An embedded DRAM architecture for large-scale spatial-lattice computations

Abstract

Spatial-lattice computations with finite-range interactions are an important class of easily parallelized computations. This class includes many simple and direct algorithms for physical simulation, virtual-reality simulation, agent-based modeling, logic simulation, 2D and 3D image processing and rendering, and other volumetric data processing tasks. The range of applicability of such algorithms is completely dependant upon the lattice-sizes and processing speeds that are computationally feasible. Using embedded DRAM and a new technique for organizing SIMD memory and communications we can efficiently utilize 1 Tbit/sec of sustained memory bandwidth in each chip in an indefinitely scalable array of chips. This allows a 10,000-fold speedup per memory chip for these algorithms compared to the CAM-8 lattice gas computer, and is about one million times faster per memory chip for these calculations than a CM-2.