Transcending resistive memory noise constraints for high-precision analogue computing

Unlocking general-purpose, high-precision analogue computing is a critical paradigm shift required to overcome the fundamental data-movement bottlenecks of von Neumann architectures. Resistive memory-based Computing-In-Memory (CIM) offers a transformative solution by integrating storage and computation. However, its expansion into error-sensitive domains is fundamentally hindered by intrinsic physical unreliability. Device-level non-idealities, particularly unavoidable write noise, pose a universal challenge to the deployment of advanced Artificial Intelligence (AI) models, especially in the context of scientific computing. This problem becomes especially severe in mainstream generative frameworks, representing a frontier for high-fidelity analogue computing, where computational errors accumulate during successive Neural Differential Equation (NDE) solving steps and lead to catastrophic degradation in generation performance. Here, we introduce a high-precision and energy-efficient CIM architecture that incorporates a training-free analogue-domain write noise compensation scheme to push computing precision toward its fundamental physical limit, which is effectively constrained only by inherent read noise. Furthermore, by leveraging the intrinsic algorithmic resilience of flow matching against read noise, our system successfully decouples high-precision analogue computation from the constraints of device-level imperfections. Validated on 180 nm macros, our system demonstrates that inherently imprecise resistive memory arrays can execute error-sensitive tasks with digital-equivalent performance. When applied to spatiotemporal turbulence generation, our system achieves both exceptional generation quality and remarkable energy efficiency. Specifically, it reduces the statistical error by 99.6% and 97.6% for the generated velocity and pressure fields compared to uncompensated analogue baselines, while delivering efficiency gains of 5.76× for latent flow matching and 5.50× for the decoder relative to a state-of-the-art GPU. This work establishes a versatile pathway for high-precision analogue computing in AI-driven scientific applications.