Connection Machine - History

History

Danny Hillis and Sheryl Handler founded Thinking Machines in Waltham, Massachusetts (it was later moved to Cambridge, Massachusetts) in 1983 and assembled a team to develop the CM-1 Connection Machine. This was a "massively parallel" hypercubic arrangement of thousands of microprocessors, each with its own 4 kbits of RAM, which together executed in a SIMD fashion. The CM-1, depending on the configuration, had as many as 65,536 processors. The individual processors were extremely simple, processing one bit at a time.

The CM-1 and CM-2 took the form of a cube 1.5 meters on a side, divided equally into eight smaller cubes. Each sub-cube contained 16 printed circuit boards and a main processor called a sequencer. Each printed circuit board contained 32 chips. Each chip contained a communication channel called a router, 16 processors, 16 RAMs. The CM-1 as a whole had a hypercubic routing network, a main RAM, and an input/output processor. It was connected to a switching device called a nexus.

In order to improve its commercial viability, the CM-2, launched in 1987, added Weitek 3132 floating-point numeric co-processors and more RAM to the system. 32 of the original one-bit processors shared each numeric processor. The CM-2 could be configured with up to 512 MB of RAM, and a RAID hard disk array, called a DataVault, of up to 25 GB.

Two later variants of the CM-2 were also produced, the smaller CM-2a with either 4096 or 8192 single-bit processors, and the faster CM-200.

Due to its origins in AI research, the software for the CM-1/2/200 single-bit processor was influenced by the Lisp programming language and a version of Common Lisp, *Lisp (spoken: "Star-Lisp"), was implemented on the CM-1. Other early languages included Karl Sims' IK and Cliff Lasser's URDU. Much system utility software for the CM-1/2 was written in *Lisp. Many applications for the CM-2, however, were written in C*, a data-parallel superset of ANSI C.

With the CM-5, announced in 1991, Thinking Machines switched from the CM-2's hypercubic architecture of simple processors to an entirely new MIMD architecture based on a fat tree network of SPARC RISC processors. The later CM-5E replaced the SPARC processors with faster SuperSPARCs. As of November 2012, the fastest system in the TOP500 list, the Titan with Rpeak of 27.1125 PFlop/s, is over 206,965 times faster than the fastest system in November 1993, the Connection Machine CM-5/1024 (1024 cores) with Rpeak of 131.0 GFlop/s.