Magnetic Bubble Memory (1967-1986)

Short answer: magnetic bubble memory was a non-volatile, no-moving-parts store invented at Bell Labs in the 1960s (Andrew Bobeck is acknowledged as its father), first sold as a chip by Texas Instruments in 1977, pushed to one megabit by Intel in 1979 and to four megabits in 1982-83, and it lost the general market by 1981 because it was hard to make and its rivals, floppy disks, hard disks and RAM, got cheaper faster than it did; it lingered in rugged niches until the late 1980s.

1960s: the idea. The New York Times (1981) says Bell Labs researchers in the mid-1960s were looking for something more efficient than disks and drums, that in 1965 nobody had a material better than a research one, and that Richard Sherwood suggested at a meeting of 70 to 80 scientists that orthoferrites or garnet would work. It says the first bubble-memory patent was submitted in 1966 with four names: Bobeck, Sherwood, William Shockley and U. F. Gianola. Digital Design (May 1977) gives a different date: it says the first patent suggesting bubbles for memory applications was granted to Bell Telephone Laboratories in 1969. These can both be true, since one is a filing and the other a grant, but neither source says so, and the card does not reconcile them. TI's application book dates the concept to 1967 at Bell Laboratories; the 1966 to 1969 spread is therefore a mix of filing, grant and introduction dates.

How it worked. TI's application book describes a thin crystalline magnetic film grown on a gadolinium gallium garnet substrate. A bias field from permanent magnets shrinks the film's serpentine domains into cylinders of 2 to 30 micrometers in diameter; two perpendicular coils make a rotating in-plane field that moves the bubbles under permalloy patterns, much like a shift register; a bubble in the detector lowers a magnetoresistive element's resistance, giving an output of about 3 millivolts. A single long loop would be simplest but would be slow, since a bit has to circulate through the whole loop, and a single defect would ruin the chip. TI therefore chose the major-minor loop layout, in which data circulates in many short minor loops and is transferred in parallel to a major loop for reading and writing; redundant minor loops let defective loops be mapped out. Digital Design (1977) describes the same layout and adds that propagation patterns evolved from T-bars to half-disks and asymmetric chevrons, which it says gave a four-fold increase in bit density for the half-disk over the T-bar at the same processing capability (a Bell Labs figure quoted from Bobeck's Electro 77 paper).

1977: the first products. TI says that in the spring of 1977 it was the first to market a 92,304-bit bubble memory. The same bit count is attached to TI part numbers TIB 0103 (Wikipedia excerpt; Found and Lost) and TIB0203 (the TI application book), so the part number is unresolved here. TI's own Silent 700 Model 763 and 765 terminals used bubble memory: the January 1978 maintenance manual says they were the first commercial application of the new memory system, with 20,000 characters standard and up to 80,000 in 20,000-character steps, used much as a floppy disk would be, and a maximum of 91,206 characters stated in the introduction. The same manual says the 765 is portable and the stored data survives power-off. Digital Design (May 1977) separately describes a Bell System application, the 13A announcement system, using a 272-kilobit serial bubble store.

1978-1979: bigger chips and the first one-megabit part. TI announced a quarter-million-bit TIB0303 for late 1978 with 3-micrometer bubbles, a 7.2-millisecond average access time, 0.9 watts of power and a 100-kilobit-per-second read rate (vendor claim, from a trade-press note transcribed on the same page). Intel, through its Intel Magnetics subsidiary, had announced the 7110 by April 1979 with a family of support chips: Intel's own authors in Electronics (26 April 1979) wrote that bubbles had a more difficult interface problem than semiconductor memories and that manufacturers had to supply a whole chip family, not only the memory. The 7110 datasheet gives 1,048,576 usable and 1,310,720 gross bits (256 loops of 4,096 bits; the surplus is for redundancy and error correction), a 512-bit page, 2,048 pages, and a 100-kilohertz maximum data rate with a 40-millisecond average access time; the 7112 variant runs at twice the rate with a 20-millisecond average access time (all vendor claims).

1982-1983: four megabits. Intel's November 1982 advance information sheet for the 7114 gives 4,194,304 usable bits (5,242,880 gross), a major-minor loop layout, and for the 7114A a maximum data rate of 400 kilobits per second and a 40-millisecond average access time (vendor claim). The Wikipedia excerpt and Found and Lost date the 7114's introduction to 1983; the datasheet read for this card is dated 1982 and is marked advance information. Bobeck had forecast in 1977, according to Digital Design, that chips of 10 to the 8th bits would be state of the art by 1985; the four-megabit parts of 1982-83 were a factor of about 24 below that (calculation: 10^8 / 4.19 x 10^6).

1981: the exits. The New York Times (20 September 1981) says Rockwell left the bubble market in February 1981, Texas Instruments in June and National Semiconductor in August, each saying the market had not developed as expected. Nature (17 September 1981) reports the same three withdrawals and says only Intel remained among the four US firms that had been publicizing plans. Nature adds that National's decision was also part of a general retrenchment after weak semiconductor demand, so it was not caused by bubbles alone. The New York Times lists the companies still in the field as Bell Labs and IBM (mostly for their own use), Intel and Motorola in the US, and Hitachi and Fujitsu in Japan. Wikipedia's excerpt adds Plessey to the 1981 closures and says that by 1984 a 'big five' of Intel, Motorola, Hitachi, SAGEM and Fujitsu pursued second-generation bubbles; that list rests on the excerpt alone.

Why it lost, as the sources give it. (1) Manufacturing: the New York Times says the garnet substrate was not consistently good enough and yields were poor (citing a Gnostic Concepts study), that Union Carbide, the material's maker, was reported unwilling to invest in better garnet for a market that had not materialized, and that Bell's own production at Western Electric plants in Columbus, Ohio and Reading, Pennsylvania from 1979 proved harder than expected; Bobeck says it is like making integrated circuits but 'tough'. (2) System complexity: the New York Times says bubbles needed cumbersome, heavy peripheral equipment, including a heavy external drive coil, and that selling the device as a single component failed because users needed a family of control chips; it credits Intel's integrated subsystem approach with outlasting rivals. Intel's own 1979 authors agree that the interface is the harder problem. Nature says Rockwell and TI were said to have concentrated on the technology and neglected the ancillary equipment. (3) Speed and rivals: Nature says bubbles were slower than silicon and that the price of competing systems was falling so fast that bubbles' advantages could not win the volume needed to lower their own price; the New York Times says the same about RAM and floppy disks. (4) Market forecasts: the New York Times says analysts cut their forecast for the mid-1980s market sharply in 1981. Prices and market sizes are not reproduced here.

After 1981. The New York Times reports Hitachi had made the largest number of bubble memories and was staying in the business, and mentions use in machine-tool programs, oil and gas wellhead recording and portable terminals. Found and Lost, a secondary summary, says Intel divested Intel Magnetics to MemTech in 1986, that MemTech was the only US seller by the late 1980s (military and satellite buyers), and that commercial interest was gone by about 1990, citing Intel and the Dvorak News Blog; those statements rest on that one summary and its citations and were not checked in the primary pages. The same page lists consumer and computer uses such as the Sharp PC-5000, GRiD laptops, HP instruments and Konami arcade boards, also single-source.

What the sources do not settle. The 92,304-bit TI part's number (TIB 0103 or TIB0203); the date of the first Bell patent (1966 filed, 1969 granted, or the 1967 concept date); the 7114's date (1982 datasheet or 1983 introduction); and the exact end date of Intel's involvement (one secondary summary says 1986). Bit-density and power figures in the table are those the vendor stated and were not measured on a common test. A direct comparison with the contemporary disk and floppy market on capacity per unit cost is deliberately omitted because the sources that give such numbers are dated 1981 and give them in mixed units.