Monolithic Silicon-Photonics Linear-Algebra Accelerators Enabling Next-Gen Massive MIMO

Tzu-Chien Hsueh, Yeshaiahu Fainman, Bill Lin

Introduction

The transition toward 6G wireless communication and immersive extended reality (XR) necessitates supporting a massive number of simultaneous users through Massive MIMO (Multiple-Input Multiple-Output) systems. A critical bottleneck in these systems is the constant estimation and inversion of "channel matrices" to decode signals. Matrix inversion possesses a cubic complexity ($O(M^3)$), which poses significant challenges for conventional digital electronics as the number of users ($M$) grows. In millimeter-wave (mmWave) settings, where channel conditions fluctuate within milliseconds, the energy and area overhead of moving data between photonic and electronic chips in heterogeneous solutions becomes prohibitive.

Monolithic Silicon-Photonics Architecture

To address these challenges, researchers at UC San Diego have developed a monolithic silicon-photonics (M-SiPh) System-on-Chip (SoC). Fabricated on a commercial 45nm SOI process (GlobalFoundries 45SPCLO), this design integrates electronics and photonics on a single substrate, eliminating the energy "tax" associated with chip-to-chip communication.

Figure 1: System architecture of the M-SiPh accelerator showing the integration of the laser comb source, DACs, micro-ring modulators (MRMs), and racetrack-resonator photodetectors (RTR-PDs).

The system leverages Wavelength-Division Multiplexing (WDM) to achieve massive spatial and spectral parallelism. A laser comb source injects multiple wavelengths into a single waveguide, where each wavelength carries distinct data. The core computation—matrix-vector multiplication (MVM)—is performed in the optical domain.

Technical Implementation

The M-SiPh accelerator employs several key technical innovations:

  • Neumann-Series Approximation: The system avoids direct matrix inversion by using a hardware-friendly iterative approximation based on repeated matrix-matrix multiplications.
  • Photonic Building Blocks: High-speed digital-to-analog converters (HS-DACs) translate digital data into electrical signals for modulation. Micro-ring modulators (MRMs) act as matrix weights, while racetrack-resonator photodetectors (RTR-PDs) perform the final summation of the dot product by converting aggregate light power back into photocurrent.
  • Circuit-Photonics Co-Design: The architecture includes transimpedance amplifiers (TIAs) and flash ADCs optimized for the photonic signal path.
Figure 2: Detailed circuit diagrams of the HS-DACs and R2R-DACs used for weight modulation and data input.

To handle the inherent non-linearity of silicon photonics, the authors implemented a 5-bit DAC and digital calibration logic. Thermal management is addressed through a hybrid approach using tungsten heaters for coarse tuning and post-fabrication trimming for fine adjustments.

Performance Evaluation

The M-SiPh SoC was evaluated using PDK simulations and compared against Google’s TPUv4 (7nm CMOS). For a vector dimension of $M=256$, the results are significant:

  • Throughput: 131 TMAC/s (Tera-Multiply-Accumulates per second).
  • Computation Density: 2.14 TMAC/s/mm², representing a 12x improvement over the TPUv4 (0.17 TMAC/s/mm²).
  • Energy Efficiency: 27.9 fJ/MAC, which is approximately 40.9x more efficient than the TPUv4's 1141.9 fJ/MAC.

The study demonstrates that the benefits of photonic computing scale more gracefully than digital CMOS as matrix sizes increase.

Limitations and Future Directions

While promising, the current design is limited to 4-bit precision, which is suitable for MIMO decoding but lower than general-purpose AI requirements. Additionally, scaling optical combs to 256 wavelengths remains a challenge. Future work will explore more complex algorithms like Gauss-Seidel or Conjugate-Gradient methods using the same monolithic building blocks to further enhance wireless sensing and communication capabilities.

Summary updated

BibTeX
@article{hsueh2024monolithic,
  author = {Tzu-Chien Hsueh and Yeshaiahu Fainman and Bill Lin},
  title = {{Monolithic Silicon-Photonics Linear-Algebra Accelerators Enabling Next-Gen Massive MIMO}},
  journal = {Journal of Lightwave Technology},
  year = {2024},
  doi = {10.1109/jlt.2024.3395408},
}
Figure