Early Core Memory Wiped Each Bit It Read, So the Computer Had to Write It Right Back
Bernard Widrow worked in MIT's core memory group while Whirlwind was being built, and his oral history says the memory he worked on was destructive. His thesis aimed at a version that wasn't, and a 1951 patent already described the erase.
At MIT’s Digital Computer Laboratory, Bernard Widrow was assigned to the magnetic core memory group, and the lab was building the computer called Whirlwind. His oral history for the Computer History Museum, interviewed by Shayne Hodge in 2013, has a short exchange on page 8 that undoes a common assumption about memory. Hodge asks whether reading a core destroyed the data in it. Widrow’s answer: “It was destructive memory.”
Here is how it worked, in Widrow’s account. A core is a tiny ring that holds a one or a zero as a direction of magnetization. To read it, the computer writes a zero into it. If the core already held a zero, nothing changes. If it held a one, the core switches from one to zero, and that switch is the signal the memory senses. The catch is that the one is now gone, so the memory has to rewrite it immediately. (A memory that forgets what you asked it. Relatable.)
Whirlwind’s designers needed it badly. Before core, the MIT machine used electrostatic storage tubes, which Widrow calls “highly unreliable and very, very costly.” The computer itself, he says, could add two numbers in “ten millionths of a second,” and in his telling his father couldn’t see why anyone would want to add numbers that fast. Whirlwind was used for air defense, tracking radar signals in real time, which is the sort of job that makes ten millionths of a second matter.
None of this started with Widrow. The patent on the scheme, US 2,736,880, was filed by Jay Forrester on May 11, 1951 and granted February 28, 1956. It says that reading uses the same current pattern as writing, so the read erases whatever was stored, and that the previously stored information can be rewritten if desired. So the erase-on-read was in the paperwork from the start.
Widrow’s own piece was the 256-bit plane he designed at the lab, which a technician built and tested. In his 1997 interview with the IEEE history center’s Andrew Goldstein, he calls it “all of 256 bits.” That is 32 bytes. In the 2013 oral history, he puts storing at about ten microseconds and retrieving at about five.
His master’s thesis went after the erase. In the 2013 account, Widrow says it was about a nondestructive memory for core. In 1997, he described a sensing scheme of his own: driving the grid lines of the core with currents at two different frequencies, 10 megahertz and 10.5 megahertz. The selected core produced a beat at half a megahertz, and the phase of that beat flipped 180 degrees depending on whether the core held a one or a zero. The trouble was noise. “The bigger you made them, you see, the poorer was the signal-to-noise ratio,” he said, and a noisy signal could cause a read error.
It was also a race. RCA’s core memory group, led by Jan Rajchman, competed with the MIT group led by Forrester, and Widrow remembers the rallying cry in his lab as “Hip-Hip-Array, Beat RCA.” Whether his nondestructive scheme was ever the one machines used is something the two interviews never say.
What this portends is a question to ask of any storage, old or new. A system that senses a bit by disturbing it needs a repair step, and a repair step is where delays and failures can hide. If the power dies between the read and the rewrite, that bit is gone for good, which is the kind of window a designer closes with extra circuitry. Nondestructive readout would close it at the source, and Widrow’s thesis, by his own account, was going after exactly that, with noise as the obstacle he describes in 1997.
Every read was a rewrite.