📝 SummarXiv

Abstract

Massive MIMO systems have the potential to significantly enhance spectral efficiency, yet their widespread integration is hindered by the high power consumption of the underlying computations. This paper explores the applicability and effectiveness of Algorithm-Based Fault Tolerance (ABFT) for massive MIMO signal processing to tackle the reliability challenge of Near Threshold Computing (NTC). We propose modifying matrix arithmetic Newton iteration MIMO algorithm to seamlessly integrate ABFT to detect any computational errors by inspecting the final result. The overhead from ABFT depends largely on the matrix dimensions, which in this context are dictated by the number of user equipments involved in the computation. NTC is a promising strategy for reducing the energy consumption in digital circuits by operating transistors at extremely reduced voltages. However, NTC is highly susceptible to variations in Process, Voltage, and Temperature (PVT) which can lead to increased error rates in computations. Traditional techniques for enabling NTC, such as dynamic voltage and frequency scaling guided by circuit level timing error detection methods, introduce considerable hardware complexity and are difficult to implement at high clock frequencies. In this context ABFT has emerged as a lightweight error detection method tailored for matrix operations without requiring any modifications on circuit-level and can be implemented purely in software.A MIMO accelerator was implemented on a reconfigurable hardware platform. Experimental results demonstrate that for sufficiently large problem sizes, the proposed method achieves a 36% power saving compared to baseline, with only an average of 3% computational overhead, at default clock frequency. These results indicate that combining ABFT with near-threshold operation provides a viable path toward energy-efficient and robust massive MIMO processors.