Thinking Machine/Connection Machines

Thinking Machines Corporation - Connection Machines

Connection Machines and their structures

The CM-1 is a SIMD machine. In SIMD, one copy of the program controls numerous processors. The data is distributed, and the processors execute a set of instructions in lockstep fashion. Processors are organized in a 12-dimensional hypercube with 16 processors at each vertex of the hypercube.

The CM-2 incorporates 65,536 simple processors. Each processing unit contains 4,096 bits of memory. There are 16 processors on a single chip, and 32 chips are put on a single printed circuit board. Overall, there are 128 boards. The 16 processors on each chip are connected by a switch which allows for a direct connection between any pair of processing units. The routing device of each chip is attached to 12 other routing devices in the whole system. Communication between different processors can be performed through many routes, and this facilitates fast transmission of data even if a route is occupied. The average speed of the CM-2 for most applications is two billion operations per second.

The CM-5 is the most widely used Connection Machine at present. It consists of processing, control, and I/O nodes and an interface connected by two scalable networks which handle communication of data and control information. Each processing node has a 64-bit RISC microprocessor, 32 MB of memory, and interface to the control and data interconnection networks. The CM-5 can have up to 16,384 processors, but the existing ones have from 32 to 1,024 processors. The memory extends from 1 gigabyte to 32 gigabytes at present and can be maximized at 512 GB. The three main features of the CM-5 are:

Scalability - The current CM-5, including networks, clocking, I/O system, and software, is designed to scale up to 16,384 processors. This means that number of processors, amount of memory, and the I/O capacity can be increased on-site, and no change in the software is necessary.
Distributed memory/Global addressing - Memory is distributed for fast local access. All of the memory is accessible by each processor through a network. The cost-effectiveness, speed, and the efficiency of distributed memory is not sacrificed.
Distributed execution/Global synchronization - Instructions are not implemented in lockstep fashion always. The CM-5 has this feature. Each processor gets its instruction from a local instruction cache and chooses to proceed only at a certain point is reached by the other processors. This allows for fast synchronization and supports the sequential control flow. Each data parallel operation is completed before the next begins.

The CM-5 is a MIMD machine although it has many features found only on SIMD machines. The processors can operate independently when synchronization of instrutions is not essential. A fat-tree topology is used instead of the hypercube of CM-2 and permits better scalability. The system supports a Unix-based operating system in which time-sharing and priority-based job queuing are possible.
Image of CM-5

Applications of Connection Machines

Connection Machines have been mainly used in research fields for the government and in international business as well as in universities for educational purposes. The CM-5 is used by NASA's Jet Propulsion Laboratory (JPL) to process data from space shuttle expeditions into high resolution color images of the earth's surface. NASA is also utilizing the CM-5 to solve Grand Challenge Problems.

About Thinking Machines Corporation

Thinking Machines Corporation is the producer of the Connection Machines. It specializes in parallel techniques, software, and systems solutions for parallel computing problems. Its headquarters is in Bedford, MA.

Report by: Anurag Chandra
Date: 16 May 1996
Acknowledgements:

"The CM-5 Connection Machine: A Scalable Supercomputer" by W. Daniel Hillis and Lewis W. Tucker. Communications of the ACM, November 1993, Vol. 36, No. 11.