| While molecular dynamics simulations (MD) are a fundamental method for gaining the understanding of chemical and biological systems, their computational cost is extremely high: Simulating macromolecules requires thousands of node hours and cell-level systems remain altogether out of reach. We address this issue by using an emerging mode of high performance computing that is based on configurable logic in the form of Field Programmable Gate Arrays (FPGAs). The problem is that, while FPGAs have often delivered 100-fold speed-ups per node over microprocessor-based systems, the applications have generally been limited to those with small regular kernels operating on low-precision integer data types. MD possesses neither. We address this problem by creating an explicitly designed FPGA-coprocessor that can be integrated into generic commercially available systems.; MD is an iterative technique: the forces on each particle are computed, then applied using the equations of motion. We use standard partitioning by computing bonded forces, motion updates, and bookkeeping on the host, while computing the remaining forces (which dominate) on the FPGA accelerator. For the short-range forces we combine the following: cell lists, systematically determined interpolation and precision, handling of exclusion, and support for models with large numbers of particles. This has required new microarchitectures for the cell list processor and off-chip memory controller; and extensive experimentation to explore the design space to optimize precision, interpolation order, interpolation mode, table sizes, and simulation quality. To perform efficient and accurate numerical computation on FPGA, we created a novel arithmetic mode that is tuned for computing high order polynomial interpolation. For the long-range forces we use the multigrid method: we show that this is an excellent match to FPGAs with the primary operations having a favorable systolic structure and taking advantage of the large number of independently addressable RAMs.; The significance of this work lies at several levels: the 5x to 10x acceleration of MD production code while retaining simulation quality; the turings of algorithms for the FPGA; the system integration; the new arithmetic mode; the numerous novel microarchitectures; and the methods for optimizing MD implementations on FPGAs. |
