As the complexity of large-scale applications like scientific computing and data analytics continues to rise, traditional methods are struggling to keep up with the ever-growing data demands. PhD thesis by Ali Piri tackles this challenge by introducing strategies to improve computational efficiency without compromising performance. The key focus of the research is Approximate Computing (AxC), which exploits the inherent error tolerance in many applications—such as multimedia processing and machine learning—to balance computational load with acceptable error levels.
The thesis starts by offering a comprehensive review of circuit-level functional approximations, particularly in multipliers, and provides key insights into accuracy metrics for evaluating the performance and reliability of these circuits. A notable contribution is the introduction of input-aware approximate computing, a novel approach designed to optimize both performance and energy consumption while maintaining acceptable accuracy for specific workloads. By automating the design of circuits based on input distributions, the research demonstrated significant improvements in accuracy and energy efficiency, particularly when tested with the FIT filter.
The study also delves into the relationship between approximate computing and deep neural networks (DNNs). It explores how DNNs, especially those implemented on custom systolic arrays, respond to faults introduced by approximate multipliers. The findings highlight that while DNNs are generally fault-tolerant, approximations can still significantly impact accuracy, with critical bits like the most significant bit (MSB) being particularly sensitive. The thesis emphasizes the importance of balancing energy efficiency with the resilience of deep neural networks to ensure optimal system performance.
This research makes a significant contribution to the field of efficient computing by introducing innovative techniques for enhancing energy efficiency while maintaining reliable performance, setting the stage for the next generation of scalable, energy-efficient systems.
Read more here: https://theses.hal.science/tel-04965256v1/file/TH_2024ECDL0052.pdf
