A DigAn™ Computing Paradigm

Physical AI requires more than just perception; it requires capability for on-device reasoning and generative decision-making. However, deploying Large Language Models (LLMs) and agentic workflows on physical edge devices (e.g., robotics, smart sensors) is currently limited by the thermal and power constraints of digital logic. This talk explores a breakthrough in "Energy Aware Intelligence" through a massively parallel, mixed-signal compute architecture.

We will discuss the implementation of a single monolithic circuit capable of computing entire neural network layers in a single cycle, utilizing a proprietary Instruction Set Architecture (ISA) optimized for AI workloads. By combining the density of analog MAC operations with the precision of digital control, and integrating them directly with 3D memory, we demonstrate a path to running complex generative models at milliwatt scale. This session will cover the architectural trade-offs, the elimination of off-chip memory traffic, and the implications for the next generation of autonomous, physical AI agents.