Performance and Energy Efficiency Analysis of Reverse Time Migration on a FPGA Platform

Reverse time migration (RTM) modeling is a computationally intensive component in the seismic processing workflow of oil and gas exploration, often demanding the manipulation of terabytes of data. Therefore, the computational kernels of the RTM algorithms need to access a large range of memory locations. However, most of these accesses result in cache misses, degrading the overall system performance. GPGPUs and FPGAs are the two endpoints in the spectrum of acceleration platforms, since both can achieve better performance in comparison to CPU on several high-performance applications. Recent literature highlights FPGA better energy efficiency when compared to GPGPU. The present work proposes a FPGA accelerated platform prototype targeting the computation of the RTM algorithm on an HPC environment. Experimental results highlight that speedups of 112x can be achieved, when compared to a sequential execution on CPU. When compared to a GPU, the power consumption has been reduced up to 55%.