Abstract:
Artificial Intelligence is applied in various domains of compute-intensive applications ranging from image recognition healthcare to statistical analysis. Additionally, recent advancements in Deep Neural Network (DNN) running on millions of devices for various AI tasks deliver an accuracy comparable to human levels. However, the accuracy in computations came with an additional cost of increased computational resources and power consumption in traditional computing units. Moreover, this problem becomes more complex while deploying computationally intensive heavy machine learning (ML) models on energy-constrained edge devices. Approximate computing has emerged as a promising paradigm for error-tolerant AI/ML applications deployed on energy-constrained edge devices where the complexity of hardware computing units can be reduced by optimizing circuit logic while slightly trading off computational accuracy. Therefore, we propose novel approximate compressors to design multiply and accumulate (MAC) hardware unit of Deep Neural Network and Convolutional Neural Network (DNN/CNN) that achieve energy-efficient and faster computations with slightly reduced precision. We also propose tunable compressors and MAC unit that support switching between two approximation modes to enable runtime adjustment of energy efficiency and accuracy for energy-autonomous edge devices. We validated and verified the design of the proposed approximate circuit at 7nm and 55nm technology nodes. The simulation result for the tunable compressor shows an average reduction of 49 % in energy consumption, and 30 % in delay compared to the state-of-the-art compressor. In addition, an average reduction of 36 % in energy consumption and 18 % in delay was observed for the MAC unit compared with the conventional MAC.