Thinking Machines Corporation (1983 - 1994)

1983 - 1994

dead us

danny hillis wanted a computer built from sixty-five thousand tiny processors instead of one large fast one, and thinking machines turned that thesis into a black cube whose front panel blinked red while it worked. for a few years around 1990 it held the top of the supercomputer list, with richard feynman having worked on the router that held it together. then washington was accused of favoring one supplier and the government orders stopped. the company filed chapter 11 on 17 august 1994 and the cube carried on as a movie prop.

~ the obituary ~

Thinking Machines was four years dead before I plugged in my first modem. Everything I know about it I read afterward, in the way you read about a building that was demolished in a city you have never visited. It belongs to a period of computing that Turkey experienced only as imported magazines. to a market, government supercomputing, that had no local edition at all.

this is mostly told from the outside. more or less. The company started in may 1983 in Waltham, Massachusetts, founded by Sheryl Handler and Danny Hillis to build the machine Hillis had described in his MIT doctoral work. for some reason. The argument was that a computer should not be one fast processor doing everything in sequence. It should be an enormous number of very simple processors doing one small thing each at the same moment, wired together well enough to pass results around. more or less. The first Connection Machine had sixty-five thousand of them, each one bit wide.

i literally was not paying attention by then, but anyway (cant remember the exact year). more or less. Richard Feynman spent time there as an early corporate fellow and worked on the analysis of the router, the part that decided how messages found their way between processors. back then. That is the detail everyone repeats about the company. it is repeated because it is true and because it says something accurate about what the place was. to be honest. Thinking Machines recruited physicists into a hardware startup and gave them the interesting half of the problem (roughly).

~ the case file ~

Born	may 1983, Waltham, Massachusetts
Peak	$65 million in sales in 1990, four of the world's fastest machines by 1993
Killed	17 august 1994
Lifespan	eleven years
Made by	Danny Hillis and Sheryl Handler
Killed by	government buyers backing away from their favorite supplier

~ the climb ~

this happened more in the english speaking internet than where i was. to be honest. The CM-1 shipped in 1985 and the CM-2 in 1987, both of them the black cube with the grid of red status lights on the front, one light per group of processors, the whole face flickering as a job ran. It is the only supercomputer of that era with a shape people outside the field can still draw. The CM-5 in '91 changed the architecture underneath to independent processors running commodity SPARC chips, which was the honest answer to where the industry had gone.

In 1990 the company booked sixty-five million dollars in hardware and software sales and led the parallel supercomputer market outright. By '93 four of the fastest computers in the world were Connection Machines. That is as close to winning as this kinda company gets.

~ how it actually ended ~

this part you had to be in the US for, and i was not. back then. The money came from Washington. DARPA and the Department of Energy bought the machines.

by 1991 the complaint inside that world was that the agencies were unfairly propping up one vendor at the expense of Cray and everyone else building the same thing. The purchases were cut back. anyway. Export controls on high-end systems narrowed what was left of the commercial market at the same time.

this part i was not really around for, but anyway. A company selling million-dollar machines to a handful of buyers has no cushion when two of those buyers step back. Thinking Machines filed a voluntary Chapter 11 petition on 17 august '94 in the bankruptcy court in Massachusetts, eleven years after it opened. Sun Microsystems took the hardware business and then the parallel software group in december 1996. though, oracle bought what was left of the data mining work in 1999. to be honest. back then. Nothing survived long enough to be renamed.

~ the funny part ~

i wasnt close to this part of the story, still, the most widely seen Connection Machines never computed anything. Jurassic Park put CM-5 cabinets in the control room in 1993 as non-functional props, red lights blinking on a set, more people than watched and that footage every real job those machines ever ran. though, mission: Impossible name-checked the company three years later.

this is mostly told from the outside (or whatever it was called). to be honest. So the shape survived and the argument survived, while the company that made both of them did not. Anyone who has seen the film has seen the machine and does not know it.

~ the long tail ~

i was not paying attention by then (cant remember the exact year). by then. anyway, hillis was right, which is the part that stings. Every graphics card in every desk today is a very large number of small processors working at the same instant on pieces of one problem. the entire machine learning industry sits on top of that idea. really though, the thing Thinking Machines could not survive was being right in '85 with hardware that cost a million dollars, in a market with about six customers, all of them funded by the same government.

this happened more in the english speaking internet than where i was (if memory serves). look, the name got taken up again in '25 by an unrelated AI startup, which is how these things go (give or take). by then. The original was a cube in a dark room with a face full of red lights. it lost because the people who bought it were told to stop (close enough).

« previous  ·  random  ·  next »