Engineering Research Associates and the Atlas Computer (UNIVAC 1101)

From Folklore
Jump to: navigation, search

Unisys History Newsletter.
Volume 3, Number 3
June, 1999
by George Gray

The first Atlas computer was developed because the U.S. government had a serious interest in secret codes. One of the ingredients of Allied success during World War II was the ability to read the secret messages of the Axis powers and, in the case of Japan, to prevent them from reading ours. Cryptology, the science of coding and code breaking, was very important and also very secret: some of the major successes were not made public until thirty years later. The U.S. had broken the Japanese diplomatic code in 1940 and it continued to read those messages throughout the war. The U.S. Navy cracked the Japanese naval code early in 1942. This was a major factor in the victory at the Battle of Midway and contributed to the success of other operations for the rest of the war. The U.S. was not the only country doing this. The British built the Colossus computers in 1943 and 1944 to help decipher German coded messages. At the end of World War II, Captain Joseph Wenger, head of the Navy's cryptoanalytic group was extremely concerned that the expertise which had been built up during the war would be lost as the scientists and engineers returned to civilian life. He encouraged a group of Navy cryptologic analysts from the Navy's Communications Supplementary Activity - Washington (CSAW) led by Howard Engstrom and William Norris to form a private company to do contract cryptologic work for the Navy. Wenger, Engstrom, and Norris attempted to get financing from the well-known New York investment firm Kuhn, Loeb, but the firm felt that the project was too risky. Just as they were giving up hope, John Parker, an investment banker who had directed a glider manufacturing company, Northwest Aeronautical Corporation (NAC), during the war, entered the picture. Parker, Engstrom, and Norris established Engineering Research Associates (ERA) in 1946, housed in the former glider factory in St. Paul, Minnesota. ERA provided work for many of the NAC employees. About 40 members of the CSAW staff moved to St. Paul, and the Navy transferred its wartime computing machine laboratory from Dayton, Ohio to St. Paul during the summer of 1946. NAC itself was dissolved in 1948.

The ERA/NAC combination started out with two research contracts given to it by Wenger. Naval personnel were assigned to the factory, which from 1946 through 1950 was designated as a Navy Reserve base commanded by a Captain, to supervise ERA's activities. Because the work was highly secret, there were armed guards at the entrances to work areas. Conditions in the factory, which had been built in 1920 as a foundry for the American Radiator Company, were rather primitive. George Champine recalled: "Some of the lower windows had no screens, so sparrows and swallows would get in and fly around the high ceiling. It was necessary periodically to chase birds out of the building. In the winter, the inside temperature dropped so low ... that programmers wore overcoats and mittens at their desks. In the summer, the temperature rose to the point that whole departments worked without shirts on, and draftsmen has trouble with sweat dripping on their drawings." The work for each contract was called a "task" and was assigned a number. The first tasks involved the development of various components, such as drum memories, which could be used to facilitate cryptologic work. In the first year of operation, the company had revenues of $1,500,000 and made a profit of $35,000. Even after the National Security Agency (NSA) took over the cryptologic work of CSAW and other groups, many of the formal contracts with ERA were done through the Navy's Bureau of Ships. As part of its program of secrecy, the NSA tried hard to pretend it didn't exist. As far as public record went, it had no budget, since its appropriations were buried in the budgets of other agencies, and it fit right in to say that ERA's products were for "the Navy." ERA tried to diversify and did obtain contracts for other projects, such as building a device for the Great Northern Railroad to weigh railroad ore cars, studying the effects of large non-nuclear explosions for the Army and Air Force, and designing the radio antenna which was used on the Boeing 707 airliner, however these activities never amounted to a large share of the company's revenue.

ERA built various cryptoanalytic drum memory devices, under the code-names Omalley, Hecate, Warlock, Demon, and Goldberg. The first Goldberg, completed in 1947, had a 34-inch diameter drum with magnetic tape bonded to the drum surface to provide a recording medium. For writing, the drum advanced slowly, one row at a time, but it was run at 50 rpm for reading. An ERA team designed and built the Demon as a crash project, urgently required by CSAW, between March and October 1948, leaving the project team exhausted. The Demon was a large machine with 8000 vacuum tubes which read data in from paper tape to drum storage and had extensive circuitry for shifting the data and searching for patterns. CSAW had specified a design which related to a particular cryptoanalytic system used by the Soviet Union and it was built into the circuits. In February 1949, the Soviet Union changed its system, rendering the Demon useless. The Goldberg II was an even more ambitious device which employed both photoelectric sensing and paper tape scanning to achieve very high input rates. It used over 7000 vacuum tubes and had 36 decimal counting circuits. The Goldberg II was not operational until 1951. In 1946 James Pendergrass, a member of the CSAW staff had attended a series of lectures on computer technology presented by the staff of the Moore School at the University of Pennsylvania. He was convinced of the need to build general purpose devices which could be reprogrammed. In 1947, CSAW asked ERA to begin work on Task 13, to produce a design for a computer. Due to bureaucratic politics within the U.S, government, where the National Bureau of Standards (NBS) was trying to gain control of anything called a computer, Task 13 referred to its machine as an "analytical engine." The plans were ready late in 1948, and ERA was told to go ahead and build the machine. It was code-named Atlas, which was the name of a mental giant in a then-popular comic strip called Barnaby, drawn by an artist named Crockett Johnson. Arnold Cohen was in charge of the logical design, while Jack Hill and Frank Mullaney led the engineering development team. The original design envisioned an electrostatic memory, but since electrostatic memories were still unreliable, the Atlas used a magnetic drum memory. ERA was more familiar with drum memory technology, and in fact had done design studies for IBM and NBS in 1949-1950 for drum memory computers, neither of which was built.

The Atlas was 38 feet long, 20 feet wide, and used 2700 vacuum tubes. Its U-shaped operations console, made up of rows of switches and lights, sat on top of an ordinary metal desk. The drum memory was 8.5 inches in diameter, rotated at 3500 rpm, and held 16,384 24-bit words. To provide fast access, the drum had 200 read-write heads. Numeric values were expressed in binary, with one bit for the sign and 23 bits for the value. The Atlas represented negative numbers in one's complement format where the negative is formed by subtracting each binary digit from one; thus +5 is 101 and -5 is 010 (namely 111-101). The use of octal (base 8) notation (where each octal digit represents three bits) is much less tedious than writing out long strings of zeros and ones, so +5 in a 24-bit word can be expressed as 00000005 and -5 as 77777772. One's complement format simplified the logical design of the processor, but it does allow for both positive and negative zero: 00000000 and 77777777. For arithmetic operations, the processor's accumulator (A) register could hold 48-bit double-word binary values as intermediate results, but only 24 bits could be stored back to memory. The accumulator was fundamentally subtractive, addition being carried out by subtracting the one's complement of the number to be added. This appeared rather devious, but the "subtractive adder" did reduce the likelihood of coming out with negative zero in normal operations. The instructions were also 24 bits long, with 6 bits for the function code, 4 bits for the "skip" value (telling how many memory locations to skip to get to the next instruction in program sequence), and 14 bits for the memory address. There were 43 different functions. The following example program, taken from a 1953 manual, finds the sum of the first fifty odd integers.

Address Contents Descriptions
37777 11010020 Load (11) A register with constant at address 10020
00000 35010022 Store (35) it into address 10022 to be counter value
00001 11010016 Load (11) A register with the value 1 (from address 10016)
00002 35010021 Store (35) current sum into address 10021
00003 43000000 Transfer (43) the value ( 1) in A register into Q register
00004 11010017 Load (11) a 2 (from address 10017) into A register
00005 42000000 Add (42) value in Q register to value in A register
00006 43000000 Transfer (43) this total from A register to Q register
00007 12010021 Add (12) previous sum from address 10021 to value in A register
00010 35010021 Store (35) new sum into address 10021
00011 11010022 Load (11) A register with counter from address 10022
00012 47000014 If A is non-zero jump (47) to address 00014 (decrement A by 1)
00013 55000000 Stop
00014 35010022 Store (35) reduced counter into address 10022
00015 45000004 Jump (45) to address 00004
10016 00000001 The value 1
10017 00000002 The value 2
10020 00000060 The value 48 (octal 60)
10021 00000000 Storage location for sum
10022 00000000 Storage location for counter

Note that the answer is left sitting in address 10021 (at instruction 10), since for the sake of brevity the example did not go into the intricacies of printing the answer. This example was also simplified in the fact that it deliberately did not take account of drum rotation time. The drum was, of course, constantly rotating while the processor was executing each instruction, so that there was no guarantee that the read-write head would be positioned at the next instruction when execution of the previous one was completed. In all likelihood the computer would have to wait until the drum rotated to the proper location. To get faster program execution, a programmer would have to allow for instruction execution time and place the next instruction in the location where the read-write head would be. Thus, logically sequential instructions would not be in physically sequential drum memory locations. The skip field in the instruction word indicated how many memory locations to skip in finding the next instruction in logical sequence. This meant that the efficient version of the program had the instructions scattered around in order to minimize unnecessary drum rotation, an approach called minimum latency coding. Naturally, this made the programming task even more complex.

ERA delivered the first Atlas to the NSA at a CSAW location on Nebraska Ave. in Washington in December 1950. The installation team set up the machine and tested it in just three days and went home in time for Christmas. NSA made extensive use of it and a second Atlas which was delivered in March 1953. The Atlases were very reliable, running 24 hours per day, seven days a week, with ten percent of each day for scheduled preventive maintenance. The work was all highly secret, however ERA obtained permission to offer a modified version of the Atlas in the commercial market. There was talk of calling the commercial version MABEL, but then Jack Hill suggested that since the Atlas had been Task 13, it be called the 1101 (thirteen in binary). This was done, and it was publicly announced in December 1951. The 1101 had no operating or programming manuals at that time, and the only input-output facilities were typewriter and paper tape. Programs were input in octal machine code punched on paper tape. Given these primitive conditions, it is not surprising that no one bought an 1101.

Sometime in 1953 or 1954, ERA installed a third 1101 at its own Washington office in a storefront at 555 North 23rd St. in Arlington, Virginia. ERA hoped to run it as a service bureau for organizations which had computer work to be done and created an advertising brochure to drum up business. In these years, computer programming (of any sort) was an extremely rare skill, so few prospective customers had any programming resources of their own. The only group of people with experience of the 1101 was hidden away at NSA, and ERA was unable to staff the Arlington office with people to create programs for customers. These circumstances made the service bureau impractical, and the company donated the 1101 to Georgia Tech in November 1954, at an imputed value of $500,000. Georgia Tech put the 1101 into storage until its new computer center building was completed in August 1955, when it was installed and put to work. By the following June it had done over 1,000 hours of work. The two machines at NSA had been retrofitted with core memory sometime prior to the middle of 1956, but Georgia Tech had to endure several cycles of denied budget requests before it came up with the necessary $39,400 in 1958. In November of that year its 1101 received 4096 words of memory designed for the 1103A computer, scaled down to 24 bits by the removal of 12 core planes. As late as 1959, Tech still used the 1101 as the primary machine for teaching programming. The computer center staff even started a project to write a FORTRAN compiler for it, but that was dropped in favor of using an existing language called Runcible on Tech's IBM 650 computer and ALGOL on the new Burroughs B220. However, the 1101 was still running student work in 1961.

In the fall of 1951, James Rand, the chairman of Remington Rand Corporation, approached John Parker with the suggestion that Remington Rand buy out ERA. Parker estimated that ERA would have to raise between $5 million and $10 million in capital in order to expand its role in the computer market beyond that of a supplier of custom-built projects for government agencies, and he did not regard that as likely to happen. He may also have been ready to cash in on his investment in ERA after a 1950 newspaper column by Drew Pearson had drawn attention to the special relationship between the Navy and ERA with the result that government auditors had been scrutinizing ERA's accounts. The relationship with the Navy's procurement offices had become so strained that Joseph Wenger was unable to get an official letter of thanks sent to ERA for having developed the Atlas. In any case, Parker agreed in December 1951 to terms of sale equivalent to $1,700,000, or $8.00 per share. When the company was set up the share price had been $0.10, and every engineer owned some stock. (This widespread stock ownership may have given rise to the story that the price had been arrived at by multiplying the number of ERA engineers (340) by $5,000.) This was solely Parker's decision, since he had the controlling interest in the company. William Norris was opposed to the sale, and indicated his disgust by giving out Remington electric shavers instead of turkeys (which had been a company tradition) to the staff at Christmas, making it plain that he thought that a turkey was a superior gift. Willis Drake said: "It was a particularly John Parker-Bill Norris confrontation... and sheer power prevailed." However other employees were happy to receive Remington Rand stock in place of ERA stock, as William W. Butler recalled: "Many of the founding engineers did hold stock and I think they were quite pleased to find their stock was worth something." After obtaining government approvals, the acquisition was announced at the end of April 1952 and became official in May. Since Remington Rand had purchased the Eckert-Mauchly Computer Corporation in 1950, it now owned both of the pioneer American computer companies.


Thanks to George Champine, former employee of ERA, and to Marjorie McNinch at the Hagley Library in Wilmington, Delaware.


Unisys, UNIVAC, and UNISERVO are registered trademarks of Unisys Corporation. Copyright 1999 by George Gray