Understanding Performance Differences of FPGAs and GPUs: (Abtract Only)

The notorious power wall has significantly limited the scaling for general-purpose processors. To address this issue, various accelerators, such as GPUs and FPGAs, emerged to achieve better performance and energy-efficiency. Between these two programmable accelerators, a natural question arises: which applications are better suited for FPGAs, which for GPUs, and why? In this paper, our goal is to better understand the performance differences between FPGAs and GPUs and provide more insights to the community. We intentionally start with a widely used GPU-friendly benchmark suite Rodinia, and port 11 of the benchmarks (15 kernels) onto FPGAs using the more portable and programmable high-level synthesis C. We provide a simple five-step strategy for FPGA accelerator designs that can be easily understood and mastered by software programmers, and present a quantitative performance breakdown of each step. Then we propose a set of performance metrics, including normalized operations per cycle (OPC_norm) for each pipeline, and effective parallel factor (effective_para_factor), to compare the performance of GPU and FPGA accelerator designs. We find that for 6 out of the 15 kernels, today's FPGAs can provide comparable performance or even achieve better performance, while only consume about 1/10 of GPUs' power (both on the same technology node). We observe that FPGAs usually have higher OPC_norm in most kernels in light of their customized deep pipeline but lower effective_para_factor due to far lower memory bandwidth than GPUs. Future FPGAs should increase their off-chip bandwidth and clock frequency to catch up with GPUs.