A new floating-point architecture for wireless 3D graphics

We present a new ARM architecture designed to speed up geometry computations for wireless 3D graphics. Our motivating problem is the graphics geometry pipeline, and we begin by examining its fundamental operations: matrix manipulations, dot products, cross products, distance computation and divisions. The operations are used to derive a minimal but powerful SIMD instruction set. The difficulty of vectorizing some of the operations leads to a novel solution using new load/store instructions. We conclude with examples showing how the graphics problem maps onto a possible implementation, and a discussion of the speedups provided over ARM's general-purpose floating-point processor.