Upcoming physical layer (PHY) processing solutions are expected to support broad transmission bandwidths and the concurrent transmission of multiple information streams, leveraging the massive multiple-input multiple-output (m-MIMO) advances. However, the inherent computational complexities of MIMO PHY algorithms pose significant practical challenges in satisfying the tight real-time processing latency constraints, regardless of the computing platform. Realizing future PHY solutions requires a fundamental rethinking of PHY processing from both architectural and algorithmic perspectives, leveraging and expanding the potential for parallelization within PHY operations.

This work takes a first step in this direction by introducing several key contributions: (a) the introduction of a first highly-scalable, fully software-based MIMO PHY design and implementation framework that can practically realize MIMO designs with large numbers of information streams in a power-efficient manner.

Specifically, on a single commercial-off-the-shelf (COTS) x86 server with up to 12 real-time processing cores, we can support up to 12 concurrently transmitted information streams at 10 MHz bandwidth and 8 streams at 20 MHz bandwidth. As we show the design is scalable to support even massive MIMO configurations (64×12) at 100 MHz.

Unlike existing approaches, it operates within a fully 5G-NR and Open-RAN-compliant framework and presents the first evaluation of the power consumption of a purely software-based PHY. (b) the first software-based realization of massively parallelizable non-linear (MPNL) MIMO processing, running in real-time and over-the-air. Compared to conventional linear approaches, MPNL achieves substantial gains in the system-level power consumption by effectively reducing the number of required antennas and RF-chains by half. (c) the design and implementation of an FPGA-based MPNL detector and the first evaluation of MPNL within a practical Open-RAN-based vehicular network. Quantified gains include improvements of over 300% in the number of concurrently transmitting vehicles and power savings in the order of hundreds of watts from the radio units. Finally, (d) the first application of neuromorphic computing to m-MIMO detection problem, with the potential to revolutionize PHY computational power consumption via an event-driven massively parallel computational framework. Specifically, as we show in this work neuromorphic processing can exceed the error-rate performance of linear MIMO processing with over 45% fewer required operations by avoiding computationally expensive channel inversions or additional overheads related to training. When implemented on neuromorphic hardware, our approach suggests power consumption reductions exceeding an order of magnitude compared to conventional detectors, making it a promising candidate for ultra-energy-efficient PHY implementations.