Supercomputing with VLSI
Supercoprocessors (SCPs), highly parallel VLSI architectures tuned to solving a specific problem class, are shown to provide a means of cost-effective supercomputing. A methodology for building SCPs for different computation-intensive problems is described: two pragmatic constraints namely problem-size-independence and limited-bandwidth constraint are imposed on special-purposes architectures; a simple but powerful model of computation is used to derive general upper bounds on the speedup obtainable using such architectures. It is shown that bounds established by other authors for matrix multiplication and sorting, using problem-specific approaches, can be derived very simply using this model. Poisson Engine-I (PE-I), a prototype SCP, is a system for solving the Laplace equation using the finite difference approximation. PE-I uses a novel approach: asynchronous iteration methods are implemented using a fixed-size, synchronous array of simple processing elements. Architectural and algorithmic extensions to PE-I are briefly considered: the solution of a wider class of PDEs and the use of more sophisticated algorithms like the multigrid method are some of the issues addressed. The SCP methodology is applied to the problems of matrix-multiplication and sorting. For sorting, an SCP with superlinear speedup is outlined. For the matrix problem, the architecture and implementation details of SMP are described in detail. SMP, realized with about fifty chips using current technology, is capable of through-puts greater than 150 Mflops, and is also unique in being optimal with respect to the lowerbound derived using the SCP model. The use of a collection of sup SCPs is advanced as a cost-effective supercomputing alternative.
- Research Organization:
- Brown Univ., Providence, RI (USA)
- OSTI ID:
- 5922874
- Resource Relation:
- Other Information: Thesis (Ph. D.)
- Country of Publication:
- United States
- Language:
- English
Similar Records
VLSI design and synthesis for a class of two-dimensional problems
Area-time efficient and fault-tolerant VLSI arrays for digital signal processing
Related Subjects
SUPERCOMPUTERS
COMPUTER ARCHITECTURE
ALGORITHMS
ARRAY PROCESSORS
COST
FINITE DIFFERENCE METHOD
INTEGRATED CIRCUITS
LAPLACE EQUATION
MATRICES
PARALLEL PROCESSING
SORTING
USES
COMPUTERS
DIFFERENTIAL EQUATIONS
DIGITAL COMPUTERS
ELECTRONIC CIRCUITS
EQUATIONS
ITERATIVE METHODS
MATHEMATICAL LOGIC
MICROELECTRONIC CIRCUITS
NUMERICAL SOLUTION
PARTIAL DIFFERENTIAL EQUATIONS
PROGRAMMING
990200* - Mathematics & Computers