It will be argued that optimizing cost-effectiveness in parallel computer design must be performed in respect of a target application. This paper quantitatively examines the design and implementation of a simple Virtual Reality application for a SIMD architecture. A uniprocessor version was designed, implemented, and analyzed. Parallel implementations were them proposed. The resulting SIMD computations were then simulated on a detailed model which allows parametric variation of the physical characteristics of the computer's subsystems. Hardware cost tradeoffs were measured against performance in simulation. Design choices and results are presented. The paper concludes with an examination of the significance of these results and the possibilities for future research.
展开▼