3D FFTs on a Single FPGA.

Area of Science:

• Computational Physics
• Computer Engineering
• Scientific Computing

Background:

• The 3D Fast Fourier Transform (FFT) is a fundamental computational kernel.
• FPGA implementations of 3D FFTs were previously considered inefficient compared to alternatives like multi-grid methods.
• Efficient 3D FFT computation is crucial for large-scale physical simulations and image processing.

Purpose of the Study:

• To investigate the efficiency of a simple FPGA-based 3D FFT design.
• To compare the performance of FPGA 3D FFTs against GPU implementations.
• To assess the viability of FPGAs for accelerating scientific computing tasks, specifically molecular dynamics (MD).

Main Methods:

• Implementation of a straightforward 3D FFT algorithm on an FPGA.
• Performance evaluation at a conservative operating frequency.
• Benchmarking against a contemporary Nvidia GPU for comparable data sizes.
• Analysis of execution times for 16^3, 32^3, and 64^3 single-precision data points.

Main Results:

• The FPGA 3D FFT achieved execution times of 4μs (16^3), 21μs (32^3), and 215μs (64^3).
• Performance for 16^3 and 32^3 data points favorably compared to a Nvidia GPU's 25μs and 29μs, respectively.
• The results challenge the notion that FPGAs are inefficient for 3D FFT computations.

Conclusions:

• A simple FPGA design provides a highly efficient 3D FFT implementation.
• FPGA-based 3D FFTs are competitive with, and in some cases superior to, GPU solutions.
• This work establishes FPGAs as a key component for building large-scale, FPGA-based molecular dynamics engines, potentially removing FFTs from the simulation critical path.