Heterogeneous reconfigurable design for TreePM N-body simulation

摘要

In this paper, we describe our implementation for N-body simulation with the aid of FPGA, GPU and CPU. TreePM, which is a common algorithm for N-body simulation and meets our demands well, is chosen here. We split treePM algorithm into several parts, and assign them to different accelerators depending on the features of the operation. Moreover, partially dynamically reconfigurable FPGA architecture is used here for fine-grained scheduling, making the assignments more balanced. Experiments showed that it has improvement in speed and energy efficiency over previous design with the same scale. Besides, it's potential for advances with multiple clusters also been demonstrated here.