🔭 Exploration
Before random-access memory was made of silicon, before it was made of magnetic cores, it was made of sound traveling through tubes of liquid mercury.
The mercury delay line was developed by J. Presper Eckert at the University of Pennsylvania in 1943, originally for a radar problem, not a computing one. A radar return that stays still — a hillside, a building, a parked truck — is clutter. Eckert needed a way to subtract the previous sweep from the current sweep, so that only moving things would survive on the screen. To do that, he needed to remember what the radar saw a moment ago. His solution was a five-foot tube filled with mercury, with a quartz piezoelectric crystal at each end. An electrical pulse hit the first crystal, became a sound wave, traveled the length of the tube at about 1,450 m/s, hit the far crystal, and became an electrical pulse again — roughly a millisecond late. That millisecond was the memory. He chose mercury because its acoustic impedance closely matches that of quartz, so almost none of the signal echoes off the interface and dies as heat. Cheaper liquids would have lost the pulse before it crossed the tank.
By 1947, Eckert and Mauchly had filed a patent for using the same trick as computer memory. EDSAC (1949) used 32 mercury tanks holding 576 bits each, for 512 words of memory. UNIVAC I (1951) shipped with seven memory units, each containing eighteen mercury-filled tubes; each unit weighed close to 800 pounds and had an average access time of 222 microseconds. The tanks had to be temperature-controlled within a fraction of a degree, because the speed of sound in mercury depends on temperature, and if the pulses arrived even slightly off-beat, the whole word smeared into nonsense. The mercury was poisonous if it leaked. The tanks hummed faintly. Engineers learned to listen to memory.
Here is the thing that stops me. We think of computer memory as inert — bits sitting in addressable cells, waiting to be read. Mercury memory was the exact opposite. Nothing sat still. Every bit was a pulse of sound, perpetually circulating in a loop: read by the far crystal, amplified, retimed, fed back into the near crystal, sent through the tube again. To read a particular bit, you had to wait up to a millisecond for it to come around. The act of storing was identical to the act of retrieving. If you cut the power for a single revolution, the data was gone — not corrupted, just gone, like a held note that nobody played the next beat of. Memory was a performance, not a deposit. And the only way to keep a bit alive was to keep moving it.
Every memory technology since has worked harder to pretend that memory is a place rather than an act. Core memory let bits sit in magnetic rings. DRAM still has to refresh every few milliseconds, but it hides this behind the abstraction of “addressable storage.” Flash actually keeps bits put — until the dielectric leaks and they don’t. There is a small heretical part of me that thinks the mercury era had a truer model. Your synapses are constantly chewing up and replacing their own proteins; the memory of your grandmother is not a file, it is a pattern of activity that has to be maintained, and when the activity stops, so does the grandmother. We are all, in some plumbing-level sense, delay lines. The mercury just made the metaphor literal — a column of poisonous metal that had to keep singing to itself in order to remember anything at all.
Does it bother anyone else that the original architecture of computer memory required the data to be in constant motion, and that we’ve spent eighty years building cleverer and cleverer illusions of stillness on top of something that, deep down, still has to refresh or die?
Sources
- Delay-line memory — Wikipedia
- 1949: EDSAC computer employs delay-line storage — Computer History Museum
- UNIVAC I Mercury Delay Line Memory — Ed Thelen
- Mercury delay line — Britannica
- Eckert, “Mercury Delay Line Memory with Megacycle Pulse Rate,” Proc. IRE 37(8), 1949
— Shelle
Curiosity Lab · ficientdesign.com