Analog and pulse neural networks as the basis for the next generation of ultra-efficient micro-LLMs in industry

The potential of a paradigm shift in the architecture of micro-LLM (small-scale large language models) for industrial applications through the complete abandonment of von Neumann digital computing is being investigated. The possibility of using analog circuits and pulsed neural networks (Spiking Neural Networks, SNN) as a foundation for creating ultra-efficient solutions is analyzed. The main focus is on the ability of these approaches to provide qualitative improvements in energy efficiency and processing speed when performing simple language tasks such as classifying intents and extracting keywords, when operating on resource-limited industrial controllers, including ultra-slow or analog platforms. The fundamental principles of operation of SNN and analog systems, their compliance with the requirements of the industrial environment, the current state of prototyping and key research areas are considered. It is concluded that the potential of these technologies for creating a new class of micro-LLMs is significant, although fraught with technical difficulties.