Model-Native Computing Architecture: Envisioning Future System Architecture Through the Lens of Computer Architecture

The paper introduces the Intelligent Computing Architecture (ICA), which conceptualizes large language models (LLMs) as analogous to traditional computer architectures, with LLMs treated as CPUs and their context windows as memory. It outlines six functional layers and proposes a dual plane architecture that separates probabilistic and deterministic computations, while also presenting design heuristics such as Semantic Locality and Context Budget to guide future developments. This framework aims to leverage established computer architecture principles to address the engineering challenges in next-generation model-native systems, providing a roadmap for researchers and practitioners working on LLMs and AI systems.