NASA is developing High Performance Spaceflight Computing (HPSC) Chiplet that will improve space computing capabilities by two orders of magnitude. A joint NASA-USAF project will develop radiation hardened, high capacity, high-speed memory components that will enable novel space missions and applications. While these efforts will provide an underlying platform, they do not provide the full range of advanced computing capabilities that will be required to support missions currently in the planning stage for the mid-2020s and beyond.
The proposed NPC will enable general purpose neural networks and other machine learning accelerators for applications such as system health management, robotic vision, and similar applications which are needed to meet the combination of performance and power requirements in future autonomous robotic systems. Some of the advantages of NPC are: High Data Rates, High Fan-In, Low Power Consumption, Low Latency [us-ns], Radiation Robustness, and EMI Robustness. The benefits of NPCs enable advancements in NASA applications and HPSC use cases which require hard real time decisions of high bandwidth signals from many analog instruments. We specifically aim to develop a HPSC-compatible NPC capable of support landers, small sat, and generally high bandwidth instruments.
The novelty of this proposal is to develop an NPC capable of system health management, specifically Long Short-Term Memory networks (LSTM) with predictive analysis algorithms. An NPC would be able to process 100s of high bandwidth (>20 GHz) input signals encoded on unique wavelengths of light to determine safety levels of critical systems and compute hard real time decisions. The need for this functionality has increased as the complexity of instruments and signals responsible for maintaining a NASA vehicle continues to scale, the combination of artificial intelligence and photonic hardware offers a cutting-edge solution for maintaining critical systems.
The proposed NPC would be designed specifically for invariant analysis in system health management. As the control and monitoring of space vehicles grows in complexity, so does maintaining the combined health of vehicle and determining when issues are imminent. The primary applications we are focusing, as defined High-Performance Spaceflight Computing (HPSC) Middleware Overview, Lander, Smallsat, and high bandwidth instrument applications.
The NPC being proposed for this proposal would be able to support other applications which require low power consumption and high throughput. The current design can be implemented for maintaining nuclear power stations. More sophisticated deep learning algorithms could be implemented for autonomous vehicle applications where performance to power ratios are critical.