Frederic C. Williams

Frederic Calland Williams

Electrical engineer · 1911–1977 · United Kingdom

Williams and Tom Kilburn developed a practical cathode-ray-tube store for early computers. The Manchester Baby tested whether this store could remain reliable while a running program repeatedly fetched instructions and changed data. The important achievement was a working system with timing, sensing and regeneration, not simply the observation that a screen could retain charge.

From radar circuits to a computer store

Williams studied engineering at Manchester, earning degrees in 1932 and 1933, and completed a DPhil at Oxford in 1936 on circuit and valve noise. He joined British radar research in 1939. The research group became TRE in 1940 and moved to Malvern in 1942, so a claim that he joined “TRE at Malvern in 1939” combines later names and locations. In 1976 he was knighted; the 1957 knighthood sometimes stated in summaries is wrong. University of Manchester biography.

After visits to American laboratories, Williams investigated CRT storage at TRE in 1946. A single-bit demonstration preceded his move to Manchester that December. Kilburn developed improved storage methods in 1947, with assistance including Geoff Tootill. By late 1947 they had maintained 2,048 bits for hours. Most of the resulting patents named both Williams and Kilburn. The institutional history explicitly recognizes Kilburn's substantial experimental and conceptual contribution. Williams-Kilburn store history.

How dot and dash storage works

A CRT steers an electron beam to locations on an insulating screen. Bombardment causes secondary emission: arriving electrons dislodge other electrons. Different beam-writing patterns leave different local charge distributions. In the dot-dash method described by Manchester, a short dot and a longer dash produce distinguishable responses on a later interrogation.

The pickup plate outside the screen senses a transient coupled to a change in charge. It is not an optical camera reading brightness, and it does not read a permanent static charge map directly. Reapplying the beam changes the stored pattern; the sensing result controls the writing pulse so that the intended bit is restored. Repeated regeneration compensates for charge leakage. Beam positioning, interference and timing all affect reliability. Technical explanation.

This is only one historical encoding method. “Charged means one, uncharged means zero” is an inadequate description of every Williams-Kilburn store. Likewise, the glow of the screen is not the stored data's essential physical basis. A separate display tube could show information while the store's face was covered by its pickup plate and shielding.

The comparison with semiconductor DRAM is useful at the functional level: both need periodic maintenance of a charge-based state. The physical structures and sensing circuits differ. Core memory and static RAM use still other storage mechanisms. Similar requirements do not establish that all memory is a direct technological descendant of one tube. CHM storage history.

Worked mechanism: memory maintenance consumes time

The Baby's main store held 32 words of 32 bits, a total of 1,024 bits. That equals 128 eight-bit units of capacity, but does not mean the machine had byte addressing. Its instruction cycle interleaved main-store access with refresh. The Manchester history describes a four-beat rhythm with refresh on two beats, instruction fetch on another, and the required operand access on the last.

Consider an original timing model in which each beat takes 300 microseconds and each refresh beat restores one successive word. An instruction takes 1.2 milliseconds. Two words are refreshed per instruction, so a 32-word refresh sweep takes 16 instructions, or 19.2 milliseconds. This is a calculation under stated timing assumptions, not a universal specification of every tube-based computer.

Now suppose a proposed I/O operation blocks all refresh for 250 milliseconds and the store's safe retention deadline is 100 milliseconds. The operation risks losing data even if no program instruction requests memory. A store's apparent logical inactivity does not imply that its physical maintenance can stop. Possible remedies include independent refresh control, shorter I/O intervals, or a storage technology with different retention behavior.

What the Baby demonstrated

On 21 June 1948 the Baby successfully executed a program from its electronic read-write store. This is the standard milestone for such stored-program execution. It was designed as an experimental test, with Williams, Kilburn and Tootill playing central roles. The program searched for a number's highest proper factor. Manchester's detailed history distinguishes the first small test input from a run on 2^18 within the following days, which took about 52 minutes. Collapsing both into one precisely described first run gives more certainty than the source supports. Manchester Baby history.

A stored program does not logically require random-access memory. Delay-line machines also executed stored instructions, with scheduling designed around serial circulation. Nor does the Baby's success prove that all later engineering was straightforward. Capacity, error rates, instruction design, I/O and maintainability still required substantial work.

Exercises

1. In the timing model above, double main memory to 64 words without changing the number of refresh beats. Find the refresh interval. A proposed change then refreshes only one word per instruction: calculate the new interval and compare it with a 60-millisecond retention deadline.
2. A circular delay line presents 32 word positions, one every 10 microseconds. Assume a read request arrives exactly at a word boundary and the requested position is equally likely to be any of the 32. Compute the mean wait before that word begins. Explain why sequential instruction placement can perform better than independent random requests.
3. An exhibit says “Williams invented RAM alone; its glowing dots were read without changing them; the first run factored 2^18 in 52 minutes on 21 June.” Identify the credit, mechanism and chronology problems. Rewrite it in three defensible sentences.

Compare your reasoning with the worked solutions

1. At two words per 1.2-millisecond instruction, 64 words need 32 instructions, or 38.4 milliseconds. With one word refreshed per instruction, the sweep needs 64 instructions, or 76.8 milliseconds. The latter exceeds the stipulated 60-millisecond deadline even before scheduling jitter or failures. Increasing capacity requires reconsidering maintenance bandwidth rather than assuming that addressability alone guarantees correctness.

2. Waits are 0,10,20,...,310 microseconds. Their mean is (0+310)/2 = 155 microseconds. This excludes the time to transmit the requested word. A sequential program can place its next instruction near the point where it will be needed and reduce waiting; a stream of unrelated addresses cannot rely on that arrangement. Serial access changes scheduling costs, not the possibility of storing a program.

3. Williams and Kilburn jointly developed the store, with other contributors to the computer. A pickup plate detected electrical transients during beam interrogation and the system regenerated the charge pattern. The first successful Baby program ran on 21 June 1948; the detailed Manchester account places the 2^18 test within the following days. These sentences preserve the achievement while avoiding unsupported exclusivity, optical sensing and a conflated timeline.