CAM-8: A computer architecture based on cellular automata
Norman Margolus · 1995
The maximum computational density allowed by the laws of physics is available only in a format that mimics the basic spatial locality of physical law. Fine-grained uniform computations with this kind of local interconnectivity (Cellular Automata) are particularly good candidates for efficient and massive microphysical implementation. Conventional computers are ill suited to run CA models, and so discourage their development. Nevertheless, we have recently seen examples of interesting physical systems for which the best computational models are cellular automata running on ordinary computers. By simply rearranging the same quantity and quality of hardware as one might find in a low-end workstation today, we have made a low-cost CA multiprocessor that is about as good at large CA calculations as any existing supercomputer. This machine’s architecture is scalable in size (and performance) by orders of magnitude, since its 3D spatial mesh organization is indefinitely extendable. Using a relatively small degree of parallelism, such machines make possible a level ∗ This research was supported by the Advanced Research Projects Agency, grant N0014-89-J-1988. of performance at CA calculations much superior to that of existing supercomputers, but vastly inferior to what a fully parallel CA machine could achieve. By creating an intermediate hardware platform that makes a broad range of new CA algorithms practical for real applications, we hope to whet the appetite of researchers for the astronomical computing power that can be harnessed in microphysics in a CA format.