I've been reading Achieving Accuracy: A Legacy of Computers and Missiles by M. W. McMurran, who helped develop the Autonetics D-17 guidance computer for the Minuteman ICBM in the early 1960s. For all its specialised features, the D-17 has at least a family resemblance to general-purpose magnetic drum computers like the IBM 650 (after the Minuteman I was retired, some surplus D-17s made their way to universities and bitsavers.org has a good collection of documents from them).
However, McMurran suggests that the D-17 was a break from Autonetics tradition, and its design was in part reaction to a competing proposal from TRW Space Technology Laboratories:
The STL "Blue Book" computer was to be a "compulsory-optimum" programmed whole-value machine. [...] Autonetics computer designers had virtually no experience with whole-value general-purpose real-time systems. But [overnight we] blocked out a general-purpose alternative to Blue Book. By then, we had developed very serious concerns about the speed of any emerging serial whole-value computer [but those concerns were addressed by features of the D-17 design]. (p. 276)
The part I'm struggling with here is what is meant by whole-value computation, and why was it a problem for real-time systems?
One possible answer comes from the fact that the first digital guidance systems were modelled after analogue computers, but were built around digital differential analysers (DDAs, not to be confused with the graphics algorithm) which were hard-wired elements for solving discretised differential equations at high speed. McMurran describes two Autonetics DDA guidance computers that preceded the D-17, NATDAN (the North American Transistorized Digital ANalyzer, developed for the Navaho cruise missile), and Verdan (the VERsatile Digital ANalyzer):
[NATDAN] was rate-limited, but could solve an amazing array of complex differential equations one pulse at a time, as long as the solution moved at a rate NATDAN could handle. [...] Since NATDAN had no whole value section, communication with guidance and control systems was a lot easier for it than communicating with people. (pp. 138-139)
The [Verdan] design started with the NADAN (sic) DDA. By 1955, it had become clear that even embedded control computers frequently needed to communicate with whole-value based systems, often with people in the loop. (pp. 142-143)
Because DDAs dealt with rates of change discretised into relatively small increments (e.g. integrating accelerometer output to compute vehicle velocity and hence position), I'm guessing this made them faster in a serial computer than dealing with full-precision variables representing physical quantities across a range of scales. However, I suspect that's too simplistic an answer: is the phrase "whole value" more than just a synonym for "general purpose" computation in this context?
EDIT: Since posting I've found a couple of early papers on DDAs that give insight into the digital techniques used in systems like NATDAN and Verdan:
- R. E. Sprague, Fundamental Concepts of the Digital Differential Analyzer Method of Computation. Mathematical Tables and Other Aids to Computation v6n37 (Jan. 1952), pp. 41-49
- John F. Donan, The Serial-Memory Digital Differential Analyzer. Mathematical Tables and Other Aids to Computation v6n38 (Apr. 1952), pp. 102-112