Inference Per Joule: A Performance Metric For Artificial Intelligence In Space Applications

The use of artificial intelligence (AI) has grown exponentially in recent years. This growth is driven in part by the significant advancements in computing capabilities, which have also increased exponentially. Computers have not only become more powerful but also smaller in size, thanks to the evolution of transistor technology. These developments have enabled AI to become a widely accessible tool, even in recreational activities such as image creation and entertainment videos. More recently, the use of AI has extended to space applications, where it can enhance and optimize various tasks. However, space conditions pose significant challenges for conventional computers due to the lack of atmosphere, exposure to high levels of radiation, shock and thermal stresses. This raises an important question: how can we measure the efficiency of a computer running AI in space? For an image processing AI model, a relevant metric is the number of inferences it can process per second. However, in the context of space, limited power consumption becomes critical, as the amount of heat generated - and the subsequent need for dissipation - directly impacts a computer's operation. To address these challenges, this work proposes a metric that combines these two factors to determine the number of inferences a compute module can process per joule of energy consumed. This metric, referred to as efficiency, serves as a key indicator of performance in space environments. To validate this concept, the following work summarizes three years of research dedicated to demonstrating the feasibility and utility of this metric.