1957: Lower the unit of access

DEC and Fairchild were both founded in 1957, but they changed different boundaries in computing. DEC's smaller computers enabled laboratories and engineers to obtain more direct control over a machine. Fairchild's semiconductor work helped develop manufacturing methods and components from which other products could be built. The comparison is about access and learning, not a claim that minicomputers and integrated circuits were the same invention or became mature in one year.

A centralized computer service can have excellent utilization yet impose long turnaround on an experimenter. Suppose a researcher submits a job, waits four hours, studies the result, and changes a parameter. Four iterations consume at least sixteen hours of waiting. A dedicated laboratory machine that returns each result in ten minutes permits the same four iterations within forty minutes of waiting, even if its raw arithmetic rate is lower. The value comes from shortening the human feedback loop, not winning every benchmark.

Direct access also creates obligations. Someone must interface instruments, write control software, handle electrical timing, and recover from failure. DEC's handbook is useful primary evidence because it describes an actual architecture and interfaces, rather than only saying that a machine was accessible. Ask which operations a user could program, what memory constrained the experiment, and how external devices exchanged signals with the processor.

Semiconductor manufacturing exposes another learning loop. A design can be logically correct while production yield is poor. If a batch produces 100 candidate devices and only 40 meet specification, each saleable device carries the cost of 2.5 candidates before other expenses. Improving yield to 80% halves that candidate cost per good device. Manufacturing measurements, process control, packaging, and reliability testing therefore belong inside the invention story.

Noyce's device-and-lead patent documents a concrete interconnection approach. It should be read as a legal and technical claim with a filing history, not as proof that one person independently supplied every element of the integrated circuit. Semiconductor progress involved several inventors, processes, organizations, and competing implementations. Spinouts can carry knowledge and experience, but they also depend on capital, suppliers, customer demand, and enforceable agreements.

A lower purchase threshold can enlarge a market because new users can justify experiments that a large centralized procurement would reject. It can also shift work onto those users. Compare total task cost: equipment, engineering time, waiting, maintenance, and the value of learning sooner. A cheaper machine that requires an unavailable specialist may not be cheaper for a particular laboratory.

The institutional mechanism is a change in who can run an experiment and how quickly its result reaches the builder. Evaluate that mechanism with a named workload and constraints. Avoid treating small size, low price, or a famous spinout network as sufficient explanations by themselves.

Reconstruction exercise

A laboratory needs six iterations. A shared machine executes each in one minute but adds a two-hour queue; a local machine executes each in fifteen minutes without a queue. Calculate elapsed machine-and-wait time for sequential iterations. Then compare manufacturing yield of 25% and 50% for a batch cost of 1000 units producing 100 candidate chips. Explain which omitted cost could reverse each apparent advantage.

Worked solution

The shared path takes 6×121=726 minutes; the local path takes 6×15=90 minutes, assuming each result must be examined before the next submission and human analysis time is equal. The local machine improves turnaround despite slower execution. At 25% yield,25 good chips share the 1000-unit batch cost, or 40 units each; at 50%,50 good chips cost 20 each before packaging and other expenses. A local machine’s purchase, programming, maintenance, or instrument-interface cost can outweigh saved waiting. A higher-yield process can require more expensive equipment, processing steps, or lower-performing devices. The comparison needs total costs and an explicit performance specification, not one favorable ratio.

Test the limits and transfer the idea

1. Scheduling calculation · allow 25 minutes

Choose by the deadline rather than arithmetic speed

An instrument experiment needs five dependent iterations. A shared machine adds a 40-minute queue and computes for 2 minutes each time. A local machine needs a one-time 30-minute calibration and 12 minutes per iteration. Human analysis takes 6 minutes after every iteration on either route. Calculate completion times and whether either meets a 120-minute deadline. Then recalculate if the five jobs are independent, all shared-machine jobs enter one queue together, and are executed consecutively before analysis. State your concurrency assumptions.

Get a starting hint

Sequential dependence repeats the queue delay; independent submission need not.

Compare your reasoning with the worked solution

Dependent shared work takes 5×(40+2+6)=240 minutes. Local work takes 30+5×(12+6)=120, exactly meeting the deadline. For the stipulated independent shared batch, one 40-minute wait, ten minutes of execution and thirty minutes of subsequent analysis total 80 minutes. That route now wins under the specified non-overlap assumption. If analysis overlaps computation, the schedule could be shorter; if each job has a separate queue delay, it could be longer. The DEC handbook is a primary architecture/interface artifact, not a measurement of these invented queues. The example demonstrates that control over a feedback loop can matter more than raw throughput, while also showing why the conclusion changes when dependencies change.

Check your reasoning

• Obtains 240, 120 and 80 minutes under stated assumptions.
• Explains why dependency structure changes the queue term.
• Avoids treating hypothetical timing as historical machine performance.

2. Yield and unit-cost model · allow 25 minutes

Follow a device through independent yield stages

A fictional fabrication run makes 200 candidate dies for 8000 units. Eighty percent pass wafer test. Packaging costs 10 per attempted package; 90% of packaged dies pass final test. Only wafer-test survivors are packaged. Compute expected shipped devices and cost per shipped device. A process improvement raises wafer yield to 90% but costs another 800 per run. Compare unit costs. Explain why Noyce's patent alone cannot establish either production yield or the profitability of the improvement.

Get a starting hint

Packaging cost depends on attempted packages, not only successful final units.

Compare your reasoning with the worked solution

Initially 160 dies reach packaging, costing 1600. Expected shipments are 160×0.9=144, so unit cost is 9600/144≈66.67. With improvement, 180 are packaged for 1800 and 162 ship. Total cost is 8000+800+1800=10600; unit cost is 10600/162≈65.43. Better yield modestly lowers expected unit cost despite increased run expense. This uses constant conditional final yield and ignores variability, rework, testing cost and capacity constraints. A patent is primary evidence of a disclosed technical claim and its date; it does not supply measured yield distributions or customer willingness to pay. Production records and controlled process comparisons are required to test the cost model.

Check your reasoning

• Counts packaging attempts and final survivors separately.
• Computes both total and per-shipped-device costs.
• States the difference between patent disclosure and manufacturing evidence.

Check your understanding

A slower dedicated machine enables more experiment iterations per day than a fast shared one. Which mechanism explains this?

• Shared computers cannot execute scientific programs.
• Reduced queueing and a shorter user feedback loop.
• Dedicated machines automatically have better numerical accuracy.
• A lower clock speed always increases throughput.

Explanation

Task turnaround includes waiting and revision, so raw execution speed alone is insufficient.

A batch costs 1000 units for 100 candidates. Yield doubles from 25% to 50%. What happens to candidate-production cost per good device?

• It falls from25 to12.5 units.
• It remains10 because there are100 candidates.
• It falls from40 to20 units before other costs.
• It necessarily halves the complete delivered product price.

Explanation

Good units rise from 25 to 50; packaging, equipment, performance, and other costs remain separate.

Sources and further reading

• Small Computer Handbook, 1967: Digital Equipment Corporation, archived by Bitsavers
• Semiconductor device-and-lead structure, US 2981877A: Robert Noyce, United States patent