Energy-efficient approximate least squares accelerator: a case study of radio astronomy calibration processing

Approximate computing allows the introduction of inaccuracy in the computation for cost savings, such as energy consumption, chip-area, and latency. Targeting energy efficiency, approximate designs for multipliers, adders, and multiply-accumulate (MAC) have been extensively investigated in the past decade. However, accelerator designs for relatively bigger architectures have been of less attention yet. The Least Squares (LS) algorithm is widely used in digital signal processing applications, e.g., image reconstruction. This work proposes a novel LS accelerator design based on a heterogeneous architecture, where the heterogeneity is introduced using accurate and approximate processing cores. We have considered a case study of radio astronomy calibration processing that employs a complex-input iterative LS algorithm. Our proposed methodology exploits the intrinsic error-resilience of the aforesaid algorithm, where initial iterations are processed on approximate modules while the later ones on accurate modules. Our energy-quality experiments have shown up to 24% of energy savings as compared to an accurate (optimized) counterpart for biased designs and up to 29% energy savings when unbiasing is introduced. The proposed LS accelerator design does not increase the number of iterations and provides sufficient precision to converge to an acceptable solution.