Connection Machine

Connection Machine

The Connection Machine was a family of massively parallel supercomputers developed by Thinking Machines Corporation and led by designers including Danny Hillis and Severo Ornstein. It emerged in the 1980s alongside developments at MIT, Bell Labs, NASA, DARPA, and Lawrence Livermore National Laboratory, influencing later projects at Cray Research, IBM, and Intel. The machines were notable for their use in research at institutions such as Stanford University, Harvard University, Caltech, and Los Alamos National Laboratory.

History

Development began in the early 1980s when Danny Hillis proposed an architecture motivated by research conducted at MIT Artificial Intelligence Laboratory and informed by earlier work at Xerox PARC and Bell Labs. Funding and partnerships involved National Science Foundation, DARPA, and industry collaborators including Texas Instruments and Massachusetts Institute of Technology Lincoln Laboratory. The first commercial systems were sold by Thinking Machines Corporation to universities and laboratories including Los Alamos National Laboratory, Lawrence Livermore National Laboratory, and NASA Ames Research Center. Throughout the late 1980s and early 1990s the project intersected with efforts at Stanford Linear Accelerator Center and standards groups such as IEEE. Financial and market pressures in the changing supercomputer industry, competition from firms like Cray Research and strategic moves by IBM, led to the company's bankruptcy and cessation of production in the mid-1990s.

Architecture

The hardware architecture employed a highly parallel topology inspired by research at MIT Media Lab and theoretical work from Carnegie Mellon University. Early models used bit-serial processors arranged in a hypercube and later models adopted a two-dimensional mesh and fat-tree variants informed by networking research at Bell Labs and AT&T Research. Processor design drew on microprocessor advances from Motorola and Texas Instruments, while communication subsystems referenced designs from Xerox PARC and Sun Microsystems. Memory was distributed across thousands of small SRAM modules, reflecting semiconductor technology trends at Intel and AMD. The system cabinets and cooling strategies were influenced by datacenter practices at IBM Research and Hewlett-Packard. The I/O subsystems and host interfaces connected to workstations from Sun Microsystems and Silicon Graphics.

Programming Model and Software

Programming models combined ideas from functional languages developed at University of Cambridge and MIT, and from parallel language research at Stanford University and Carnegie Mellon University. Languages and tools included variants of Lisp used at Xerox PARC, parallel extensions influenced by work at University of California, Berkeley, and compilers drawing on optimizations studied at Princeton University and University of Illinois Urbana–Champaign. The runtime and operating environment connected to host systems such as Unix System V and workstation environments from Sun Microsystems and Silicon Graphics; system software design consulted methods from Bell Labs Research. Researchers from Brown University and Yale University developed libraries for scientific computing, while visualization interfaces were inspired by projects at Ivory Tower National Laboratory and NASA Jet Propulsion Laboratory.

Applications and Performance

The machines were applied to computational fluid dynamics projects at NASA Glenn Research Center and Los Alamos National Laboratory, to molecular modeling at Harvard University and Stanford University, and to artificial intelligence research at MIT and Carnegie Mellon University. Image processing work drew interest from NASA Ames Research Center and Lawrence Livermore National Laboratory, while pattern recognition collaborations involved teams at Bell Labs and AT&T Research. Performance benchmarks referenced contemporary comparisons with systems from Cray Research and later IBM servers; scaling studies were published with collaborators from University of California, Berkeley, Caltech, and Princeton University. Users reported strong throughput on data-parallel workloads in projects funded by NSF and DARPA.

Variants and Successors

Variants included early bit-serial hypercube machines and later mesh-based models influenced by networking research at Bell Labs and AT&T Research. Successor concepts fed into architectures at Cray Research and commercial products developed at IBM and Intel Research. Academic spin-offs and research groups at MIT Media Lab, Stanford University, University of California, Berkeley, and Carnegie Mellon University extended ideas into multicore and GPU designs, informing projects at NVIDIA and AMD. Legacy efforts influenced parallel programming models taught at Massachusetts Institute of Technology and University of Cambridge, and hardware ideas resurfaced in exascale research at Oak Ridge National Laboratory and Argonne National Laboratory.

Category:Supercomputers