Tunable Approximate Multiply Accumulate circuit

Developing smart resource constrained devices ca-pable of internal data processing is crucial due to computing power and energy limitations. Approximate computing design paradigms offer opportunities to enhance computational effi-ciency by introducing negligible variations in results without compromising the Quality of Results (QoR). Redesigning Multi-ply Accumulate (MAC) units to enhance energy efficiency is a key focus, with approximation techniques playing a pivotal role. We propose a Tunable Approximate MAC (AMAC) unit offering computational efficiency, cost-effectiveness, and flexibility. Tun-able AMAC used a 3-bit tune Signal value in the design, which enables the circuit operations to be adjusted dynamically based on specific needs. Results show that our proposed architecture consumes only 36% of the power compared to accurate MAC units and is 68% faster. In image processing applications, utilizing our architecture results in PSNR values ranging from 26 to 32dB and SSIM values ranging from 77 to 86%. Similarly, in machine learning applications, accuracy was maintained with minor variations of 2 to 8%.