Zuse Z3: Program Control, Relay Arithmetic, and the Limits of ‘First’

The Zuse Z3 is often compressed into one sentence: “the first computer.” That shorthand hides the useful engineering questions. What did it compute, where did it keep data, how did an operator change its program, and what exactly is meant by “programmable”? The best answer is not a leaderboard. It is a bounded reconstruction of a particular electromechanical architecture, with each claim tied to the kind of evidence that supports it.

The Z3 was completed in Berlin in 1941 by Konrad Zuse. The Zuse Institute Berlin’s computer museum describes a working relay machine with a floating-point arithmetic unit and a memory of 64 words, each 22 bits wide. It attributes 600 relays to the arithmetic unit and 1,400 to memory. The Deutsches Museum describes the original as destroyed during the war and identifies its present display as a functional reconstruction authorized by Zuse. Those facts establish an important milestone, but not a universal answer to every “first computer” question: other machines qualify under different requirements, such as electronic switching, general-purpose control, or a stored program.

Separate instruction medium, data store, and arithmetic unit

The most clarifying architectural distinction is between the program and the numbers it operates on. In the Z3, the program was supplied on punched tape and read sequentially; the 64-word relay memory was not a modern unified store holding both arbitrary instructions and data. A 1997 technical history by Raúl Rojas, based on a close evaluation of Zuse’s 1941 patent application and a software simulation of the logic, describes the tape instructions as eight-bit rows and the memory as a 64-word address space. Those instruction-format details are attributed to Rojas’s technical analysis rather than inferred from a museum exhibit.

That distinction changes how “program in memory” should be used in a comparison. A punched-tape sequence was still a way to alter the machine’s operations without rewiring its arithmetic circuits. But it was external instruction storage, unlike the later arrangement in which instruction words and data reside together in addressable electronic memory. Describing the Z3 as program-controlled is well supported; silently equating that phrase with a modern stored-program architecture is not.

The arithmetic machine was not a general-purpose integer ALU. The Zuse Institute lists addition, subtraction, multiplication, division, and square root. Rojas analyzes the Z3 as a binary floating-point machine and describes a small instruction set divided among input/output, memory, and arithmetic operations. This made it useful for numerical work while imposing a sharply defined computational model. The design’s compact 64-word memory also made address width a real architectural limit, not an abstract footnote.

Reading the tape as control, not as a modern executable

A punched tape is a physical sequence. In Rojas’s analysis, each instruction occupies an eight-bit row; memory instructions use six lower bits for a word address, consistent with a 64-location store. Arithmetic operations are carried out by the machine’s relay-based datapath and control sequencing. The tape reader advances through the sequence rather than fetching an instruction from an addressable program counter in memory.

This model is easy to misunderstand when a modern emulator or diagram turns the tape into an array of byte values. Such a representation is convenient for software, but it can erase the original separation between the operator’s program medium and the machine’s data storage. It can also encourage readers to infer features the interface does not establish, such as arbitrary conditional control flow. A careful account should identify the source of each layer: the archive record for documentary provenance, the museum description for high-level machine characteristics, and a technical reconstruction for instruction decoding and cycle behavior.

The “no conditional branch” statement needs the same care. Rojas’s 1997 architecture paper says the documented instruction set lacks conditional branching and describes a loop made by physically joining the ends of the punched tape. Its final discussion considers how an external interpreter changes the abstract system being analyzed; that is not yet the Z3-specific simulation proof. In a separate paper, dated September 5, 1997 and published in IEEE Annals of the History of Computing in 1998, Rojas argues that a fixed arithmetic loop on the Z3 can simulate a Turing machine with bounded tape by implementing conditional selection and indirect addressing arithmetically. The construction requires a punched-tape program whose size grows to cover the modeled memory, and Rojas distinguishes this theoretical result from practical Z3 programming. Thus “no conditional-branch instruction” describes the machine’s direct instruction model; “universal for bounded-memory computation” describes a deliberately expanded abstract interpretation, not an ordinary feature of its program tape.

Relays make the control path visible

The relay count is not merely a spectacle. The Zuse Institute’s summary assigns the 600-relay arithmetic unit and 1,400-relay memory to different functions. That split illustrates how the design budget was spent: storing values in a relatively large relay array required substantially more switching elements than the arithmetic unit in the museum’s account. A modern diagram can make the processor, memory, and control look like clean blocks, but in the original machine these were physical electrical and mechanical assemblies whose timing and wiring constrained the usable instruction sequence.

Do not merge the figures into a precise total for every subsystem. Rojas describes a 600-relay processor and says the memory used three times as many relays, while the Zuse Institute museum page reports 1,400 relays for memory. The sources cited here do not explain that discrepancy. Keep each count attributed to its source instead of inferring an unreported subsystem or presenting a single reconciled total.

Reconstruction is evidence, but not the 1941 original

The original Z3 was destroyed during the war. The Deutsches Museum says its current functioning exhibit is a Zuse-authorized reconstruction. This is valuable: it can show the intended design in motion and make an otherwise lost machine tangible. It is still a reconstruction, not an untouched surviving 1941 artifact. When a demonstration is used to support a claim, distinguish what the reconstruction demonstrates from what is documented by original plans, notes, patents, or contemporary records.

The Deutsches Museum research project also explains why provenance matters. It describes the archival expansion and critical reassessment of Zuse’s legacy, including the need to examine the completeness, authenticity, and dating of documents rather than relying only on retrospective autobiography. The instruction-format analysis here is attributed to Rojas’s published technical work and author manuscript; it is not inferred from an inaccessible catalog record or from the reconstruction alone.

A practical checklist for comparing early machines

When a source calls the Z3 “the first” something, ask which requirement is being measured:

Working machine: Was the design completed and demonstrated as a functioning device?
Programmability: Could the operator change an operation sequence without rebuilding the logic? If yes, how was that sequence represented and supplied?
Automatic operation: Which parts of input, execution, and output proceeded without human intervention, and where did an operator still pause or restart the machine?
Technology: Was the switching electronic, electromechanical, or mechanical?
Program representation: Were instructions external, internally stored, or physically configured?
Control flow: Did the instruction set provide conditional branching, or only a fixed sequence and other forms of repetition?
Evidence quality: Is the claim derived from a contemporary artifact, a later reconstruction, an institutional summary, or a historian’s analysis?

These questions make the Z3 more interesting, not less. It was a working relay-based digital calculator with a separated arithmetic unit and data memory, floating-point operations, and externally sequenced program control. The surviving historical picture combines archival records, a machine reconstruction, museum summaries, and later technical analysis. Those sources do not all answer the same question. Keeping their roles separate is the difference between a careful account of an important design and an unsupported “first computer” slogan.

Related:

Sources: