The memory wall, power wall, and serial bottleneck of the von Neumann architecture have become fundamental physical constraints restricting the development of intelligent computing. Existing compute-in-memory and neuromorphic chips represent local optimizations under fixed architectures. They fail to break through the constraints of preset rules and fixed topology at the foundational level of the computing paradigm, and cannot natively support endogenously evolving intelligent systems. Based on the axiomatic system of Gradient-Relational Ontology, this paper reconstructs the ontological essence of computation and proposes a novel non-von Neumann computing paradigm called gradient computing. The essence of computation is not symbolic manipulation under preset rules, but the conduction, condensation, and bootstrapped evolution of gradient differences. There is no absolute dichotomy between program and data, instruction and operation, architecture and function, which fundamentally eliminates the root cause of the von Neumann bottleneck. This paper further presents a four-layer hardware system architecture for gradient computing, designs three types of native hardware primitives—conduction, coupling, and condensation—and proves that the architecture possesses three core properties: native compute-in-memory, dynamic topological evolution, and architectural bootstrapping. It has the potential to deliver orders-of-magnitude improvements in computing density and energy efficiency over traditional architectures for evolutionary intelligent tasks. Compared with existing brain-inspired computing schemes, gradient computing realizes rule autogenesis and architectural bootstrapping at the hardware level for the first time, provides a native physical substrate for autopoietic agents, and promotes a fundamental leap of the computing paradigm from "executing preset programs" to "endogenously evolving order".
Y Cao (Sat,) studied this question.