Runtime Accuracy Tunable Approximate Floating-Point Multipliers

Due to the widespread popularity of data-driven applications such as big data analytics, machine learning, and computer vision, modern systems need extremely high computational performance to meet users’ requirements. However, the conventional design of arithmetic units is not able to satisfy users with high performance and high energy efficiency simultaneously. Approximate computing is the new paradigm to design arithmetic units for high performance and high energy efficiency. For guarantees on a certain range of error rate from the approximate computing, approximate arithmetic units need an appropriate error estimator. An error estimator based on machine learning (ML) classifiers is a good option since it can foresee the detailed feature of upcoming input data with high accuracy. On the other hand, an error estimator based on ML classifiers requires a training phase, which consumes extra computational power and energy. Besides, for different applications, the error estimator is often required to have another training phase to increase its accuracy. To overcome this shortcoming, we propose Runtime Accuracy Tunable Approximate Floating-point Multipliers in this paper. Our proposed design does not require data profiling, training, and re-training phases. When the error tolerance margin of target applications is 7%, our experimental results indicate that our Runtime Accuracy Tunable Approximate Floating-point Multiplier can save the average delay by 6.4% and energy by 2.9% compared to a reconfigurable approximate floating-point multiplier with an error estimator. Compared to the exact floating-point multiplier, our design can save the average delay by 45.2% and energy by 36.3%.