to one equation, any more than a pressure cooker is used to boil one three-minute egg. They are used, instead, to get a long series of answers to the same equation where the variable factors change a little bit each time. A gradual change in a variable should produce a gradual change in answers, which, if charted, would show up as a nice smooth curve.
A mathematician can sense this curve without charting it, just as an old fisherman can tell when his boat is drifting without taking bearings. Any sudden departure from the smooth curve shows that the calculator is making mistakes. Occasionally, though, there are some surprises.
Some of the people at Harvard got a surprise during the war. One day they received from Washington a formula to run on Mark I, with no further information than orders to do the job right away. They put it in the machine, and for a while the answers built up slowly in a nice, normal curve. All of a sudden they went completely haywire; the curve stopped being a curve and became a mountain peak. Then the peak itself got higher and higher.
The watchers tested the machine up and down, but it looked all right. They thought that the machine had gone crazy but couldn’t prove it.
Months later an atomic bomb burst on Hiroshima. That made the astounding answer to the problem plausible.