A 354 Mb/s 0.37 mm2 151 mW 32-User 256-QAM Near-MAP Soft-Input Soft-Output Massive MU-MIMO Data Detector in 28nm CMOS

Charles Jeon, Oscar Castañeda, Christoph StuderESSCIRC 2019 · DOI

TL;DR

This paper presents a novel ASIC (Application-Specific Integrated Circuit) that implements a near-optimal data detector for massive MIMO (Multiple-Input Multiple-Output) wireless systems. It achieves unprecedented performance for 32 users transmitting 256-QAM signals, offering significant improvements in error rate and efficiency compared to existing solutions, especially under realistic channel conditions.

What Problem Does This Paper Address?

Massive MIMO technology promises to dramatically increase wireless communication spectral efficiency. However, a major bottleneck is the complexity of the data detection process at the base station, especially as the number of users and antennas grows. Traditional methods for optimal detection (like Maximum A Posteriori - MAP) become computationally intractable. Existing low-complexity algorithms and ASICs often compromise on performance, particularly in real-world scenarios with correlated channels and path loss, or when the number of users equals the number of base station antennas. This paper tackles the challenge of designing a practical, efficient, and high-performance data detector that overcomes these limitations for massive MIMO systems.

Methods

The core of this work is a novel ASIC implementing a modified version of the Large-MIMO Approximate Message Passing (LAMA) algorithm. LAMA is known for its near-optimal error-rate performance in the large-antenna limit. The authors introduce several key innovations:

  • Algorithm-Level Optimizations: They adapt LAMA to handle realistic channel conditions, including correlation and per-user path loss. This involves transforming the algorithm to operate on the Gram matrix (HHH) instead of the channel matrix (H) for reduced complexity, and employing message damping techniques to mitigate performance loss in finite-dimensional systems.
  • Hardware-Friendly Architecture: The ASIC design focuses on efficiency and high throughput. It uses coarse-grained pipeline interleaving to process two detection problems simultaneously, reducing the overall processing time.
  • Low-Complexity Mean and Variance Estimation: For the Mean and Variance (MV) estimation unit, they employ a bit-domain computation approach, leveraging Gray-mapping symmetries and the max-log approximation. This significantly reduces the required arithmetic precision and complexity compared to symbol-domain methods.
  • Efficient Interference Cancellation: The Interference Cancellation (IC) unit utilizes a 32-MAC matrix-vector multiplier and a simplified version of Cannon's algorithm for efficient computation.
  • Soft-Input Soft-Output (SISO) Capability: The detector is designed to provide soft-input soft-output capabilities, allowing for iterative detection and decoding with the channel decoder, further improving overall system performance.

Datasets

The paper evaluates the performance of the proposed ASIC using simulated data based on realistic channel models:

  • i.i.d. Rayleigh fading channel: A standard benchmark for wireless communication channel modeling.
  • WINNER II typical urban micro [11] channel model: A more realistic model that accounts for correlation and path loss, representing a challenging propagation environment.

The authors also compare their results against existing ASICs and algorithms, implicitly using their reported performance metrics as benchmarks.

Key Insights and Ideas

  • Near-Optimal Performance is Achievable in Practice: The paper demonstrates that it's possible to achieve near-MAP error-rate performance in massive MIMO systems with a practical ASIC, even under realistic and challenging channel conditions.
  • Algorithm-Level Optimizations are Crucial for Real-World Performance: Simply applying theoretical algorithms like LAMA to hardware isn't enough. The authors' modifications to handle channel correlation and path loss are critical for bridging the gap between theoretical performance and practical implementation.
  • Bit-Domain Computation for Efficiency: Shifting computations to the bit-domain for mean and variance estimation is a clever way to drastically reduce hardware complexity and power consumption without significant performance degradation.
  • SISO Capability is Key for Iterative Systems: The inclusion of SISO functionality is essential for integrating the detector with channel decoders, enabling iterative processing that further boosts system reliability.
  • Trade-offs in Area vs. Energy Efficiency: While the proposed ASIC achieves excellent performance and a good balance, the paper highlights that achieving the absolute highest energy efficiency might involve operating at lower voltages and frequencies, trading off peak throughput.

Key Results

  • Throughput: The fabricated 28nm CMOS ASIC achieves a throughput of 354 Mb/s for 32 users transmitting 256-QAM.
  • Area: The LAMA detector core occupies a small area of 0.37 mm².
  • Power Consumption: The ASIC consumes 151 mW at its nominal operating point.
  • Performance Improvement: For a 32x32 massive MU-MIMO system under realistic urban micro channel conditions, the ASIC achieves over 11 dB SNR improvement compared to linear MMSE detectors.
  • Area Efficiency: The ASIC offers 4x higher throughput per area compared to a recently proposed message-passing detector.
  • Near-Optimal Error Rates: The detector achieves near-optimal error rates, outperforming existing algorithms like CHEMP, which suffer from error floors.
  • Fixed-Point Performance: Extensive simulations indicate that the implemented fixed-point design achieves near-floating-point performance.

Limitations and Future Work

  • Nominal Energy Efficiency: The paper notes that the nominal energy efficiency is inferior to some other designs due to the increased arithmetic precision required for realistic channel conditions and symmetric massive MU-MIMO systems. However, they show that significant energy savings are possible by operating at lower voltages.
  • Future Work: While not explicitly detailed, the success of this ASIC in achieving near-optimal performance under realistic conditions opens doors for further optimization and deployment in future wireless communication standards. The ability to support up to 32 UEs with relatively small BS antenna arrays is highlighted as a critical advantage for BS providers.

Summary updated

BibTeX
@inproceedings{jeon2019mb,
  author = {Charles Jeon and Oscar Castañeda and Christoph Studer},
  title = {{A 354 Mb/s 0.37 mm2 151 mW 32-User 256-QAM Near-MAP Soft-Input Soft-Output Massive MU-MIMO Data Detector in 28nm CMOS}},
  booktitle = {ESSCIRC},
  year = {2019},
  doi = {10.1109/esscirc.2019.8902889},
}
Figure