3D FFT for FPGAs

The 3D FFT is critical in electrostatics computations such as those used in Molecular Dynamics simulations. On FPGAs, however, the 3D FFT was thought to be inefficient relative to other methods such as convolution-based implementations of multigrid. We find the opposite: a simple design using less than half the chip resources, and operating at a very conservative frequency, takes less than 50us for 32³ and 200us for 64³ single precision data points, numbers similar to the best published for GPUs. The significance is that this is a critical piece in implementing a large scale FPGA-based MD engine: even a single FPGA is capable of keeping the FFT off of the critical path for a large fraction of possible MD simulations.