Êtes-vous un étudiant de l'EPFL à la recherche d'un projet de semestre?
Travaillez avec nous sur des projets en science des données et en visualisation, et déployez votre projet sous forme d'application sur GraphSearch.
Resistive Random Access Memory (RRAM)-based FPGAs are predicted to outperform conventional FPGAs architectures in area, delay and power over a wide range of voltage operations, allowing novel energy-quality trade-offs for reconfigurable computing. The opportunity lies in that RRAMs can realize the functionality of a Static Random Access Memory (SRAM) and a transmission-gate in a unique device. However, most of predictive analysis shown in the state-of-the-art are achieved by using analytical models. Unfortunately, while analytical models have been intensively refined for conventional FPGA architectures, their accuracy on RRAM-based FPGAs has not been carefully investigated. Consequently, misleading conclusions may be caused by using inaccurate analytical models. In this paper, we rely on electrical simulations and semi-custom design tools to perform detailed area and power comparison between SRAM-based and RRAM-based FPGAs. To enable accurate analysis, we develop a synthesizable Verilog generator for both SRAM-based and RRAM-based FPGAs and also enhance FPGASPICE to support most recent advanced RRAM-based circuits and FPGA architectures. The area analyses are based on fullchip layouts of SRAM-based and RRAM-based FPGAs, which are produced by a semi-custom design flow. We consider a full FPGA fabric, including core logic, configuring peripherals and I/Os, which is more realistic than analytical models. The power analysis are based on SPICE simulation results by considering the twenty largest MCNC benchmarks. Simulation results identify that the target RHRS of RRAM-based FPGAs should be at least 20M Ω to guarantee energy improvements over SRAM-based FPGAs. Experimental results present that at nominal working voltage, RRAM-based FPGAs can improve up to 8% in area, on average 22% in delay and on average 16% in power respectively, as compared to SRAM-based counterparts. Compared to SRAMbased FPGAs working at nominal voltage, near-Vt RRAM-based FPGAs can outperform close to 2 × in Energy-Delay Product without delay overhead. As a result, RRAM-based FPGAs are more capable of trading-off energy and quality than the SRAM-based counterparts.