wm@ogccse.ogc.edu (Wm Leler) (10/08/88)
COGENT RESEARCH DELIVERS PARALLEL SUPERCOMPUTER IN "PC PACKAGE" Beaverton, Oregon -- October 3, 1988 -- Cogent Research, Inc. today introduced a modular parallel processing supercomputer that can be user-configured to provide over 45 MFLOPs (millions of floating point operations per second) and 150 VAX MIPs (millions of instructions per second) in a PC-sized package. Called the XTM, Cogent's Transputer-based desktop hardware can be reconfigured from 2 to 400 processors to provide a wide range of processing, user interface and input/output capabilities. The XTM also employs the Linda software environment developed at Yale University. "The Linda environment truly represents a technology breakthrough. In a single stroke, it eliminates many of the software hurdles that have limited widespread acceptance of parallel processing," said Charles Vollum, Cogent's president and co-founder. "Users can invoke Linda by adding four simple statements to familiar programming languages. They can choose the language best suited to their application -- LISP for AI, Ada for embedded systems, Fortran for scientific programs, C and C++ for systems work." By providing automatic task allocation and automatic process synchronization, Linda frees the programmer from dealing with the details of the underlying hardware. This significantly reduces development time and results in programs that can run without change on many different architectures with varying numbers of processors. Unix library compatibility allows straightforward porting of existing sequential software. Parallel applications developed on the XTM are easily portable: Linda implementations exist on such systems as the Sequent Balance and Symmetry, Encore Multimax and Intel iPSC. The XTM's dynamically reconfigurable architecture is achieved via a unique parallel bus structure and intelligent switch, which Cogent is patenting. The bus structure provides high-speed communication between processors and takes advantage of the transputer's built-in communication links. Each XTM contains two INMOS T800-20 transputers, VLSI microprocessors with enhanced RISC instruction sets, 4 Kbytes of on-chip RAM, links for point-to-point communications to other transputers, and 4 Mbytes of local RAM. The T800-20 has a throughput of 5 MIPs/1.5 MFLOPs. The XTM's extensible architecture employs high-speed fiber optic interconnects that allow computation and storage resources to be efficiently shared regardless of their physical location. "The XTM is general purpose enough to span a broad range of applications," commented Vollum. "It's particularly well-suited for financial and scientific data analysis, transaction processing, image and signal processing, commercial graphics, factory automation, modeling and simulation -- from neural networks to environmental studies." Pricing, which is dependent on configuration, starts at $19,800 for a two-processor system with three MFLOPs and 10 MIPs. A 30-processor system with 45 MFLOPs and 150 MIPs sells for $225,300. Many configurations are available. The XTM is available immediately. Cogent Research, founded in August 1986, is located in the Oregon Graduate Center Science Park at 1100 NW Compton Drive, Beaverton, OR 9706-6998, (503) 690-1450. ____________________ LindaBus is a trademark of Cogent Research, Inc. VAX is a trademark of Digital Equipment Corporation Transputer is a trademark of Inmos Corporation Balance and Symmetry are trademarks of Sequent Computer Systems, Inc Multimax is a trademark of Encore Computer Corporation iPSC is a trademark of Intel Corporation Technical Overview - Linda Linda is based on the use of tuples, or collections of related data. Elements of a tuple are fields that contain actual values, or formals. These tuples exist in an abstract space called a tuple space, which can span multiple processors. Processes exchange messages by adding or removing tuples from tuples space. Linda is implemented in conventional programming through the addition of four new statements. These are: in, rd, out, and eval. In removes a tuple from tuple space for use by a process. Rd is similar to in except it reads the data in the tuple without removing it from tuple space. The out operation creates a passive tuple and places it into tuple space. Finally, eval creates an active tuple, one that evaluates its data upon entry into the tuple space and is then available for consumption or reading by other active tuples. To run a Linda program, an active tuple is dropped into tuple space where it can create other active tuples. The active tuples, which are executing simultaneously, exchange data by generating, reading and consuming passive tuples. An active tuple that is finished executing becomes a passive tuple. Technical Overview - Lindabus The Lindabus (patent pending) communications strategy is a hybrid architecture consisting of a parallel bus and an intelligent switch. In the XTM, processors have access to two types of communication: via a 32-bit parallel bus and through the Transputer's high speed serial links. A processor places a connection request on the 32-bit parallel bus, where it is received by the switch controller. The controller then connects the requesting processor and the supplying processor via their high speed serial links. The result is that any processor enjoys direct connection with any other processor within an XTM system. The parallel bus can also be used for broadcast/multicast communication by the processors.