Optimal dynamic data layouts for 2D FFT on 3D memory integrated FPGA

Abstract

FPGAs have been widely used for accelerating various applications. For many data intensive applications, the memory bandwidth can limit the performance. 3D memories with through-silicon-via connections provide potential solutions to the latency and bandwidth issues. In this paper, we revisit the classic 2D FFT problem to evaluate the performance of 3D memory integrated FPGA. To fully utilize the fine grained parallelism in 3D memory, optimal data layouts so as to effectively utilize the peak bandwidth of the device are needed. Thus, we propose dynamic data layouts specifically for optimizing the performance of the 3D architecture. In 2D FFT, data is accessed in row major order in the first phase whereas, the data is accessed in column major order in the second phase. This column major order results in high memory latency and low bandwidth due to high row activation overhead of memory. Therefore, we develop dynamic data layouts to improve memory access performance in the second phase. With parallelism employed in the third dimension of the memory, data parallelism can be increased to further improve the performance. We adopt a model based approach for 3D memory and we perform experiments on the FPGA to validate our analysis and evaluate the performance. Our experimental results demonstrate up to 40x peak memory bandwidth utilization for column-wise FFT, thus resulting in approximately 97% improvement in throughput for the complete 2D FFT application, compared to the baseline architecture.