Efficient Coarse-Grained Reconfigurable Array architecture for machine learning applications in space using DARE65T library platform

Abstract

With the increasing use of satellites, rovers, and other space exploration devices, Artificial Intelligence (AI) is also becoming an important tool for space exploration, allowing autonomous decision-making and operations in harsh environments. As a result, there is an increasing demand for reliable and energy-efficient processing platforms in the space industry. Among all processing architectures, Coarse-Grained Reconfigurable Arrays (CGRAs) are becoming popular, particularly in data-intensive applications like machine learning, demonstrating a substantial improvement in the energy efficiency of inference operations while preserving a good degree of versatility. In high-level class space missions, the hardware platforms incorporate radiation-hardened Field Programmable Gate Arrays (FPGAs) and microcontrollers, which do not meet the performance requirements for the aforementioned AI applications. The use of CGRA architectures in space missions is still not widely studied. The main contribution of this work is a comprehensive Design Space Exploration (DSE) activity with our highly parameterized CGRA architecture, exploring the costs associated with various design parameters when targeting AI in the space domain. We evaluated performance, power consumption, and area occupation after synthesis on the radiation-hardened DARE65T standard cell library developed by imec, based on a commercial 65 nm technology process. We characterize different CGRA configurations, comparing them with state-of-the-art solutions used for the acceleration of the AI algorithms. This work highlights Performance, Power, and Area (PPA) results that range from $\mathrm{100}\mathrm{MHz}$ (up to $\mathrm{600}\mathrm{MOps}$), $\mathrm{2.43}\times {\mathrm{10}}^4\mu m^2$ cell area occupation and $\mathrm{0.699}\mathrm{mW}$ power consumption, to $\mathrm{625}\mathrm{MHz}$ (up to $\mathrm{3.75}\mathrm{GOps}$), $\mathrm{2.43}\times {\mathrm{10}}^5\mu m^2,\mathrm{46.5}\mathrm{mW}$. During DSE activity, we highlight the optimal solutions in terms of area efficiency (up to $\mathrm{313.1}{\mathrm{GOps/mm}}^2$) and energy efficiency (up to $\mathrm{289}\mathrm{GOps/W}$) of each CGRA configuration.