Traditional computing architectures face severe limitations due to the Von Neumann bottleneck, where data movement between memory, CPU, and RAM incurs orders-of-magnitude higher energy and latency penalties than the actual processing. Furthermore, silicon-based binary switching structures degrade significantly when executing large-scale matrix multiplications required by modern artificial intelligence applications, frequently leading to thermal breakdown. To eliminate these bottlenecks, this paper proposes the Junction-Based Flow Architecture (JBFA), an asynchronous, clockless computing paradigm that unifies memory and logic tracking directly through continuous electromagnetic wave propagation. Within the JBFA framework, scalar values are mapped to wave amplitudes, opcodes/instructions are encoded as specific harmonic frequencies, and mathematical signs are regulated by spatial phases. A single master polarisation equation, gated by Heaviside step functions, governs all four arithmetic operations within one physical junction. Memory retention is structurally managed using Phase-Change Materials (PCM), while second-order non-linear susceptibility (^ (2) ) substrates execute single-transit multiplication. Due to non-linear constraints, division is processed through an analytical multi-state carrier dynamics sequence stabilized by third-order Auger decay, where oversaturation and channel reset are shown to be the same physical process. The fundamental operations of addition, subtraction, and multiplication are validated through Finite-Difference Time-Domain (FDTD) simulations within MEEP, while division is established via analytical verification. We further propose a three-sector chip architecture (Input/Intermediate/Output) in which PCM provides synchronisation without a global clock. Finally, we demonstrate that while standard Von Neumann architectures retain absolute advantages in general-purpose computing, JBFA establishes clear superiority as a high-efficiency, domain-specific hardware accelerator optimized for complex AI workloads.
K R Naga Abishe Kumar (Mon,) studied this question.