In 1956, while on the SAGE computer project and on sub assignment as a liaison to Gene Amdahl, I was asked by Steve Dunwell, a former associate of NSA, if I would like to join a project to plan a special computer for NSA. I accepted. Three of us were sent to Washington for polygraph test. I was the only one who passed the test. Shortly afterwards, I was established, alone, except for secretary in a secure area in the 701 building in Poughkeepsie.
Steve Dunwell, my boss, said that we would have a machine similar to SLED with two word length input registers and two word length output registers all of which would have multilevel index capabilities and could address to the bit level. The machine would have a 60 bit word because 60 is divisible by all integers from 1 through 6 which would aid in handling characters of different bit lengths. A logic unit between the two input and output registers would operate on two characters at a time of 6 bits or less. With that outline, I was to study problems NSA would provide and determine how well such a machine would perform. The Army courier provided by NSA was Nick Blazenski, later, an IBM employee.
After studying the problems supplied by NSA, I determined that about half of the problems could be managed by sorting the problems into manageable size. For these problems, I convinced Dunwell that the proposed configuration would make a very efficient sorting machine. The remainder of the problems required tables with such a large number of cells that if a word was used for each cell in the table, the tables would be too large for the proposed memory. At this time in computer development, working memory for the CPU used magnetic core memory and was manufactured in packages of 4096 words. This memory, also, had to contain the operating system.
Since the table entries were not expected to require any cell the size of a full word of 60 bits, I considered dividing the 60 bit word into cells which would permit counts the size of 2 to the power of some integer N.
The missing 4 addresses, 60 through 63, in address space at the end of each word created a problem. To do the arithmetic to correct for the missing 4 addresses would slow the s performance. I told Dunwell that if he would permit a 64 bit word and an 8 bit character, logical table cells of the size of 2 to the power N could be created by shifting the address, created by the operation on the two characters, N spaces to the left and adding N zeros to the right. I assumed that for speed, tables could begin on a power of 2 large enough to contain a table and that the address of the entries, obtained by the operations performed by the logic unit, could be shifted and ORed into the zeros to create the address. Dunwell said that if I could sell it to NSA, we would go with a 64 bit word and an 8 bit character. NSA accepted the suggestion at a meeting in the Summer of 1956.
During the project, I was given a list of names given to me by NSA for consideration for the machine. One of the names in the list was Harvester. I suggested shorting it to Harvest. I continued to work on estimating the time the conceptual equipment would require solve the problems submitted to me. A set of problems was given to me for an estimate without telling me what kind of equipment was creating the encipherment. When I gave an estimate that the conceptual machine would take about a week, some strange looks crossed the faces of the people receiving the report. I deduced that the problems probably related to U.S.A. equipment.
In the Fall of 1956 Paul Herwitz joined the project and was assigned to write up the logical connectives which were to be provided with the machine.
In the Spring of 1957, a committee was assembled to create a specification manual for Harvest to obtain NSA funding to continue the project. The committee consisted of the Harvest associated people, me as chairman, Paul Herwitz, and Bill Lawless, a former associate of NSA, and a person whom I had never seen before or afterwards but whom Dunwell said was a former NSA analysts whose name I do not know; and people associated with Stretch, Werner Buchholtz, Fred Brooks, and Jim Pomerene. At the beginning of the meeting, Werner Buchholtz suggested that the 8 bit character be called a bite and be spelled byte. I accepted the suggestion. The manual was written and dated May 1, 1967. An interesting side note is that the manuals were numbered starting with 001 which I kept. Later Fred Brooks asked me to make a manual with the number of 000 and give it to him, which I did. NSA wanted the first numbers. So I had a series of covers created with an HNN number where the NN was a series of numbers starting with the number 1.
I left the project shortly after the first Harvest manual was created. A short while later, I was asked by Bill Lawless what would be needed to start the tables at a location other than at a power of 2. I told him that rather than logically OR the results together, a full adder would be required.
Last updated June 17, 2009
Web page copyright 2009 Eric Smith