PiM-enabled Inference Co-Processor for Long-Range Missions (PICO-LRM)

Onboard AI/ML processing is becoming increasingly prevalent when finding solutions to latency, bandwidth, and power constraints. PICO-LRM is a co-processor for AI/ML workloads designed specifically for deep space and long-term missions. Processing-in-Memory (PiM) computing eliminates the need to move data between separate memory and processing units, which constitutes the majority of power consumption in traditional computing architectures. Utilizing non-volatile memory (NVM) arrays and radiation-hardened FPGA, PICO-LRM is radiation resistant at the hardware level. PICO-LRM adapts NVIDIA's Deep Learning Accelerator (NVDLA) to offer an end-to-end software/firmware/hardware framework that will allow scientists to develop models using familiar tools, such as PyTorch/TensorFlow + CUDA, for optimized deployment on spacecraft. PICO-LRM is designed for seamless integration into NASA's High-Performance Spaceflight Computing (HPSC) ecosystem as an AI/ML coprocessor and can enable NASA's target of 100x improvement in computational capacity within the same power budget.