Richard Greenblatt. The Lisp Machine. Working paper 79. MIT AI Laboratory. 1974.

A description of the architecture later built and sold by LMI.

A visionary description of a machine to run Lisp at a “non-prohibitive cost” of $70,000 per system. (The web tells me that this is the equivalent of $440,000 in 2024.) A completely integrated system based on a Cons processor and using a PDP-11 as a console. A relocating, compacting, garbage collector with invisible forwarding pointers. Non-linear spaghetti stacks to support advanced programming constructs.

One of the many notable facets of this design is one of its proposals (the “frame pushdown list”) to avoid excessive conses and garbage collection overhead by maintaining lists of blocks of particular sizes. It uses this approach to manage the creation of call stack frames; modern Python implementations use this scheme throughout to avoid heap fragmentation through allocating and re-allocating blocks of different sizes. (Lisp has this feature generally, of course, with its use of fixed-size cons cells.)

It’s also fascinating to see that, even at this early stage, hardware design was being seen as an exercise in refinement to maintain software-level abstractions even as features were being pushed into hardware – while still providing optimised implementations that were invisible to the programmer.