The Rekursiv: How a 1988 Scottish Computer Anticipated Modern Memory Safety, Garbage Collection, Persistence, and Domain‑Specific Silicon
The Rekursiv: How a 1988 Scottish Computer Anticipated Modern Memory Safety, Garbage Collection, Persistence, and Domain‑Specific Silicon
The Rekursiv’s Core Idea Was Right, Even If the Timing Was Wrong
The 1988 Rekursiv, a custom silicon box built by Scottish hi‑fi company Linn, implemented hardware‑enforced memory safety, garbage‑collected persistent objects, and a domain‑specific instruction set—features that only became mainstream in the 2020s. The story matters because it shows how technical convictions can outlive the economics that initially kill them.
A Hi‑Fi Company Builds a Computer
Linn Products, famous for the Sondek LP12 turntable, ran its factory on VAX‑11 mainframes in the early 1980s. Founder Ivor Tiefenbrun disliked the VAX software stack and wanted every physical object in the factory to have a shadow software object that recorded its full history. To achieve this, Linn hired programmers and University of Glasgow lecturer David Harland to create an object‑oriented language called LINGO (a Smalltalk‑like language with Algol‑style syntax).
LINGO ran too slowly on the VAX, so the team concluded that the hardware was the bottleneck. While traveling on a delayed train, Harland and engineer Bruno Beloff sketched a custom chip architecture that would implement the language directly in silicon. The project, later named Rekursiv, was funded in 1984 by Linn, a £10 million Department of Trade and Industry grant, and fabricated by LSI Logic as four 1.5 µm CMOS gate arrays: NUMERIK, LOGIK, OBJEKT, and KLOK.
Objects All the Way Down
The Rekursiv eliminated visible addresses. Every object received a 40‑bit identifier; the OBJEKT chip translated identifiers to physical locations via a hashed page table, checking type and bounds on every access. This hardware‑enforced capability model prevented out‑of‑bounds reads, forged references, and allowed objects to be moved without updating pointers.
Because only OBJEKT knew an object’s location, the chip could perform hardware garbage collection: a two‑space compacting collector slid live objects between DRAM halves while the CPU continued execution. Memory and disk formed a single persistent object store; if an object was not in DRAM, the processor stalled mid‑instruction while an external disk controller fetched it, then resumed transparently.
The instruction set was not fixed; it was loaded as a microcode artefact. Linn supplied a C microcode, James Lothian wrote a Prolog microcode, and other groups ported Scheme and PS‑Algol. The Rekursiv could therefore execute high‑level language primitives (e.g., a CONS cell in 2 µs, Prolog unification as a single instruction) far faster than contemporary Lisp workstations—though those performance claims came from Linn’s own simulations and were never independently verified.
Why the Rekursiv Failed
The Rekursiv arrived just as the industry shifted from complex instruction set computers (CISC) to reduced instruction set computers (RISC). The seminal RISC papers (Patterson & Ditzel, 1980) argued that most programs use only a small subset of complex instructions, and that simple instructions combined with caches and good compilers yield higher performance. By 1988, commodity microprocessors (SPARC, 386, 486) outpaced the Rekursiv’s four‑year‑long design cycle.
Moore’s Law was still strong, delivering ~52 % annual performance growth from 1986‑2003—a period dubbed the attack of the killer micros. The Rekursiv’s custom silicon could not compete with rapidly improving off‑the‑shelf CPUs, and only about 20‑30 boards were built, mostly for universities. A Sun‑3 running threaded‑code LINGO was roughly twice as fast as the Rekursiv it was meant to surpass.
The final blow came from a personal dispute: a delivery van crashed into Harland’s Porsche, Linn refused to pay for repairs, Harland resigned, and he dumped the remaining hardware and backup media into the Forth and Clyde Canal. One board survived and is now displayed in the Jim Austin Computer Collection near York.
Four Design Decisions That Became Mainstream
| Rekursiv Design Decision | Modern Realisation |
|---|---|
| Hardware memory safety – OBJEKT enforced type and bounds on every access. | CHERI capability architecture (Cambridge/SRI) and ARM Morello prototype (2022) and the Memory Tagging Extension in Android phones. |
| Garbage collection in silicon – two‑space compacting collector ran alongside execution. | Azul Vega processors with hardware read‑barrier support for pause‑free Java GC; the algorithm now runs in software on commodity x86. |
| Single‑level persistent store – memory and disk formed one object store. | IBM i (System/38 lineage) maintains a single‑level store; Intel’s Optane attempted a similar model, and CXL memory fabrics are pursuing persistent memory today. |
| Domain‑specific silicon – custom gates for a specific object‑oriented workload. | Modern TPU, Groq, Cerebras, and Etched transformer ASICs—all built to accelerate a narrow class of workloads (linear algebra, streaming, or transformer models). |
All four ideas have either shipped in production silicon, are being deployed at scale, or define the direction of current research. The Rekursiv was simply four years too early.
Lessons for Today’s Hardware Landscape
- Technical correctness is not enough – market timing and economic conditions can doom even visionary designs.
- Abstractions above silicon survive – IBM’s TIMI virtual ISA allowed the System/38’s semantics to outlive the original gate arrays, showing that putting ideas in a portable abstraction layer can extend their lifespan.
- Domain‑specific architectures are now mainstream – With Dennard scaling gone, the industry embraces workload‑specific silicon, exactly the path the Rekursiv anticipated.
- Hardware memory safety is finally viable – Capability machines and memory‑tagging extensions prove that the Rekursiv’s safety model is practical and valuable.
Community Insight
"Thing is capability machines (like this and the 432 and lots of research machines) were very much the thing at the time – cutting edge even. The research literature was full of them. What ate them up was ‘what can you fit all on a chip with not many pins’, followed by ‘what can you fit along with a cache on a chip with more pins’…" – Taniwha (Hacker News comment)
"It’s crazy what a small group of smart people can achieve if they’re not captive to conventional wisdom. They got crushed by the commodity curve and Moore’s Law… AI has ended the scarcity of programmers so dedicated purpose hardware is more viable than ever." – JSR_FDED
The Rekursiv’s Relevance to AI‑Driven Languages
The article argues that today’s AI‑generated or model‑centric languages (e.g., Vera, agentlanguages.dev) are reshaping themselves to fit the hardware, just as the Rekursiv reshaped hardware to fit its object‑oriented language. The cycle is complete: the substrate (silicon) now adapts to the authors (LLMs), not the other way around.
Conclusion
The Rekursiv was a bold experiment that anticipated four major trends: hardware‑enforced memory safety, built‑in garbage collection, a unified persistent object store, and domain‑specific silicon. Although it sank in a Scottish canal, its ideas have resurfaced in modern architectures such as CHERI, ARM Morello, persistent memory systems, and AI‑focused ASICs. The story illustrates that technical foresight can survive decades of economic change, and that the current end of Moore’s Law makes revisiting such contrarian hardware designs both feasible and profitable.
요약 1988년 Linn Products가 만든 Rekursiv 컴퓨터는 하드웨어 메모리 안전성, 가비지 컬렉션이 적용된 영구 객체, 그리고 워크로드‑특정 실리콘이라는 아이디어를 선구적으로 구현했으며, 이러한 개념들은 수십 년이 지나서야 주류 실리콘에 도입되었습니다.
제목 The Rekursiv: How a 1988 Scottish Computer Anticipated Modern Memory Safety, Garbage Collection, Persistence, and Domain‑Specific Silicon