John Atanasoff

John Vincent Atanasoff

Physicist and computing pioneer, 1903–1995. Study focus: linear-equation elimination, regenerative storage, and the scope of invention claims.

Atanasoff and graduate student Clifford Berry developed the Atanasoff-Berry Computer at Iowa State College between 1939 and 1942. The ABC is important for combining electronic binary arithmetic with capacitor-based regenerative storage in a machine designed for linear equations. It was not a general-purpose stored-program computer. Its significance does not depend on pretending those categories are interchangeable.

The problem behind the machine

Atanasoff studied electrical engineering at the University of Florida, mathematics at Iowa State, and physics at Wisconsin, completing his doctorate in 1930. Numerical work motivated his search for more effective calculation. His later account describes experimentation with alternative representations and a working addition/subtraction prototype. It is valuable first-person evidence, but recollections and claims of absolute priority must be evaluated separately. Atanasoff's account and biography

Berry was an essential collaborator in turning the design into hardware. The surviving historical accounts describe rotating storage drums, electronic arithmetic circuits, and an intermediate recording system. The Computer History Museum identifies the machine as a nonprogrammable linear-equation solver and documents 30 add/subtract circuits and drums storing 30 numbers each. The existence of parallel arithmetic channels is compatible with processing successive bits of each number serially. Berry biography and correspondence, ABC artifacts

The project ended when its builders left for wartime work in 1942. Descriptions of its intended equation capacity should not be confused with evidence that the original machine was a dependable production system for every such case. Intermediate recording was a practical weakness. A later reconstruction can demonstrate that a reconstructed design works; it cannot retroactively prove every historical claim about the original machine's reliability.

Worked mechanism: eliminate an unknown

The following is an original mathematical teaching example, not an instruction trace from the ABC. Consider:

2x + y = 5
 x + y = 3

Represent each equation as a row of coefficients including its right-hand side: [2,1|5] and [1,1|3]. Subtract the second row from the first to get [1,0|2], so x=2. Substituting in the second equation gives y=1. The key is that the same row operation acts on every coefficient and on the right-hand side. Subtracting only the left-hand coefficients would silently change the problem.

For a less convenient pair, [3,2|12] and [2,1|7], eliminate x by forming twice the first row minus three times the second. This yields [0,1|3], so y=3; then 2x+3=7 gives x=2. Exact elimination maintains the solution set when its transformations are reversible. A zero pivot requires choosing another row or recognizing dependence or inconsistency.

In a physical machine, the mathematical transformation must be decomposed into available operations. Coefficients need a representation, arithmetic must handle carries and signs, and intermediate rows must survive until the next stage. A fast adder alone is not a complete equation-solving system. This is the engineering lesson the ABC makes unusually visible.

Binary does not mean two perfect voltages

For unsigned binary addition, each position combines input bits a, b, and a carry c. The output bit is their sum modulo two; the outgoing carry is one when at least two inputs are one. In the original example 0111 + 0001, the lowest position computes 1+1=2, writes zero, and carries one. That carry propagates through the next two ones. The final result is 1000.

Circuit designers implement ranges of voltage as logical states and leave margins between them. Actual signals take time to change and are subject to noise. Digital reasoning works because circuits restore and discriminate those states within specified tolerances, not because electricity literally has only two possible values.

Decimal arithmetic does not require ten voltage levels. A decimal digit can use multiple binary-valued devices, as historical decimal computers did. Therefore the ABC's binary choice should be explained as a representation and engineering decision, not as proof that all electronic computation must be binary. Nor does binary representation make all real numbers exactly representable in finite memory.

Regeneration is a scheduled computation too

A capacitor stores charge, but leakage changes it. Regeneration senses the stored state and restores it before the signal becomes ambiguous. The ABC used capacitors on rotating drums; the movement repeatedly brought storage locations into contact with sensing and restoring circuitry. Modern DRAM also requires refresh, although its structure and control are different. Similarity of principle alone does not establish a direct historical lineage. ABC storage and operational descriptions

Imagine a constructed store whose worst-case readable lifetime is 8 milliseconds. A scan revisits every cell every 5 milliseconds, but an arithmetic operation can delay scanning by 4 milliseconds. A worst-case cell may wait 9 milliseconds, exceeding the 8-millisecond bound. Average scan time is insufficient to guarantee correctness. Either the controller must constrain interruptions, refresh earlier, or provide a different storage design. These numbers illustrate scheduling and retention; they are not ABC specifications.

A patent judgment is a particular kind of evidence

John Mauchly visited Atanasoff in 1941. In the later Honeywell v. Sperry Rand dispute, a federal court invalidated the ENIAC patent and made findings concerning derivation from Atanasoff's work. The judgment also addressed legal grounds such as public use. The court's conclusion is evidence of what the court decided under the issues and record before it. It is not a definition that makes the ABC general-purpose, nor proof that every later computer descended from it. Museum account of the dispute, judgment

Iowa State's archive preserves design manuscripts, Berry's work, and litigation materials. Read those document types for different questions: a manuscript describes a design; correspondence can document transmission; a reconstruction tests an interpretation; a judgment determines legal issues. The unsupported all-computers genealogy and tavern-as-certain-invention-moment narrative are unnecessary. Archive inventory

Exercises

1. Solve 3x+2y=12 and 2x+y=7 using elimination, then verify both original equations. What would 0x+0y=1 mean if it appeared during elimination?
2. Trace unsigned four-bit addition 1110+0011, including the carry beyond the stored word. Explain the difference between wraparound and detecting overflow.
3. A museum reconstruction completes a 20-equation run successfully. A newspaper says this proves the original machine ran all 29-equation problems reliably in 1941 and therefore was general-purpose. Identify three unsupported inferences and the evidence each would require.

Compare your reasoning with the worked solutions

1. Twice the first row minus three times the second gives y=3. Substitute into the second: 2x+3=7, hence x=2. Check: 3(2)+2(3)=12 and 2(2)+3=7. A row 0=1 is inconsistent, so the original system has no solution if the row operations and arithmetic were correct. One must distinguish an inconsistent problem from a hardware or transcription error.
2. From the right: 0+1=1; 1+1=2 produces zero and carry one; 1+0+1=2 again produces zero and carry one; the final position does the same. The mathematical result is 10001, or 17. A four-bit store keeps 0001 under modular wraparound. The outgoing carry can separately flag unsigned overflow. Discarding it does not make 17 equal to 1 as an ordinary integer.
3. A present reconstruction is not direct observation of a 1941 run; contemporary logs or reliable reports would be needed. Success on one 20-equation input does not establish reliability across every 29-equation case; tests, failure data, and numerical conditions matter. Solving a fixed class of problems does not establish general-purpose programmability; that requires an account of the machine's configurable instruction/control capabilities. These are three different claims, not one stronger version of the same claim.