Konrad Zuse

Konrad Zuse

Computer designer and engineer · 1910–1995 · Germany

Zuse developed binary calculating machines and the Plankalkül programming-language design. His work is particularly useful for separating four properties often collapsed into “a computer”: digital representation, automatic instruction sequencing, stored programs, and universal computation.

Machines, collaborators and historical setting

The mechanical Z1 was completed in the late 1930s but had serious reliability problems. The Z3, demonstrated in 1941, used relays and program instructions on punched film. The Z4 survived the war and entered service at ETH Zurich in 1950. A history by ETH mathematician Martin Gutknecht describes its mechanical store as 64 words of 32 bits, not the 512-word figure sometimes repeated in summaries. Z3 archive overview, ETH history.

Zuse's achievements were not work in complete isolation. Friends and collaborators helped build machines; firms supplied components; wartime institutions supported projects. His S1 and S2 machines served Henschel's military aircraft and guided-bomb work. Explaining these connections is more informative than either portraying a lone inventor without predecessors or declaring his institutional setting irrelevant. Surviving documents and later reconstructions also need different evidential labels. A replica can demonstrate a reconstructed mechanism without being the original 1941 artifact. Berlin museum exhibition, Deutsches Museum research project.

Program control and its limits

A relay opens or closes an electrical connection under electromagnetic control. Combinations of such components can implement binary storage and arithmetic, while mechanical motion limits switching speed. The Z3's tape specified a sequence of floating-point operations. It had no ordinary conditional branch instruction. Reading instructions from external tape is program control, but it is not the same arrangement as fetching changeable instructions from the machine's working memory.

In 1998 Raúl Rojas showed how a greatly expanded loop of arithmetic operations could simulate a Turing machine whose tape has a specified finite bound. Arithmetic selection can replace explicit branches. This is a mathematical construction, not evidence that Zuse's customers ran such a simulator or that the finite original machine held an unbounded tape. Program size and available storage remain central limitations. Rojas's paper, author-uploaded text.

An original toy example isolates one idea. If c is exactly either 0 or 1, the expression c·a + (1−c)·b selects a when c = 1 and b when c = 0. For a = 7 and b = 3, the two results are 7 and 3. No branch appears in the expression. This does not by itself implement a whole computer: the condition still must be computed, all required state must be represented, and evaluating unused alternatives must not cause invalid operations. Floating-point rounding and exceptional values can also invalidate algebraic shortcuts.

Floating point: range is not exactness

Binary floating point represents a nonzero number using a sign, a significand and a power of two. This general organization also appears in later standards, but similarity does not establish that IEEE 754 is the Z3 format or that the standard descended directly from it.

Use an explicitly invented teaching format with three significant binary digits, an adequately wide exponent, and round-to-nearest with ties to an even trailing digit. At magnitude 1, adjacent values include 1.00₂ = 1 and 1.01₂ = 1.25. The exact sum 1 + 0.125 lies halfway between them, so it rounds to 1.00₂. A nonzero addend disappears from the stored result. At magnitude 8, adjacent values are 1.00₂×2³ = 8 and 1.01₂×2³ = 10: the gap is now two. The same digit budget gives wider absolute spacing at larger magnitudes.

This example is not a reconstruction of Zuse's rounding hardware. It teaches why a format's bit layout, normalization rules, rounding, overflow handling and exceptional values all matter. A label such as “22-bit floating point” is not a complete numerical specification.

Plankalkül as a language design

Zuse's 1940s work on Plankalkül described data structures and operations at a level above an individual machine's instructions. Bauer and Wössner's 1972 analysis explains Boolean primitives, composite arrays and records, and restrictions on meaningful values. Their article reconstructs the design using later terminology; a modern pseudocode example should not be passed off as original Plankalkül syntax. 1972 scholarly analysis.

A full manuscript edition appeared in 1972, but saying nothing was published or known before then is wrong: the archive records Zuse's earlier article on the general Plankalkül. Nor does chronological anticipation automatically establish influence on FORTRAN or ALGOL. That requires evidence that particular designers encountered and used the earlier work. Earlier article record.

For a modern teaching analogy, an eight-bit field has 256 bit patterns. A type representing a month restricts meaningful values to 1 through 12. Merely reserving eight bits does not enforce that semantic restriction. A useful language design must specify what operations preserve it and how invalid data is handled.

Exercises

1. Evaluate c·a + (1−c)·b for a = 10, b = 4, and c = 0, 1, and 0.25. Why does the third result fail to act like a Boolean choice? Explain why replacing an if-statement whose unused branch divides by zero can be unsafe.
2. In the three-significant-digit teaching format, round 8 + 1 and 10 + 1. Both are halfway cases. Explain why the rounded results differ in direction and why this is not historical evidence about the Z3.
3. A catalog calls an artifact “the original 1941 stored-program universal electronic Z3.” Audit the claims “original,” “stored-program,” “universal,” and “electronic.” What would an accurate label say for a later relay reconstruction? Explain how the Rojas result should be described separately.

Compare your reasoning with the worked solutions

1. The results are 4, 10 and 5.5. Selection requires the invariant c ∈ {0,1}; the third input blends values instead. Moreover, arithmetic masking normally evaluates both a and b first. Multiplying an undefined division by zero does not make the invalid computation legitimate. A correct branch-elimination transformation must control these issues as well as produce the desired final value.

2. The midpoint 9 lies between 8 = 1.00₂×2³ and 10 = 1.01₂×2³. Ties-to-even chooses 8 because its final stored significand bit is zero. The midpoint 11 lies between 10 = 1.01₂×2³ and 12 = 1.10₂×2³, so it chooses 12. Rounding can go downward or upward. These rules were stipulated for a teaching model and cannot establish the behavior of another machine without its arithmetic specification.

3. “Original” requires provenance and fails for a reconstruction. “Stored-program” misdescribes the normal tape-controlled design. “Universal” omits the model assumptions and expanded program used in a later theoretical construction. “Electronic” should be electromechanical for relay logic. A suitable label identifies the reconstruction's actual date and says it represents Zuse's tape-programmed electromechanical Z3, demonstrated in 1941. A separate note can explain Rojas's bounded-tape simulation result and its resource costs without claiming that construction ran historically.