Source: https://www.wikiwand.com/en/Lisp_machine (via https://en.wikipedia.org/wiki/Lisp_machine)
A Lisp machine represents a category of general-purpose computers engineered to efficiently execute Lisp as their primary programming language, typically via dedicated hardware support. These systems exemplified high-level language computer architecture principles and served as early commercial single-user workstations.
Despite modest production numbers—approximately 7,000 units by 1988—Lisp machines pioneered several technologies now considered standard, including networking innovations such as Chaosnet and effective garbage collection mechanisms. Major manufacturers during the 1980s included Symbolics (3600, 3640, XL1200, MacIvory models), Lisp Machines Incorporated (LMI Lambda), Texas Instruments (Explorer, MicroExplorer), and Xerox (Interlisp-D workstations).
Artificial intelligence research during the 1960s and 1970s demanded substantial computational resources. The Lisp symbolic programming language exacerbated these demands, as commercial hardware was optimized for assembly and Fortran rather than symbolic computation. As integrated circuit technology advanced and AI memory requirements exceeded the address space of machines like the DEC PDP-10, researchers pursued a novel solution: computers specifically designed for large AI program development with Lisp language semantics. These systems typically operated as dedicated single-user workstations rather than shared resources.
In 1973, Richard Greenblatt and Thomas Knight at MIT's Artificial Intelligence Laboratory initiated the MIT Lisp Machine Project. They constructed a computer with hardwired support for basic Lisp operations using a 24-bit tagged architecture, featuring incremental garbage collection capabilities.
The initial prototype, called the CONS machine (named after Lisp's cons operator), became affectionately known as the "Knight machine." The design underwent refinement into the CADR version, based on essentially identical architecture. Approximately 25 prototype CADRs sold for roughly $50,000 each, gaining rapid acceptance within the hacker community. Following positive reception at an MIT AI conference in 1978, DARPA provided developmental funding.
In 1979, Russell Noftsker proposed commercializing Lisp machine technology to Greenblatt, envisioning significant market potential. Disagreements over business philosophy led the decision to the MIT AI Lab community itself. The lab ultimately sided with Noftsker's venture-backed approach over Greenblatt's proposed independent startup.
This division spawned two competing enterprises. Symbolics, Noftsker's company, eventually dominated the market with strong support from 14 AI Lab hackers, while Lisp Machines Inc. (LMI), backed by consultant Alexander Jacobson and founded by Greenblatt with F. Stephen Wyle, competed with fewer resources.
Richard Stallman, notably absent from both firms, spent 1982-1983 independently replicating Symbolics programmers' output to prevent monopolistic control over lab computers.
Both companies commercialized CADR-based systems: Symbolics marketed the LM-2 (approximately 100 units at $70,000 each), while LMI sold the LMI-CADR. Second-generation products followed—the Symbolics 3600 (delayed one year, expanding to 36-bit words and 28-bit address space with accelerated microcode operations) and the LMI-LAMBDA (1983, compatible with CADR microcode but with hardware differences). Texas Instruments licensed the LMI-LAMBDA design, producing the TI Explorer with dual Lisp and Unix processors.
Symbolics continued advancing the 3600 family and its Genera operating system, developing the Ivory VLSI implementation. From 1987 onward, multiple Ivory-based systems emerged: Sun and Mac boards, standalone workstations, and embedded systems (Symbolics XL-400, UX-400, MacIvory II; later platforms included Symbolics XL-1200, MacIvory III, UX-1200, Zora, NXP1000). LMI developed the K-Machine before bankruptcy prevented market introduction. TI miniaturized the Explorer into the MicroExplorer for Apple Macintosh II.
These machines provided hardware support for Lisp primitives (type testing, CDR coding) and incremental garbage collection, enabling efficient execution of substantial Lisp programs. The Symbolics machine competed favorably against commercial minicomputers but found limited conventional applications. Applications extended beyond AI into computer graphics, modeling, and animation.
MIT-derived systems ran Lisp Machine Lisp (descended from Maclisp), with operating systems written entirely in Lisp using object-oriented extensions. Later implementations supported Common Lisp variants (Flavors, New Flavors, CLOS).
Bolt, Beranek and Newman (BBN) developed a machine named Jericho running Interlisp, which never reached market. The frustrated AI group subsequently joined Xerox Palo Alto Research Center, which had independently developed Lisp machines for Interlisp execution (later supporting Common Lisp). This identical hardware also ran Smalltalk and the Xerox Star office system.
Xerox machines included the Dolphin (1979), Dorado, Dandelion (1981), Dandetiger, and Daybreak models. The Xerox systems featured advanced development environments (InterLisp-D), the ROOMS window manager, early graphical interfaces, and novel applications like NoteCards—among the earliest hypertext systems.
Xerox also pursued RISC-based Lisp machine designs using the Common Lisp Processor, planning 1987 market entry that never materialized.
During the mid-1980s, Integrated Inference Machines built Lisp machine prototypes designated Inferstar.
In 1984-85, the UK firm Racal-Norsk (jointly owned by Racal and Norsk Data) attempted repurposing the Norsk Data ND-500 supermini as a microcoded Lisp machine running CADR software, called the Knowledge Processing System (KPS).
Japanese manufacturers pursued market entry: Fujitsu's Facom-alpha mainframe co-processor, NTT's Elis, Toshiba's AI processor (AIP), and NEC's LIME. University research produced working prototypes including Kobe University's TAKITAC-7, RIKEN's FLATS, and Osaka University's EVLIS.
France saw two projects: M3L at Toulouse Paul Sabatier University and later MAIA. Germany's Siemens designed the RISC-based COLIBRI co-processor.
With the AI winter's onset and the microcomputer revolution's acceleration, cheaper desktop PCs soon executed Lisp programs faster than specialized hardware, eliminating profit margins. Most manufacturers disappeared by the early 1990s, leaving only software firms like Lucid Inc. or hardware makers transitioning to services.
As of January 2015, besides Xerox and Texas Instruments, "Symbolics is the only Lisp machine firm still operating, selling the Open Genera Lisp machine software environment and the Macsyma computer algebra system."
Multiple open-source emulator projects have been undertaken: CADR Emulation, Symbolics L Lisp Machine Emulation, the E3 Project (TI Explorer II), Meroko (TI Explorer I), and Nevermore (TI Explorer I). On October 3, 2005, MIT released the CADR Lisp Machine source code as open source.
In September 2014, Alexander Burger announced PilMCU, implementing PicoLisp in hardware. The Bitsavers PDF Document Archive maintains extensive documentation for Symbolics, TI Explorer, MicroExplorer, and Xerox Interlisp-D machines.
Lisp machines primarily served artificial intelligence applications alongside computer graphics, medical image processing, and related domains. Major commercial expert systems included Intellicorp's Knowledge Engineering Environment (KEE), Carnegie Group's Knowledge Craft, and Inference Corporation's Automated Reasoning Tool (ART).
Initially designed as personal software development workstations, Lisp machines operated without multi-user capability. Systems provided large monochrome bitmap displays, keyboards, mice, network adapters, local hard disks (exceeding 1 MB RAM), serial interfaces, and local bus extension card support. Optional color graphics cards, tape drives, and laser printers enhanced functionality.
The processor did not directly execute Lisp but operated as a stack machine with instructions optimized for compiled Lisp. Early systems used microcode for instruction provision. Hardware performed type checking and dispatching for various operations—a single addition instruction accommodated integers, floats, rationals, and complex numbers, producing highly compact compiled code representations.
Operating systems utilized virtual memory for expanded address spaces, with garbage collection managing memory. All code shared a single address space; all data objects carried type tags enabling runtime type determination. Multiple execution threads, termed "processes," ran within this unified address space.
All operating system software was written in Lisp—Xerox employed Interlisp while Symbolics, LMI, and TI utilized Lisp Machine Lisp (MacLisp descendant). Common Lisp support emerged subsequently, with system software ported or newly written in Common Lisp.
Later machines (TI MicroExplorer, Symbolics MacIvory, Symbolics UX400/1200) no longer constituted complete workstations but rather boards embedding in Apple Macintosh II or Sun-3/Sun-4 systems.
Some Lisp machines, particularly the Symbolics XL1200, featured extensive graphics capabilities via specialized graphics boards. These found application in medical image processing, 3D animation, and CAD domains.