A bit of history: how computers remembered

Ferrite cores

The first computers needed memory that could be written quickly, read quickly and wouldn't forget when switched off. In 1953 at the Massachusetts Institute of Technology the Whirlwind computer got magnetic-core memory made of ferrite cores (Jay Forrester's group), and for the next twenty years almost every computer remembered this way.

A ferrite ring the size of a poppy seed can be magnetised clockwise or counterclockwise — that is the bit.

A ferrite ring is one bit: magnetised clockwise or counter-clockwise. Wire X, wire Y and the sense wire are threaded through the ring's hole. X and Y each carry half the current needed to flip it — only the ring where both half-currents meet flips. If the ring flipped, a pulse jumps through the sense wire.

To pick one core out of thousands, coincident current was invented: each wire carries only half of the current needed to flip the magnetisation. A core flips only where two wires meet.

A memory plane: rings at wire crossings. An X wire threads every ring of its row, a Y wire every ring of its column. The address picks one X and one Y — only the ring at their crossing flips. A 64×64 plane holds 4096 bits. Such planes were woven by hand: thin wires threaded through rings the size of a poppy seed.

Reading destroyed the bit: to find out what was in a core, the computer tried to write it to 0 — if a pulse appeared in the read wire, there had been a 1, and it was immediately written back. On the other hand, the cores forgot nothing when switched off.

Memory planes were woven by hand: thin wires were threaded through the cores under a microscope. Thousands of hours of painstaking work for every kilobyte.

Hand-woven ROM

The Apollo Guidance Computer, which steered the spacecraft to the Moon in 1969, also had read-only memory — also made of cores, but built differently:

A ROM woven by hand: the program memory of the Apollo Guidance Computer. A wire threaded through a ring is a 1, a wire routed around it is a 0. The program was not written but woven in at the factory; changing it meant re-weaving.

The program wasn't written in, it was woven in at the factory. A bug in the program was fixed by re-weaving.

Memory on a chip

In 1969 the young company Intel releases its first memory chips: the 3101 with 64 bits and the 1101 with 256 bits — static memory (SRAM): each bit is held by transistors as long as there is power. In 1970 the Intel 1103 comes out — dynamic memory (DRAM) of 1024 bits. A bit in it is the charge of a tiny capacitor; the charge leaks away, so the memory is constantly "recharged" by reading and rewriting it. But the cell is tiny, and within a couple of years the 1103 became the best-selling chip in the world — ferrite cores were gone.

The cell of one transistor and one capacitor was invented by Robert Dennard of IBM (1966). Almost all the main memory of computers and phones is still built this way.

Why our cell is so big

How a bit is stored. A ferrite ring — by its magnetisation. Static memory (SRAM) — two inverters holding each other plus two switches: 6 transistors. Dynamic memory (DRAM) — by the charge of a capacitor behind one transistor; the charge leaks, so it is constantly refreshed. Our cell is a latch made of first-kit gates plus a "tap": 22 transistors.

In real static memory a bit is 6 transistors: two inverters hold each other, and two switches connect them to a shared data wire. To write, the outside circuit simply "overpowers" the cell. Our simulator can't do that: its switches are ideal, a transistor has no "strength", and two parts pulling a wire in different directions produce not a winner but a conflict. There is no slowly discharging capacitor in the simulator either.

So our cell is a latch made of the first kit's gates plus an AND "tap": 22 transistors instead of 6. It honestly stores a bit, works in every simulator mode and is built from parts you already know. And we'll replace the shared data wire with OR gates — why, you'll see in lesson 2.

And in the Game & Watch?

The Sharp SM5 microcontroller inside the Game & Watch also has static memory, and it sits right on the chip, next to the processor: a few dozen bytes of 4-bit cells. Some of the cells are screen segments: write a bit, and a segment lights up. Our machine's screen memory works exactly the same way.