Machine Learning at the Edge Using Neural Network Processor

Increasingly, power, space, and cooling constrained embedded computing assets need to run machine learning applications to make autonomous, low latency decisions within operational environments. Where GPUs, CPUs, and FPGAs aren't suitable for these constraints, custom ASICs with their non-standard development processes and frameworks are the best available option for bringing the power of machine learning to the edge. This paper discusses a neural network processing architecture and why it surpasses all size, power, throughput, and weight requirements while maintaining performance and allowing interoperability with commercial vendors license comprehensive development environment and machine learning libraries. We compare this neural network processor architecture with existing AI and machine learning platforms that leverage common architectures such as GPUs, CPUs, and FPGAs. The plug and play nature of the neural network processor architecture is optimized for applications such as EO-IR (Electro-optical/Infrared) Sensor and Radar Systems without respinning the hardware, reducing costs. This paper explores best practices for deploying common AI and machine learning models for object detection, super resolution, and natural language processing on a neural network processor at the edge. The neural network processor offer an alternative for deployment of autonomous machine learning at the edge that brings the power of a robust development ecosystem together with an architecture favorable for power and space constrained use cases.