PGA HPC implementation of microtubule brownian dynamics simulations

This paper presents high performance simulation of microtubule molecular dynamics implemented on Xilinx Virtex-7 FPGA using high level synthesis tool Vivado HLS. FPGA implementation is compared to multicore Intel Xeon CPU and Nvidia K40 GPU implementations in terms of performance and energy efficiency. Algorithm takes into account Brownian motion thus heavily uses normally distributed random numbers. Original sequential code was optimized for different platforms using OpenMP for CPU, OpenCL for GPU and Vivado HLS for FPGA. We show that in terms of performance FPGA achieved 17x speed up against CPU and 11x speedup against GPU for our best optimized CPU and GPU versions. As to power efficiency, FPGA outperformed CPU 227 times and GPU 75 times. FPGA application is developed using SDK, which has Board Support Package including FPGA project framework where accelerator kernel (designed in Vivado HLS) IP core is to be integrated, and host-side libraries used to communicate with FPGA via PCI Express. Developed flow does not require expert FPGA skills and can be used by programmer with little knowledge of hardware design methodology that could use C\C++ language for complete development of FPGA accelerated solution.