Computing Models for FPGA-Based Accelerators with Case Studies in Molecular Modeling

Field Programmable Gate Arrays (FPGAs) are widely being considered as accelerators for compute intensive applications – and not just as add-on components as been possible for many years, but also tightly integrated with the main processor. This interest has been motivated, in part, by reported speedups of . This performance, however, have sometimes failed to translate to production applications. Moreover, the unique issues in programming FPGAs has left some application writers waiting for better tools before giving them serious consideration. In our work we have found that a critical phase of FPGA application development is finding, and mapping to, the appropriate computing model. This leads to the central point of this article: that there are several such models, and that they differ significantly from models generally used in programming. For example, whereas parallel computing models are often based on thread execution and interaction, FPGA computing can take advantage of additional degrees of freedom than available in software. This enables models based on the fundamental characteristics from which FPGAs get their capability, including highly flexible fine-grained parallelism and associative operations such as broadcast and collective response. After a review of FPGA architecture, systems, and constructs, we present five such models. The final part of the paper illustrates their use in applications involving molecular modeling. A version of this work has been accepted for publication in Computing in Science and Engineering. This work was supported in part by the NIH through awards #R21-RR020209-01 and and #R01-RR023168-01A1, and facilitated by donations from XtremeData, Inc., SGI, and Xilinx Corporation. Web: http://www.bu.edu/caadlab. EMail: herbordt@bu.edu. Currently with The MathWorks, Inc. Currently with Altera, Inc. Currently with MIT Lincoln Laboratory