Decentralized Equalization With Feedforward Architectures for Massive MU-MIMO

Abstract

Linear data-detection algorithms that build on zero forcing (ZF) or linear minimum mean-square error (L-MMSE) equalization achieve near-optimal spectral efficiency in massive multi-user multiple-input multiple-output (MU-MIMO) systems. Such algorithms, however, typically rely on centralized processing at the base-station (BS) which results in (i) excessive interconnect and chip input/output (I/O) data rates and (ii) high computational complexity. Decentralized baseband processing (DBP) partitions the BS antenna array into independent clusters that are associated with separate radio-frequency circuits and computing fabrics in order to overcome the limitations of centralized processing. In this paper, we investigate decentralized equalization with feedforward architectures that minimize the latency bottlenecks of existing DBP solutions. We propose two distinct architectures with different interconnect and I/O bandwidth requirements that fuse the local equalization results of each cluster in a feedforward network. For both architectures, we consider maximum ratio combining, ZF, L-MMSE, and a nonlinear equalization algorithm that relies on approximate message passing. For these algorithms and architectures, we analyze the associated post-equalization signal-to-noise-and-interference-ratio (SINR). We provide reference implementation results on a multi graphics processing unit (GPU) system which demonstrate that decentralized equalization with feedforward architectures enables throughputs in the Gb/s regime and incurs no or only a small performance loss compared to centralized solutions.Index Terms-Data detection, decentralized baseband processing, linear and nonlinear equalization, general-purpose computing on graphics processing units (GPGPU), massive MU-MIMO.

Design Trade-offs for Decentralized Baseband Processing in Massive MU-MIMO Systems

“…Such feedforward architectures avoid the repeated exchange of information among clusters, which mitigates data transfer latency issues [9]. Furthermore, the theoretical analysis in [16] revealed that feedforward equalization architectures with linear algorithms are able to achieve the same or similar spectral efficiency as their centralized counterparts.…”

“…While the literature describes a number of feedforward DBP architectures and algorithms [7], [9], [16], a systematic tradeoff analysis under different system configurations and channel conditions is missing. Such a trade-off analysis, however, is critical to making design decisions for practical massive MU-MIMO systems that rely on DBP.…”

“…The FD architecture performs local equalization and requires data fusion of the local equalization outputsx ul c and noise varianceσ ul c to produce the equalization outputx ul . See [15], [16] for the details.…”

“…We then calculate a matrix inversion and perform equalization locally to obtain the local estimatex c and noise variance σ 2 c . Finally, we fuse the local estimateŝ x c together with the local noise variances σ 2 c at the centralized processor, which yields the global equalization output [15], [16] for the details.…”

“…While the data transfer size and efficiency of the PD and FD architectures depend on system parameters and channel conditions, the PD architecture always outperforms the FD architecture in terms of error-rate performance, as it is able to achieve the same performance as centralized data detection [15], [16]. Figure 3(a) compares the uncoded bit error-rate (BER) of PD-MMSE and FD-MMSE data detection for a singlecarrier system with a cluster size of B c = 32, C = 4 clusters, and U = 16 UEs with 16-QAM.…”