GPU@SAT, the AI enabling ecosystem for on-board satellite applications

Today's space missions perform increasingly complicated tasks on-board satellites, ranging from processing data to running artificial intelligence algorithms to enable new applications. As these algorithms have greater computational complexity, their execution may require dedicated hardware accelerators to provide timely results. This paper present GPU@SAT, a general-purpose GPU-like IP-core fitted on a radiation-hardened FPGA, with the purpose of enabling AI applications onboard satellite. The presented IP-Core is compatible with OpenCL 1.2 programming model, simplifying the development of computational intensive algorithms and their reusability. The design is highly parametrized, in order to adapt to different mission and platform requirements. Finally, a use case in which an AI algorithm will be deployed on GPU@SAT is presented, offering an interesting alternative to custom hardware accelerators.