The Computer at the Bottom of a Canal: How Rekursiv Was Forty Years Ahead of Its Time

At the bottom of the Forth and Clyde Canal, in the silt, rests a box of custom chips that was right about almost everything. Alasdair Allan revisits the story in an essay for Negroni Venture Studios.

How a Turntable Maker Built a Computer

In 1972 Ivor Tiefenbrun founded Linn Products in Glasgow, famous for its Sondek LP12 deck. By the early 1980s Tiefenbrun had come to loathe the software running his factory: he wanted a system in which every physical object — right down to an individual turntable moving along the line — would have a software twin accumulating its history.

And so was born LINGO, an object-oriented language, a blend of Smalltalk and Algol. It worked, but on a VAX it was hopelessly slow. Tiefenbrun's conclusion was very much in character: the hardware was to blame, so Linn would build the hardware. The idea of turning the prototype into custom chips was sketched out by engineers David Harland and Bruno Beloff on a delayed train — which is why Harland later joked that “British Rail is to blame for Rekursiv.”

Objects All the Way Down

LSI Logic's fab produced four gate arrays with names in the company's house style: NUMERIK, LOGIK, OBJEKT and KLOK. The oddity of the machine was not that it executed object programs, but that neither the programmer nor the compiler ever saw an address. Every object carried a 40-bit number, and the OBJEKT chip translated numbers into physical addresses, checking the type and bounds of every access in hardware. Run past the end of an array or forge a reference, and the machine simply refused. Garbage collection went into silicon too, and memory and disk formed a single object store: if the object you needed wasn't in memory, the processor froze mid-instruction, waited for it to be paged in, and carried on as if nothing had happened. The instruction set was loadable — the machine ran C, Prolog, Scheme, PS-algol.

The Attack of the “Killer Micros”

Rekursiv belonged to a respectable tradition: pulling the hardware up toward the language to narrow the “semantic gap.” The trouble was that the tradition was publicly demolished at exactly the moment Linn bet on it: Patterson and Ditzel made the case for RISC in 1980, while commoditized microprocessors were getting cheaper and faster by about 52% a year. Development took four years; conceived in the age of the VAX, the machine shipped in 1988 into a world of SPARCstations and 386s. Two or three dozen boards were built and scattered among universities; the Linn assembly line, the whole reason for the exercise, never ran.

The ending was pure Glasgow farce: a Linn van dented Harland's Porsche, Tiefenbrun refused to pay for the repair, Harland quit — and on his way out hurled his hardware and backup media into the canal.

Right on Every Count

The author runs through Rekursiv's four key decisions and shows that all of them came back: hardware memory protection lives on today in CHERI and Arm Morello; garbage collection as an architectural problem, in Azul's work; end-to-end persistent storage runs from the IBM System/38 to CXL; and silicon tailored to a specific task became the norm in the era of TPUs and transformer ASICs. His moral is subtler than the usual “don't argue with the curve”: IBM won with the same ideas because it invested them in a virtual instruction set on top of the hardware, whereas Linn soldered them into four gate arrays. Ideas lifted above the silicon outlive any process node. Rekursiv guessed the technology perfectly and the timing catastrophically. Being right too early is not a business model.

Source: Lobsters