As I described in Dimond Rings and Read-Only Ropes, both of the methods for constructing rope read-only memories had their problems.
Those using magnetic toroids for the bit transformers were reliable, but labor intensive to string, and the inevitable stringing errors frequently required extensive rework to correct. Each time the contents were altered also required another test cycle, which led to manufacturing delays and added to the costs.
The ones using E or U shaped magnetic cores suffered somewhat in reliability due to problems aligning and maintaining the keeper bar over the open end of the magnetic path. This caused large variations in mutual inductance between the word wires and the sense winding, which made it difficult to predict a sensing threshold with any repeatability.
The open core method did, however, allow for the use of automation in weaving the wire pattern for the contents. A typical technique was the use of X-Y wire wrap machines such as those used to automatically wire computer backplanes, which could be programmed to lay in the desired pattern and do it flawlessly every time.
Engineers at Quadri Corporation wanted the efficiency of automated content stringing IF it could be had with a consistent, predictable, magnetic path loss. They settled on the idea of using a completely open magnetic path. Compared to a closed magnetic loop, this would have extremely high flux loss, hence low coupling between the word wire and the sense winding and a very low voltage induced pulse during the read operation. It WOULD, however, be predictable, thus reliably detectable.
The final form factor selected was to wind a many-turn sense winding on the end of a small ferrite rod. Word wires storing a ONE in a bit position would make an approximate one-turn loop about the rod, while those storing a ZERO would be routed straight past the rod. Sense voltage from the single-turn ONES would be extremely low, but Quadri had experience in sensing small signals based on their ferrite core memory product line, so that wasn’t much of a problem.
Automated weaving of the wire rope was the goal here, of course, not difficult to sense signals. To address that problem, several of us visited a textile factory to observe a Jacquard loom in operation. These devices were used to weave complex patterns in fabric for blankets and brocades, all automatically under punched card control.
Our stringing machine, based loosely on the Jacquard principle, had 256 spools of magnet wire feeding 256 needles, which were arranged side by side at about .2″ spacing. These were actual sewing needles, welded to the ends of thin rods. Each rod had a small solenoid mounted perpendicular to it, near it’s lower end. A static Hollerith card reader energized all solenoids corresponding to ONES. This pushed the end of the rod out far enough for a pneumatically-driven bar to raise each selected needle.
The stringing operators inserted a card in the reader, pushed a button that closed the reader and energized the solenoids, followed after a slight delay by actuation of the driving bar. This created a hole across the web of wires, with ONES above and ZEROES below. They then stuck a long skewer through the hole capturing the pattern, released the card reader, and moved on to the next bit position. After all bits had been skewered, the wires were pushed along the skewers into a plastic form that maintained their alignment. The plastic form was subsequently placed over the ferrite rods with their sense windings, and the assembly was complete.
Quadri had limited success with this technique but they did sell it to makers of 2K x 48-bit microcode memories for IBM-compatible disk drives. Years later I learned that Nixdorf used a similar design for microcode stores in their IBM-compatible mainframes. Both Nixdorf and Quadri called it “rod” memory.
Note: One curious requirement for the data cards used in the stringing machine reader was the inversion of a 256 x 48 matrix. We received desired memory contents from the customer on cards, one per word, with 48 columns punched either 0 or 1. This had to be converted, for each 256 words, to a deck of 48 cards (one per bit) with 256 locations (one per word) of hole or no hole. How we performed this operation is a topic for another story.