More FLOPS, less mass: enabling future AI missions in Space

Space-based artificial intelligence (AI) is advancing rapidly with compute efficiency rapidly becoming a critical factor in measuring system capabilities. This abstract analyzes AI compute efficiency—measured in floating-point operations per second per kilogram (FLOPS/kg)—across platforms from 3U CubeSats (~4 kg) to 1000 kg satellites. Historically, larger spacecraft relied on radiation-hardened processors that had very limited compute (a couple hundred MIPs). Recent advances in edge compute have led to the ability of nanosatellites to have gigaflop to teraflop compute capabilities. This comparison of onboard edge AI compute will evaluate the tradeoff with size and compute, demonstrating the closing gap of FLOPS/kg. Space-based missions often evaluate design trade-offs such as form factor power, thermal management, and compute architecture; this work seeks to find the appropriate compute solution for workloads and then compare those to missions capable of running those loads. Early benchmarks show small-satellite GPUs achieving ~14× speedups over legacy processors. Overall, low-mass platforms in LEO are approaching the AI performance per mass of larger systems, enabling new autonomous capabilities and specialized compute workloads. At the end of this work, sample workloads for missions will be explored and will demonstrate how multiple satellites can autonomously carry out larger missions. This work seeks to demonstrate that AI capabilities including increased autonomy and onboard edge computer vision are possible on the smallest space-based platforms.