Performance Characterization of Gemini APU Processing-in-Memory Devices for Space

Autonomous space missions are dependent on their onboard computing capability to carry out their prescribed tasks. Ideal flight hardware systems must be capable of performing mission-critical tasks in real time while satisfying size, weight, power, and cost (SWaP-C) constraints of the mission. Addition-ally, space hardware must also be robust to the unique hazards of space environments. To expand the onboard computing capability for future space missions, new hardware platforms must be evaluated to determine their computational performance, power usage, and radiation sensitivity. One such hardware platform is the Gemini Associative Processing Unit (APU), which features a unique processing-in-memory (PIM) architecture to limit memory transfer operations. With this architecture, computations can be performed directly in memory, eliminating the need for excess data transfers which can negatively affect performance. This study conducts a performance comparison of Gemini APU devices with both modern desktop hardware as well as current generation flight hardware. Based on the results collected in this research, Gemini APU devices can provide much higher performance per watt than modern terrestrial CPUs, and offer uniquely scalable performance at a power profile similar to modern flight hardware.