A Neuro-Stochastic Embedded Architecture for Energy-Adaptive and Error-Resilient Systems

Traditional approximate computing architectures rely on static truncation and rigid mode switching fundamentally failing to adapt to transient data volatility and often incurring high post-correction overheads. To overcome these critical limitations a Neuro-Stochastic Adaptive Precision Multiplier (NS-APM) is proposed achieving reconfigurability at a continuous fine-grained level. By embedding a hardware-native, gate-level Neuro-Controller the architecture transitions from discrete binary truncation to a continuous probabilistic approximation. This neural agent dynamically controls novel Probabilistic Approximation Cells (PACs) adjusting decision thresholds on a cycle-by-cycle basis to stochastically sculpt the partial product matrix for optimal accuracy. Implemented on a Zynq UltraScale+ MPSoC ZCU104 Evaluation Kit the design demonstrates a highly fluid Power-Performance-Accuracy navigation scaling dynamic power from 61.00 mW down to 14.52 mW across a 4-bit error budget range. Compared to static exact multipliers this methodology achieves up to a 96.9% reduction in energy-per-operation while maintaining a Mean Relative Error Distance as low as 0.0023%. The architecture successfully establishes a new paradigm for dynamic error-generation management by exploiting instantaneous data volatility.