Abstract: Optical and optoelectronic approaches of performing matrix-vector multiplication (MVM) operations have shown the great promise of accelerating machine learning (ML) algorithms with unprecedented performance. The incorporation of nanomaterials into the system can further improve the performance thanks to their extraordinary properties, but the non-uniformity and variation of nanostructures in the macroscopic scale pose severe limitations for large-scale hardware deployment. Here, we report a new optoelectronic architecture consisting of spatial light modulators and photodetector arrays made from graphene to perform MVM. The ultrahigh carrier mobility of graphene, nearly-zero-power-consumption electro-optic control, and extreme parallelism suggest ultrahigh data throughput and ultralow-power consumption. Moreover, we develop a methodology of performing accurate calculations with imperfect components, laying the foundation for scalable systems. Finally, we perform a few representative ML algorithms, including singular value decomposition, support vector machine, and deep neural networks, to show the versatility and generality of our platform.
Introduction
As machine learning models, particularly Deep Neural Networks (DNNs), continue to grow in complexity, they demand astronomical amounts of computational power. Traditional electronic hardware is increasingly constrained by the "power wall" resulting from the end of Moore’s Law and Dennard scaling. Optical computing presents a compelling alternative, offering massive parallelism and speed-of-light calculations. However, the practical deployment of large-scale optical hardware is often hindered by the inherent non-uniformity and device-to-device variation of nanomaterials like graphene, despite their superior light-matter interaction properties.
Graphene-Metamaterial Device Design
The proposed system utilizes two-dimensional arrays of Spatial Light Modulators (SLMs) and photodetectors. These components are engineered by integrating graphene monolayers with "Extraordinary Optical Transmission" (EOT) metallic metamaterials, specifically concentric ring apertures.
By adjusting the Fermi level of the graphene via electrostatic gating, the transmission of the SLMs and the responsivity of the photodetectors can be precisely tuned. This mechanism allows for nearly zero static power consumption during inference, as the electrostatic control requires minimal current. Furthermore, the gapless nature of graphene enables these devices to operate across a broad spectrum, from visible light to the mid-infrared.
Matrix-Vector Multiplication Architecture
The fundamental operation of the accelerator is Matrix-Vector Multiplication (MVM), achieved through a hybrid optoelectronic process:
- Vector Encoding: The input vector is encoded into the optical transmission levels of the SLM array.
- Matrix Encoding: The weights of the matrix are mapped to the tunable photoresponsivity of the photodetector array.
- Multiplication and Addition: When light passes through the SLM and strikes the detector, the resulting photocurrent represents the product of the light intensity (vector element) and the detector's sensitivity (matrix element). Photocurrents from the same row are summed electronically to complete the MVM operation.
Notably, the system utilizes incoherent light, which simplifies optical alignment and eliminates the complex interference effects associated with coherent laser sources.
Overcoming Hardware Imperfections
To address the ~20% manufacturing variation typical in graphene devices, a "minimum tuning range" strategy was implemented. This methodology normalizes the output of each row based on the smallest available tuning range among its components. This software-defined approach allows the system to map physical currents back to accurate algebraic values, making the architecture highly tolerant of fabrication noise and non-uniformity.
Experimental Results
The platform was benchmarked against several standard machine learning tasks, demonstrating performance competitive with traditional digital processors:
- MNIST Digit Classification: Achieved 88.7% accuracy (vs. 92.3% on an Intel Xeon processor).
- Fashion-MNIST: Achieved 76.8% accuracy (vs. 78.7% on a standard processor).
- Support Vector Machines (SVM): Clustering results on synthetic datasets matched general-purpose processors with a loss difference of less than 0.2%.
- Image Reconstruction: Successfully performed high-resolution reconstruction using Singular Value Decomposition (SVD).
While the system showed slight accuracy degradation in DNNs due to the sensitivity of near-zero weights to finite 8-bit precision, the overall robustness to device variation remained high.
Conclusion and Future Outlook
This graphene-based optoelectronic accelerator demonstrates a path toward scalable, low-power AI hardware. Future developments will focus on full monolithic CMOS-compatible integration and scaling the array size beyond the current 8x8 blocks. Given graphene’s high carrier mobility, future iterations could reach switching speeds in the range of 1 GHz to tens of GHz, significantly outperforming current analog electronic accelerators.
Summary updated