Der Ursprung des von Neumann-Computers &
Norbert Bartos
Grundlagen
& von Neumann-Computer
Computer Architecture - Concepts and Evolution; Gerrit A. Blaauw, Frederick P. Brooks jr.; Addison-Wesley; ISBN 0-201-10557-8; ca. 1200 Seiten; ATS 625,-
Norbert Bartos
11.1
Princeton IAS
11.1.1
Highlights
History
We start with the complete Princeton IAS design as published in [Burks et al., 1946], even though the IAS was not built exactly as described.
Dates. Part 1, the most often quoted paper is dated June 28/ 1946. Of Part II Volume 1 is dated April 1, 1947; Volume 2, April 15, 1948; and Volume 3, August 16, 1948. Volume 4 of Part II was announced in Volume 3, but apparently it was never published. Part II deals primarily with programming, but contains some changes to the Original proposal. We describe the Original design, but note the later changes.
Six years after the publication of the first IAS paper Bigelow completed the IAS Computer at the Institute of Advanced Study; and the Moore School of Electrical Engineering at the University of Pennsylvania completed the EDVAC, based on the IAS concepts.
Family tree. The IAS design inspired the designs of many other computers. We describe in this Zoo the Cambridge EDSAC (1949) and the IBM 701 (1953). Some others were the Los Alamos Maniac (1952), the University of Illinois Illiac (1952), the Ordvac (1952), the MIT Whirlwind (1953), the Rand Corporation
Johnniac (1954), the Stockholm Besk (1954), and the Amsterdam Mathematical Centre Arra (1954). The Manchester Ferranti Mark 1 (1951) and Univac I (1951), although influenced by the IAS, are nevertheless pioneering efforts of their own; we place them in the Pioneer House (Chapter 10).
Unfortunately all IAS-type machines were different, so programs were not interchangeable.
Noteworthy
Stored program. By uniting the instruction and the data store into one memory, the IAS solved the most bothersome programming problem of early Computers: addressing elements within an array. But this solution was almost immediately made obsolete by Kilburn’s invention of address modification through indexing. The basic ability to treat instructions as data, however remained; it cleared the way for programming languages.
Consistency. The IAS design excels through consistency. To be sure, some of this consistency is the result of omitting necessary functions, such as logic and I/O, as well as of ignoring bit efficiency. Nevertheless, the design is
strikingly clean and thorough. Furthermore, this consistency is not accidental: Von Neumann explicitly emphasized it.
In a sense, the IAS design also exemplifies the difficulties encountered in achieving consistency: the allocation of incommensurate objects and pressures from the implementation. (See the following comments on rounding.)
Incommensurate units. The key architectural problem of the von Neumann design is the incommensurateness of its data and instructions - exactly the two items one would like to treat alike in the stored-program design. The IAS design solves this problem by using 40 bits for all data and 20 bits for all instructions, with addressing resolution always 40 bits. The cost of this solution is a Left and Right version of Branch and of Store Address. Most other members of the house choose to gain bit efficiency by having halfword data - at the expense of complication.
Load. A small example of clean design is the Load instruction. We might overlook it as obvious, but no design tradition had been established at this point; the EDSAC is an example of a less clean alternative: lt has no Load, but combines clearing with storing.
Peculiarities
No logic. Having no logic also means having no logical shift. Logic emerges only slowly in the von Neumann house (another contrast with Kilburn’s Manchester design). The arithmetic shifts are accurately called Multiply By Two and Divide By Two, deemphasizing their shifting property.
No integers. All numbers are treated as fractions; even such obvious integers as addresses are left-aligned in the instruction.
No multiple lengths. Product and dividend are single-length data - a defensible decision.
Rounding. As an outstanding numerical analyst, von Neumann was fully aware of the rounding problem, and he treated it thoroughly. The product is a rounded high-order fraction with the remainder of the product placed in the MQ.
For the quotient, however von Neumann used “making odd" to simplify the implementation. He said that “this occasionally produces results with unfamiliar and even annoying aspects (e.g., when quotients like 0 / y or y / y are formed), but it is nevertheless unobjectionable and self-consistent on the basis of our general principles" [Burks et al., 1946].
