Cost/performance trade-off in floating-point unit design for 3D geometry processor

Abstract

Geometry processing in three dimensional (3D) graphics application is characterized by a large amount of inherent parallelism and floating-point instructions. This processing is accelerated with multiple geometry processors that have fast floating-point unit (FPU). There are many design alternatives in the geometry processor design that are suitable for multiple configurations. With these alternatives, designers have to consider design cost and complexity. In this paper design considerations and trade-off factors are evaluated with floating-point arithmetic unit organization and implementation. First, geometry-processing steps are described and consideration factors are summarized to find design considerations of FPU for geometry processing Then, based on these design considerations, implementation trade-off factors are evaluated. In addition, floating-point division algorithms and their implementation are evaluated in the point of trade-off. Among the design alternatives for floating-point arithmetic units, the best organization with minimal investment is separate adder/multiplier and radix-16 SRT divider. And split register file permits area saving and instruction issue rate increase. In the processing of whole geometry pipeline stages, 45.5% of execution time improvement is achieved with this configuration. It is a cost-effective design. In addition, execution time and throughput trade-off must be considered for high-end 3D graphics system design.