Core memory

from Wikipedia, the free encyclopedia
Close-up of toroidal cores with read / write lines

The core memory , magnetic core memory , or ferrite core ( English magnetic-core memory or English ferrite-core memory is) an early form of non-volatile memory random access by computers . It consists of hard-magnetic toroidal cores threaded onto wires , which can be magnetized and read out by electrical currents in the wires. The sign of the magnetic remanence of the individual toroidal cores represents their memory content.

Core memories were used in the computers that were common at the time from around 1954 to 1975.

Fixed-value core memory (core rope memory), very likely from a Nixdorf System 820 , storage capacity max. 256 wires per 16 lines, results in 4096 handwired command words of 18 bits

The core memory described here must be distinguished from the core rope memory , which works as a ROM and in which the program is determined by the type of wiring. This also uses toroidal cores, which in this case do not store any information magnetically, but only work as a transmitter .


The first work was carried out in 1949 by the Shanghai-born physicist An Wang at Harvard University . In contrast to MIT , Harvard was not interested in patenting its own inventions. Wang himself acquired the patent under the name pulse transfer controlling device .

Jay Forrester's group, who had worked on the Whirlwind project at MIT, learned about Wang's work. Whirlwind needed a fast storage system for a real-time flight simulator. Until now, runtime memory had to be used for this. So-called storage tubes , based on cathode ray tubes such as the Williams tube or the Selectron , never achieved a significant market position due to manufacturing difficulties and poor operational reliability and were replaced by core storage in the mid-1950s.

Two key inventions led to the development of core memory, which only allowed the development of the computers known today. The first, An Wangs, was the write-after-read cycle , which solved the problem that reading out information also destroys it: the magnetic polarity of the toroidal cores can only be determined by reading them be remagnetized.

The second, Jay Forresters, was the coincident-current system , which made it possible to control a large number of magnetic cores with a small number of wires (see below, How it works). Core memories were "threaded" manually. The work was carried out under the microscope and required special manual dexterity. In the late 1950s, factories were built in Asia in which low-wage workers made the core memories. The prices were reduced to such an extent that both the inexpensive but low-output drum storage and the expensive high-performance systems with electron tubes could be replaced in the early 1960s. Due to the mechanical complexity and the resulting voluminous design, the capacity of the core memory was limited. Versions up to a few megabytes were built. For this, however, several control cabinets were required, with less space and less effort, less than 100 kilobytes were achieved.

Although the production of the core memory was discontinued shortly before its automation, the costs followed Moore's law, which was still unknown at the time . The technology costs from initially around one dollar per data bit sank to around 0.01 dollar per data bit until the core memory was replaced by the silicon-based RAM in the early 1970s . Wang's patent was not approved until 1955, when the technology was already in use. Several lawsuits caused IBM to buy the patent from Wang for several million dollars. Wang used the money to expand Wang Laboratories .

Core memories belonged to a family of technologies that made the magnetic properties of materials their own. In addition to the toroidal core memories, the design with magnetic stick memories was also used (for ICL computers in the 1970s). In the 1950s, electron tubes were already fully developed, but still fragile and, because of the heated filaments, short-lived, unstable and too high in energy consumption. Core storage was an energy-saving, miniaturized and reliable alternative. However, it was essential that, like the drum storage unit, it did not lose its contents even when the operating voltage was switched off. After a further miniaturization step, the so-called bubble memories , it was only really replaced by non-volatile semiconductor memories (EEPROM / Flash memory) .



Simplified ring core memory with write and read wire
The readout process is illustrated by the hysteresis curve. Left: "1" is read, right "0" is read.

A core memory essentially consists of a large number of magnetizable, hard-magnetic ferrite cores, which are shaped into rings and are therefore referred to as toroidal cores . Hard magnetic in this context means that each core can store a data bit with the sign of the remanent flux density B r . To read or write to the cores, at least two electrical and mutually insulated wires run through the ring openings, as shown in the adjacent figure on a core.

The electrical current I m in the writing wire must be so large for writing that the magnetic coercive field strength H c generated by the current I m is exceeded in the magnetic circuit of the core. As a result, the hard magnetic material of the ring, which has an almost square hysteresis loop , stores the state with the sign of the remanent flux density. The remanent flux density can assume two stable points, which in the hysteresis loop are denoted by B r and -B r .

Simultaneously with the writing process, a voltage pulse is induced in the second wire, the read line, on the basis of which the orientation of the remanent flux density originally stored in the core can be determined. The information can only be read out destructively. The possibly remagnetized core must then be rewritten in order to restore the original data content.

The readout process is illustrated in the illustration of the hysteresis curves on the right: If a positive remanent flux density + B r was previously stored in the core , when a "0" is written, the large change in the magnetic flux density in the core on the read line results in a voltage pulse U l in the A few hundred mV. The course of the flux density is shown in bold in the hysteresis curve on the left. After switching off the current I m in the write line, the core remains the remanent flux density - B r , which corresponds to the state “0”. If a negative remanent flux density - B r was already stored in the core , the flux density only runs through a small part of the hysteresis curve and the rate of change is minimal. As a result, the voltage pulse on the read line is also minimal. In both cases, the core is in the "0" state after reading out, and the original memory content must be rewritten if necessary using an inverse current - I m .

In addition to the core memories, read amplifiers are required for operation, which convert the low voltage pulses on the read line into suitable logic voltage levels. Power sources are required for writing .

Arrangement in a matrix

Scheme of a matrix arrangement

So that each core does not need two separate wires and its own read amplifier, the following trick is used: The current I m of the write line is divided into two wires, which each carry only half the current strength required for magnetization reversal. These X and Y wires are arranged in a lattice structure (matrix) and have a core at each crossing point. If a certain core is to be addressed, half of the required current is contributed by the relevant X-wire and the relevant Y-wire. This means that other nuclei are either only reached by half or no field strength at all and do not change their state.

To implement a 16 kbit memory, 2 × 128 wires and just as many controllable power sources are required.

Read Write

For reading and writing in a matrix, two more wires are required, which are looped through all cores - the sensing wire ( sense-line , S ) or also called S-wire, and in earlier core memories also the blocking wire ( inhibit-line , Z ).

A read and a write cycle are always carried out at the same time. In the read cycle, the X and Y wires are used to remagnetize the corresponding core towards the logic "0". If the core has already saved a “0”, nothing happens in the read cycle, whereas with a “1” a pulse is induced in the S-wire due to the reversal of magnetization. In the write cycle, the core is magnetized again in the "1" direction. In the case of a previously saved “0”, a current is sent in the opposite direction through the blocking wire during the write cycle. This is sufficient to weaken the field strength of the X and Y wires to such an extent that the core is not remagnetized in the "1" direction.

Since the sensing wire and blocking wire are never used at the same time, later systems used only one wire. An additional control switches between the two functions.

Computer systems with core memory often take advantage of the fact that not every read value has to be retained at all. If, for example, a value is to be added to a data word, it is initially only read (read cycle). The write cycle waits until the addition is completed. Then the original value is not written, but the result of the addition. In this way, the speed of certain operations can be doubled.

The total time used for a read / write cycle was called the cycle time ; it was a measure of the speed with which a core memory could be operated. In computer systems of the 1960s, it was often a rough measure of the overall performance of the system, as was the CPU clock rate later.

Physical Properties

Core storage element
On a section of core memory that can store eight bytes in 64 cores is a microSD HC card that provides billions of times as much storage space in a much smaller space, 8 billion bytes (8 GB).

Early systems had cycle times (read and write back) of around 20 µs; it fell to 2 µs in the early 1960s and reached 0.3 µs in the early 1970s. The possible clock rates between 50  kHz and 3  MHz were thus roughly the same as those of the home computers of the late 1970s and early 1980s, for example the Apple II and the Commodore 64 .

Data words with 32 data bits were distributed over 32 levels (one XY grid each), so that an entire data word can be accessed in one read-write cycle.

Core memories are non-volatile memories - they receive the information for an unlimited time without electricity. Core memories are also robust against electromagnetic pulses , high temperatures and radiation. These are important advantages in military applications such as combat aircraft , but also in space vehicles . Core memories were used here for several years after the availability of semiconductor memories began.

Characteristic of core memories: they react to the current, not to the voltage - the reading wire, however, delivers a voltage pulse. This was an important prerequisite for high clock rates with relatively large geometrical dimensions of the memory.

The half select current I m / 2 was typically 400 mA for the later smaller and faster memories. Earlier storage facilities required larger flows.

The diameter of the toroid is in the order of 1 mm to 0.25 mm with the shortest access time.

A negative property of the core memory is the dependence of the hysteresis on the temperature. The control system adjusts the selection current - the temperature is measured with the aid of a sensor. The programmed data processor PDP-1 from Digital Equipment Corporation is one example. Other systems circumvented the problem by placing the memory in a temperature-controlled container. Examples are the IBM 1620 (it took up to 30 minutes to reach the operating temperature of 41 ° C) or the memory of the IBM 709 , which is housed in a heated oil bath .

Other designs

The destructive read-out process and the forced rewriting of a read bit in the classic toroidal core memory led to a series of further developments in the 1960s and 1970s that remedy this disadvantage. One possibility is to construct the magnetic circuit in such a way that the directional dependence of the magnetic flux density is used. These cores are also known as Biax cores .

Two holes are made on the cuboid core, which are orthogonal to each other. The interrogation wire is passed through one hole and two wires through the other: the writing wire and the reading wire. When reading out the core, a current pulse is only sent through the interrogation wire, which, depending on the remanent flux density in the core, causes a positive or negative voltage pulse on the reading wire. However, this does not lead to any permanent change in the magnetic flux density in the core, and the core does not lose its memory content. For writing, two sufficiently large current pulses must be sent simultaneously in the corresponding direction both through the interrogation wire and through the writing wire that is orthogonal to it. Only then is the amount of the coercive field strength exceeded and the new state is saved.

Other core designs that use different flux density distributions in magnetic circuits are referred to as transfluxors . Two holes of different sizes are placed asymmetrically on the toroid, thereby leading three cables. Due to the different widths of the legs, due to the asymmetry and size of the bores in the magnetic core material, there is also the possibility of non-destructive reading of the memory content with this design variant.


The term core dump , which is common in technology jargon and which is used in hardware- related programming as an expression for a memory dump in the form of a snapshot, is derived from the core memories. For troubleshooting, all cores ( cores ) read ( dump ) to be able to find all the possible bugs.


  • F. Dokter, J. Steinhauer: Digital electronics in measurement technology and data processing . tape 2 . Philips reference books, Hamburg 1970, ISBN 3-87145-273-4 , p. 276-313 .


  • Patent US2667542 : Electric connecting device (matrix switch with iron cores). Registered September 25, 1951 , published January 26, 1954 .
  • Patent US2708722 : Pulse transfer controlling devices. Registered October 21, 1949 , published May 17, 1955 , inventor: To Wang .
  • Patent US2736880 : Multicoordinate digital information storage device (coincident-current system). Filed May 11, 1951 , published February 28, 1956 , inventor: Jay Forrester .
  • Patent US3161861 : Magnetic core memory (improvements). Registered November 12, 1959 , published December 15, 1964 , inventor: Ken Olsen .
  • Patent US4161037 : Ferrite core memory (automated production). Published July 10, 1979 .
  • Patent US4464752 : Multiple event hardened core memory (radiation protection). Published August 7, 1984 .

Web links

Wiktionary: core memory  - explanations of meanings, word origins, synonyms, translations
Commons : core memory  - collection of images, videos and audio files

Individual evidence

  1. N. Metropolis et. al (Ed.): History of Computing in the Twentieth Century . Academic Press, 1980, ISBN 0-12-491650-3 , pp. 465 to 469 .
  2. R. Rost: Kristalloden Technik . 2nd supplementary volume. 2nd Edition. Verlag von Wilhelm Ernst & Sohn, 1960, p. 56
  3. a b Digital memory with ferrite cores, Robert Schmitt, Verlag Siemens AG Berlin-Munich 1971
  4. CJ Quartly: Circuit technology with rectangular ferrite . Philips Technical Library (company publication), Eindhoven 1965.