NEXTUMAX: A Physics-Accurate In-Memory Computing Framework for Sub-Nanosecond Edge AI Acceleration on TSMC 40nm HfOx Memristor Crossbars

The relentless expansion of deep neural networks and Generative AI workloads has exposed the critical power and latency bottlenecks of conventional Von Neumann computing architectures, primarily dictated by memory-bus data movement. Compute-in-Memory (CiM) architectures utilizing Resistive Random-Access Memory (ReRAM) offer an unprecedented paradigm shift by executing analog matrix-vector multiplications (I = G * V) in-situ via physical laws (Ohm's Law and Kirchhoff's Current Law). In this paper, we present NEXTUMAX, a comprehensive, physics-accurate ReRAM silicon and edge AI hardware simulation platform. Calibrated directly against experimental TSMC 40nm HfOx metal-insulator-metal (MIM) semiconductor parameters using the Stanford-PKU Compact Physics Model, NEXTUMAX accurately captures atomic oxygen vacancy filament kinetics, continuous conductance switching across an experimentally verified 112,116x resistance window (RLRS = 8.92 kOhm, RHRS = 1.0 GOhm), cycle-to-cycle (C2C) log-normal stochasticity, device-to-device (D2D) spatial wafer variations, quantum Random Telegraph Noise (RTN), and 10-year power-law retention drift. We demonstrate multi-modal in-memory AI accelerators across three benchmark tasks: (1) a 784x10 full-HD 28x28 OCR vision crossbar achieving >98% classification accuracy on physical optical inputs; (2) a 6-bit differential acoustic speech Keyword Spotter (KWS) with native Windows microphone integration and Voice Activity Detection (VAD) achieving 95.9% GPU / 93.9% ReRAM hardware agreement; and (3) a 5-tile In-Memory Transformer Self-Attention Mesh (Attention(Q,K,V) = softmax(QK^T/sqrt(d))V) executing auto-regressive next-token generation at 0.12 nJ/token, achieving a 2,800x energy efficiency improvement over modern datacenter GPUs. Full source code and models are open-sourced under GNU GPLv3 at https://github.com/NEXTUMAX/reram-silicon-simulator.