An IMPLY-based Semi-Serial Approximate In-Memristor Adder

To alleviate the Von Neumann bottleneck, new technologies and computing paradigms have been a hot topic in research and development in recent years. Memristors offer new innovative possibilities from technological and computational points of view. They can store data well and are suitable for in In-Memory Computation (IMC) since they are able to perform logical operations in memory. Another emerging computing paradigm to reduce computing time and area consumption is approximate computing, which is used in error-resistant applications. Here, we propose a novel approximated full adder hat uses the stateful logic Material Implication (IMPLY) in a semi-serial structure. We embedd this full adder in a Ripple Carry Adder (RCA) that we then evaluate on the circuit-level. The error metrics were evaluated and compared to State-of-the-Art (SoA) IMPLY-based adders. At 8-bit our approach requires up to 29% fewer steps and up to 34% less energy compared to the exact algorithm, while the Normalized Median Error Distance (NMED) is less than 0.01 for most scenarios. The proposed adder is applied in image processing and the respective quality metrics are calculated. All of the tested approximation degrees create a satisfactory result since the Peak Signal-to-Noise Ratio (PSNR) is over 30 dB. Thanks to the proposed approach, we save more than 13.5mJ of energy in gray-scale filtering of a 684x912 8-bit image compared to the exact calculations.