Pressures the Radiation Transport Problem Places on Future PIM-Based Supercomputer Designs

Thomas Christopher

Abstract

For many algorithms that simulate physical processes by discretizing space and time, the data per cell remains constant, allowing PIM architectures to maintain the same ratio of cells per processor over generations of application size and of semiconductor technology. Radiation transport algorithms, however, also discretize the angles at which particles may move among cells and the energy levels they can occupy, and the cell size is proportional to their product. For large problems, the greatest need will be for data storage, rather than processors. We examine a large radiation transport algorithm, and using ITRS tables, project its performance on PIM systems, with and without external DRAM, and an architecture resembling the "Red Storm" computer. A PIM system without external DRAM can be expected to run an order of magnitude faster than the other systems, but unless PIM prices drop to at least an order of magnitude less than those of MPU chips, the cost effectiveness of the systems may be approximately the same. Even that, however, is assuming the PIM chip area devoted to processors is only 2.5% of the total.