RapproX: An Adaptive Approximate Adder with Lookback for Efficient Edge AI via Memristive In-Memory Computing

As silicon scaling nears its physical limits and digital systems process ever-growing amounts of data, integrating computation directly within memory is emerging as a key strategy to overcome the constraints of conventional Von Neumann architectures. Approximate In-Memory Computation (IMC) with memristors offers a promising path toward energy-efficient processing for data-intensive applications. Recent adaptive approximate adders exploit operand magnitude to dynamically switch between exact and approximate computation, but typically ignore carry propagation across approximation boundaries, which can significantly degrade application-level robustness. This work introduces RapproX, a family of adaptive memristive approximate adders featuring a lightweight carry lookback mechanism that approximates carry interaction between exact and approximate regions. The proposed approach improves arithmetic robustness while introducing only minimal overhead and enabling resource-efficient implementations through memristor reuse. Experimental results demonstrate that the proposed approaches achieve superior arithmetic quality compared to State-of-the-Art (SoA) memristive approximate adders. More importantly, the carry lookback mechanism translates into substantial application-level benefits. In image processing, RapproX reduces energy consumption by up to 30.9% compared to the most competitive SoA design and by 50.3% compared to exact computation while maintaining roughly 43 dB Peak Signal-to-Noise Ratio (PSNR). Across a range of machine-learning workloads, including k-means, AlexNet on MNIST, and multiple CIFAR-10 models, RapproX preserves near-exact inference accuracy for the evaluated models at low-to-moderate k and maintains the energy advantages of adaptive approximation, while SoA approximations degrade markedly under the same conditions. These simulation-based results suggest that lightweight carry-aware approximation can improve the robustness of adaptive approximate in-memory computing with only marginal hardware overhead.