$105/Gflops Astrophysical N-Body Simulation with Reconfigurable Add-in Card and Hierarchical Tree Algorithm
Atsushi Kawai, Toshiyuki Fukushige · 2006
As an entry for the 2006 Gordon Bell price/performance prize, we report an astrophysical N-body simulation per-formed with reconfigurable computing systems and a hierar-chical tree algorithm. Each of our systems consists of a host PC and a reconfigurable add-in card attached via high-speed interface, PCI Express or PCI-X. The reconfigurable add-in card houses one FPGA chip, into which we integrated 16 pipeline processors specialized for gravitational force calcu-lation. Other operations, such as tree construction, tree tra-verse and time integration, are performed on the host PC. On our system, we performed a cosmological N-body simulation with 2.1 million particles, which sustained a performance of 15.39 Gflops averaged over 4.33 hours. The price of our sys-tem is $2,384 in total and, price per performance obtained is $158/Gflops, which is 44 times better than that obtained by