Gene Cloud remembers that you could make far more bits in an IC RAM than in magnetic core and far cheaper, but it took a long time to get there. If you needed external storage, you probably used a disk drive (developed by IBM in the 1950s). If you needed random-access memory (until the early 1970s) you used ferrite-core memory (developed by IBM in the 1950s). Ferrite cores were wonderful. They were compact, they were cheap, they were fast. They were getting faster and cheaper and they were reliable.

Before ferrite-core memories were developed, if you needed random-access memory you could design one with flip-flops. The more prevalent memories in the late 1950s and early 1960s were magnetic drums. These were serial, rather than random-access, memories; they were slow and they were big.

Magnetic drums stored a lot and typically rotated at 3,600 rpm. The leading manufacturer was Bryant Computer Products of Walled Lake, Mich. It was possible to speed a drum's access time by repeating data or a program in a ring around the drum or by mounting several read/write heads around a ring, so the same data could be accessed several times during one rotation. But that, of course, reduced the amount of data one could store. Lengthening the drum (sometimes to several feet) or using a drum with greater diameter added memory. It also added mechanical problems.

Many computer houses built their own core memories, but there were several independents, Electronic Memories and Magnetics being among the largest. The cores themselves were easy to fabricate—they were just magnetic doughnuts made of special ferrites. The harder part was making core planes, with the individual cores arranged in a pattern of alternating diagonals.

The tricky part of core

Wiring a core plane was the tricky part. This was done in the Far East by women (mostly women because they worked cheaper) who toiled for nickels an hour. They used hypodermic syringes to shoot fine wires through rows and columns of core holes for X and Y addressing, then painstakingly threaded sense wires through hundreds, sometimes a few thousand, cores. A giant core plane, which might be a foot square, might have an array of 64 x 64 cores for a total memory of 4,096 bits, a staggering number.

Core planes could be wired together, one above another, to form a core block or stack. Some of these blocks could occupy the volume of a refrigerator and could weigh many pounds. But computers were big, too. As cores got smaller, they got faster. Toward the end of their lifetime, cores' inside and outside diameters came to around 10 mils.

Toward the end of the 1960s, it was quite clear that cores would be the memory of forever, though major progress was being made in other computer technologies. Computer manufacturers were already using transistors. In l964, IBM introduced its powerful 360 computer, which used thick-film hybrid circuits, with transistors solder-balled on small ceramic wafers, instead of vacuum tubes and discrete components. And small ICs were beginning to see use in some advanced computers.

Along came integrated-circuit memory. One of the earliest was a bipolar RAM designed by Tom Longo at Transitron in 1966. It could store 16 bits. Other IC memories followed, none of them creating much of a stir in the marketplace. In 1969, year-old Intel introduced its first product, the 3101, a 64-bit bipolar memory. In 1970, Longo, then at Fairchild, managed a group that developed 1-kbit and 4-kbit bipolar memories. These generated some interest but no impact on the market. Some bipolar memories cost $2 a bit.

Intel's 256-bit 1101, a six-transistor pMOS SRAM, may have been the first semiconductor memory to achieve a degree of success in the marketplace. But it wasn't a world beater. It was expensive. Recollections differ but range around 3 cents per bit.

The core killer

The Intel 1103, introduced in late 1970, was the chip that broke the back of ferrite cores. It was a fully decoded, 1,024-bit dynamic RAM using silicon-gate pMOS technology. In an 18-pin ceramic or plastic DIP, the 1103 sold at first for about a penny a bit, matching core's price. But then core prices cratered to 0.3 cent a bit. Intel's prices declined further. Intel boasted that this chip, whose core equivalent occupied about a square foot and weighed almost a pound, was the core killer, and Intel was right.

By mid 1971, a dozen semiconductor companies swore that they were the only one actually producing 1-kbit semiconductor memories. But cores were still strong, so much so that many people tended to refer to their semiconductor RAMs as "core."

While the 1103 knelled the death of core, core didn't suffer an instant demise. Ferrite core put up a valiant battle. It had some powerful arguments.

It had proven success, while IC memories couldn't quite be trusted. Core was nonvolatile; if you pulled the plug, a core stack would retain its memory while a semiconductor memory would forget everything. In fact, you had to refresh the 1103 every 2 ms or it would lose its memory. (The capacitors that stored a bit or no bit would discharge.)

Further, core was far down the learning curve so semi memories would never approach the cost of core. What's more, core was getting faster, thanks to smaller and smaller cores, and cheaper, thanks to cheap labor in the Far East.

Perhaps most important, the 1103 used MOS technology, a very shaky technology that many people didn't trust—often with good reason. It was just difficult to make fault-free MOS transistors. You certainly wouldn't bet that every transistor would be good in a large memory, like a mammoth memory with 1,024 bits.

At Texas Instruments, L.J. Sevin created a 256-bit experimental MOS RAM, but he had to fight a management that felt that the way to beat cores was with thin-film memory. And besides, management had a strong bias toward bipolar. MOS couldn't be trusted.

While Intel boasted that the 1103 was the core killer, there was a great deal of nervousness and trepidation inside Intel. Ron Whittier, who was then Intel's vice president and general manager of the Memory Components Division, remembers.

"The 1103," he says, "was tough to make, tough to design, tough to use. But it was absolutely a critical part of the semiconductor legacy in solid-state memories."

As with any new technology, the 1103 began to get less expensive and easier to use, though it took a while for engineers to accept the need to refresh every 2 ms. But as the 1103 moved down to a lower price point so did magnetic core. It was hard to compete, says Whittier.

The 1103 had an advantage in speed. Access time was down to a brisk 300 ns and full cycle time was 580 ns while fast core was in the millisecond range.

Fortunately, there were forward-looking companies. Honeywell, for one, had been looking into solid-state memories. Engineers there saw little likelihood that a six-transistor cell like that in the 256-bit 1101 would ever make economic sense, but the 1103, with one-transistor cells, looked promising. It looked feasible, though perhaps far-fetched, that semi memories might in time cost less than magnetic ones.

And even as early as 1970, Gordon Moore was preaching what became known as Moore's Law: complexity would double every 18 months. "In our mind," says Whittier, "it was just a matter of time when we would reach the crossover point.

"We were hot for MOS," Whittier recalls, "because it was an inherently simple process. It was incredibly simple. But MOS was very much subject to ion contamination and other problems. If we could only figure out how to whip the reliability problems, then, with a simple process and a one-transistor cell, we could forecast really attractive pricing."

But a one-transistor cell, dynamic memory and refreshing were very aggressive concepts. Even if you could produce anything like that consistently, Whittier remembers, it wasn't clear that anybody would buy it. There were many unknowns, even inside Intel. "We knew it was risky. It was one of the all-time high-risk propositions. We were working 18 hours a day."

Gene Cloud, formerly a vice president at Micron Semiconductor, now the dominant manufacturer of DRAMs in the country, says that you could make far more bits in an IC RAM than in magnetic core and far cheaper, but it took a long time to get there.

But now we're there. What next?

The Century of the Engineer: Misunderstood Milestones