Abstract: Massive Multi-User MIMO (MU-MIMO) is a critical technology for 5G, yet the computational complexity of uplink signal detection remains a significant hurdle. This paper introduces Optimized Coordinate Descent (OCD), a hardware-friendly iterative algorithm that eliminates the need for high-complexity matrix inversions. By incorporating a non-linear box-constraint and a deeply interleaved VLSI pipeline, the proposed architecture achieves a throughput of 376 Mb/s on a Xilinx Virtex-7 FPGA. Experimental results show that the OCD detector outperforms state-of-the-art designs by more than 2.6× in throughput-per-area while maintaining near-optimal error-rate performance.
Introduction
In Massive MU-MIMO systems, a base station (BS) equipped with hundreds of antennas serves multiple users simultaneously. While this increases spectral efficiency, it creates a massive computational bottleneck in the uplink, where the BS must recover individual user signals from a high-dimensional received vector. Optimal detection methods like Sphere Decoding are computationally prohibitive for real-time systems. Linear approximations, such as Minimum Mean Square Error (MMSE) detection, require inverting large matrices with $O(U^3)$ complexity, which scales poorly as the number of users $U$ grows.
Optimized Coordinate Descent
To address these challenges, the authors propose Optimized Coordinate Descent (OCD). Unlike standard Coordinate Descent (CD) which can be computationally heavy per iteration, OCD restructures the mathematical operations to minimize recurrent calculations. By pre-calculating column norms and gains, the complexity per iteration is reduced by a factor of $U/2$ compared to standard CD.
A key feature of the OCD approach is the "BOX-constrained" equalization mode. In this mode, the estimated signal is constrained to the physical boundaries of the constellation (e.g., the 64-QAM square). This non-linear refinement significantly improves the packet error rate (PER), especially in scenarios where the ratio of BS antennas to users is relatively low, providing a performance boost over standard linear MMSE.
Hardware Architecture and Implementation
The sequential nature of CD typically limits hardware throughput, as the result for one user is required to calculate the next. To overcome this, the authors implemented a 24-stage interleaved pipeline that processes 24 different subcarriers simultaneously. This ensures maximum hardware utilization and high clock frequencies. The design utilizes 16-bit fixed-point arithmetic, which the authors demonstrate results in less than 0.2 dB loss compared to ideal floating-point performance.
Performance and Results
The OCD detector was evaluated using simulated WINNER-Phase-2 channel models across various antenna configurations (32, 64, and 128 BS antennas serving 8 users). The algorithm converges to near-optimal MMSE performance in just three iterations, resulting in a low latency of approximately 3.08 $\mu$s.
As shown in the performance curves, OCD consistently approaches the MMSE bound. In terms of hardware efficiency, a 128-antenna, 8-user system implementation achieved 15,597 throughput-per-LUT, representing a 2.6× improvement over the next-best Conjugate Gradient (CG) based detector. Because the design is compact, multiple instances can be placed on a single chip to reach multi-Gb/s speeds.
Conclusion
The Optimized Coordinate Descent algorithm and its corresponding VLSI architecture provide a scalable and efficient solution for massive MIMO data detection. By co-designing the algorithm for hardware efficiency—specifically through residual-form updates and pipeline interleaving—the authors achieved record-breaking throughput-per-area, making it a viable candidate for high-speed 5G base station deployments.
Summary updated