Return to "Visible Storage"

*** Please note: This website (comp-hist) was completed before I found out about Wikipedia in 2002.
Since then I have added material occasionally.
Items are certainly not complete, and may be inaccurate.
Your information, comments, corrections, etc. are eagerly requested.
Send e-mail to Ed Thelen. Please include the URL under discussion. Thank you ***

Connection Machine -1-2-5

Manufacturer Thinking Machines Corporation
Identification,ID CM-1
Date of first manufacture1985
Number produced -
35 *
Estimated price or cost-
location in museum -
donor David Yakimischak, Dow Jones & Company, Inc.
Laboratory for Computer Science, M. I. T.
Laboratory for Computer Science, M. I. T.
* from

Contents of this page:

Connection Machine-1
Connection Machine-5


  • 12 dimension hypercube
  • Architecture
  • from
    For Danny Hillis, a student working on problems in human cognition at the Massachusetts Institute of Technology's Artificial Intelligence Laboratory in the late '70s, existing sequential supercomputers were simply inadequate for the problems that interested him. Even the fastest supercomputers were unable to recognize human faces, use language at the level of a 5-year-old child, or perform other tasks that humans, equipped with brains much slower than any supercomputer, could solve with ease. He became convinced that it was necessary to design a parallel computer with a structure closer to that of a human brain.

    In order to build the first of these new machines, Hillis helped found Thinking Machines Corporation in 1983, which introduced the CM-1 in 1986 and the higher performance version, the CM-2, in 1987. (Since the CM-2 quickly replaced the CM-1, being a faster version of the same computer architecture, as well as using the same external package, I will speak only of the CM-2 from now on.) These machines had 65,536 simple 1-bit processors that could simultaneously perform the same calculation, each on its own separate data set. For problems involving the separate but interrelated actions of many similar objects or units, such as movement of atoms, fluid flow, information retrieval, or computer graphics, this "data-parallel" structure brought tremendous increases in speed while also being easy to program. Many problems that seemed impossibly complex when analyzed with sequential logic fit naturally into a parallel data structure. (3)

    This type of massively parallel architecture had been tried before, but what enabled the CM-2 to succeed where other designs had failed was an extremely flexible and fast communications network between the processors. Using the model of the human brain, Hillis's design placed importance not so much on the processors themselves, but rather on the nature and mutability of the connections between them, hence the name "Connection Machine."

    many identical interconnected processors under the supervision of a single control unit, see figure 7.1.2. The control unit transmits the same instruction, simultaneously, to all processors.

    All the processing elements simultaneously execute the same instruction and are said to be 'lock-stepped' together. Each processor works on data from its own memory and hence on distinct data streams. (Some systems also provide a shared global memory for communications.) Every processor must be allowed to complete its instruction before the next instruction is taken for execution. Thus, the execution of instructions is said to be synchronous.

    This category corresponds to the array processors discussed in section 2.3.3 and examples include; ILLIAC-IV, PEPE, BSP, STARAN, MPP, DAP and the Connection Machine (CM-1).

  • The Connection Machine CM-1 has 65,536 simple 1-bit processors connected into a hypercube and each having 4Kbits of memory. Every processor is connected to a central unit called the 'microcontroller' which issues identical 'nanoinstructions' to all of them. This unit can be regarded as a control unit.
            Processors      Grain  Topology  Control Multiplicity
                V             S    Hypercube           V

Special features
  • It should be noted that the CM-1 is not very efficient when doing floating-point calculations. In the development it turned out that floating-point performance was very important for the commercial success of such a massively parallel computer.
  • Therefore in the CM-2, the memory size was increased (to 64K or 256K per processor) and special floating-point accelerators were added. These chips were added one for each 32 1-bit processors, corresponding to the 32-bit width of one floating-point variable.
  • In further development, the principle of using custom build hardware for the processors was abandoned, because commercial microprocessors were gaining in speed much faster than could be achieved with custom hardware. The next generation [CM-5] of connection machines used standard microprocessors, thus abandoning the SIMD principle.
  • By now (1996) Thinking Machines Corporation has abandoned the hardware development, and now focuses on software development for massively parallel computers constructed by other companies.

Historical Notes

Danny Hillis writes first description of the Connection Machine
architecture (appears as memo from Artificial Intelligence Lab at
MIT).  (BMB: TMC, Connection Machine)


DARPA starts Strategic Computing Initiative, which helps fund such
machines as Thinking Machines Connection Machine, BBN Butterfly, WARP
from Carnegie Mellon University and iWarp from Intel Corp.  (MW:
TMC demonstrates first CM-1 Connection Machine to DARPA.  (BMB: TMC,
Thinking Machines Corp. ships first Connection Machine CM-1 (up to
65536 single-bit processors connected in hypercube).  (GVW: TMC, CM-1)

TMC introduces CM-2 Connection Machine (64k single-bit processors
connected in hypercube, plus 2048 Weitek floating point units).  (GVW:
TMC, CM-2)

Gordon Bell Prize for absolute performance awarded to a team from
Mobil and Thinking Machines Corporation, who achieved 6 GFLOPS on a
CM-2 Connection Machine; prize in price/performance category awarded
to Emeagwali, who achieved 400 MFLOPS per million dollars on the same
platform.  (GVW: Gordon Bell Prize)

Gordon Bell Prize in price/performance category awarded to Geist,
Stocks, Ginatempo, and Shelton, who achieved 800 MFLOPS per million
dollars in a high-temperature superconductivity program on a 128-node
Intel iPSC/860; prize in compiler parallelization category awarded to
Sabot, Tennies, and Vasilevsky, who achieved 1.5 GFLOPS on a CM-2
Connection Machine with Fortran 90 code derived from Fortran 77.
(GVW: Gordon Bell Prize)
Thinking Machines Corporation produces CM-200 Connection Machine, an
upgraded CM-2.  MIMD CM-5 announced.  (BMB: TMC, CM-200)

This Artifact

Interesting Web Sites

Other information
From: Steve Pothier
Date: Wed, Nov 23, 2016
My wife asked you something about CM and found you were looking for stories. Here's mine:

In 1987 I was do some work in Artificial Intelligence. I became a big fan of Marvin Minsky. That led me to someone I had not heard of before: Danny Hillis. I bought Danny's PhD dissertation (about the Connection Machine) and read it. About that time I interviewed at a company that didn't seem all that appealing at the time: Science Applications International Corporation. During the interview, a manager mentioned that they probably had access to every kind of super-computer except a Cray in the office. I more or less mumbled: "probably not a Connection Machine". That manager heard (good ears!) and said: "Did you say Connection Machine? We have one downstairs. Do you want to see it.". I want from "I don't want this job" to "I'll sell my first-born to get it" in a flash! When I got an offer to work on the CM at SAIC I told my wife I would do it for free. She thought that was not the best negotiating position and suggested I keep that idea to myself. Smart lady.

When was hired, I was a "Connection Machine" expert despite the fact that the only thing I new about it was from reading Danny's dissertation. I did make that clear to SAIC but as is often the case "expert" just means "knows more than anyone else in the room". Before SAIC, I was programming in Lisp and the CM was programmed in "*Lisp" so it was a great fit. I spent several years doing my dream job. It was very difficult to pull myself away and go home at the end of the day. Our use of the CM was through DARPA. Eventually contracts moved on, "our" CM was taken away. I did some other CM work on a remote machine for awhile after before switching to more traditional architectures.

If you hear of any actual CM-2's that are up an running anywhere, I would love to know about them! I still long to do that kind of data-parallel programming and would consider just about anything to be able to do it again.

If you have comments or suggestions, Send e-mail to Ed Thelen

Go to Antique Computer home page
Go to Visual Storage page
Go to top

Updated February 3, 2003, Nov 2016