A Proposal for Research on VLSI Engines for Lattice Calculations

这项研究涉及设计和建造实用的、基于VLSI的格子计算定制计算机的问题。这些计算问题的特点是迭代的，定义在规则的点阵上，在空间和时间上一致，局部，并且相对简单。例如，微分方程组的数值解、迭代图像处理和细胞自动机。最近发明的格子气体自动机是流体动力学的微观模型，是这项工作的试验台。设想的机器--格子引擎--通常由一个定制芯片和一个用于支持的通用主机的许多实例组成。在许多实际情况下，这类机器的性能受到限制，不是受实际处理单元的速度和大小的限制，而是受片上和片外的通信带宽以及芯片的存储容量的限制。具体研究分三个阶段同时进行：(1)现有体系结构的性能分析；(2)扩展现有理论，给出作为特定格子的芯片I/O带宽和存储容量的函数的可实现吞吐速率的界限；(3)原型机器的构建和测试。适合于格子计算的计算问题的例子包括微分方程的数值解、迭代图像处理和细胞自动机。这些都是非常重要的问题。这个项目的重点是专门针对这类计算设计和建造实用的基于VLSI的定制计算机的问题，提供了比通用计算机更快的问题解决方案的可能性。

This research deals with the problems of designing and building practical, custom VLSI-based computers for lattice calculations. These computational problems are characterized by being iterative, defined on a regular lattice of points, uniform in space and time, local, and relatively simple. Examples include numerical solution of differential equations, iterative image processing, and cellular automata. The recently invented lattice gas automata, which are microscopic models for fluid dynamics, are a testbed for the work. The machines envisaged--lattice engines--would typically consist of many instances of a custom chip and a general-purpose host machine for support. In many practical situations, the performance of such machines is limited, not by the speed and size of the actual processing elements, but by the communication bandwidth on- and off-chip, and by the memory capacity of the chip. The specific research consists of three phases to be conducted concurrently: (1) performance analysis of existing architectures, (2) extension of current theory to give bounds on achievable throughput rate as a function of chip I/O bandwidth and memory capacity for the particular lattices of interest, and (3) construction and testing of prototype machines. Examples of computational problems amenable to lattice calculations include the numerical solution of differential equations, iterative image processing, and cellular automata. These are all very important problems. The focus of this project on the problems of designing and building practical custom VLSI-based computers aimed specifically at this class of calculations offers the potential of much faster problem solutions than on general-purpose computers.