Source-linked AI summary
In-memory photonic dot-product engine with electrically programmable weight banks
Wen Zhou, Bowei Dong, Nikolaos Farmakidis, Xuan Li, Nathan Youngblood, Kairan Huang, Yuhan He, C. David Wright, Wolfram H. P. Pernice, Harish Bhaskaran
TL;DR
Existing hybrid photonic–electronic processing lacked a demonstrated computationally successful implementation. This work builds an electrically programmable GST in-memory photonic–electronic dot-product engine and demonstrates image-processing and MNIST convolutional inference with reported accuracies of 86% and 87%.
Problem
Hybrid photonic–electronic processing using electronically reprogrammable phase-change materials had not yet achieved computational success.
Method
The system decouples electronic programming and storage of GST weights from wavelength-multiplexed optical computation in non-resonant SOI waveguide microheaters.
Results
The system achieves record-high 4-bit encoding, 158.5% switching contrast, 1.7 nJ/dB crystallization energy, and MNIST inference accuracies of 86% and 87%.
Takeaways & Limitations
The demonstrated hardware performs parallel image multiplication and convolutional processing with low error and enhanced contrast-to-noise performance.
Abstract
from arXiv · showhide
Electronically reprogrammable photonic circuits based on phase-change chalcogenides present an avenue to resolve the von-Neumann bottleneck; however, implementation of such hybrid photonic-electronic processing has not achieved computational success. Here, we achieve this milestone by demonstrating an in-memory photonic-electronic dot-product engine, one that decouples electronic programming of phase-change materials (PCMs) and photonic computation. Specifically, we develop non-volatile electronically reprogrammable PCM memory cells with a record-high 4-bit weight encoding, the lowest energy consumption per unit modulation depth (1.7 nJ per dB) for Erase operation (crystallization), and a high switching contrast (158.5%) using non-resonant silicon-on-insulator waveguide microheater devices. This enables us to perform parallel multiplications for image processing with a superior contrast-to-noise ratio (greater than 87.36) that leads to an enhanced computing accuracy (standard deviation less than 0.007). An in-memory hybrid computing system is developed in hardware for convolutional processing for recognizing images from the MNIST database with inferencing accuracies of 86% and 87%.
In-memory photonic–electronic computing platform
The platform combines electrically programmable, non-volatile GST weight banks with wavelength-encoded optical inputs to perform in-memory photonic–electronic dot products.
- In-memory photonic–electronic computing platform: GST memory cells serve as electrically programmable, non-volatile weight banks for the photonic dot-product engine.Each cell's material state maps to a weight w_i.
- In-memory photonic–electronic computing platform: The input vector is encoded in probe-light amplitudes across wavelengths, and each GST cell performs scalar multiplication between weight and input.The probe energy remains below the level required to switch the GST phase.
- In-memory photonic–electronic computing platform: A wavelength-division multiplexer combines the optical outputs incoherently to compute the summed dot product Σw_i·x_i.The platform therefore separates weight storage and programming from optical signal processing.
Multilevel GST cells for scalar multiplication
Electrically pulsed GST waveguide cells provide reversible multilevel transmission control, enabling over-4-bit scalar multiplication with low measured error.
- Multilevel GST cells for scalar multiplication: Over 4-bit non-volatile encoding was demonstrated with 16 transmission levels controlled by pulse-amplitude modulation.Write pulses partially amorphize GST, while a 3 V erase pulse recrystallizes it to baseline transmission.
- Multilevel GST cells for scalar multiplication: 158.5% switching contrast was measured in a 3-μm-doped, 2.5-μm GST/SiO2 microheater, while multi-bit programming supports photonic-computing precision.The selected doping length balances switching contrast against optical loss.
- Multilevel GST cells for scalar multiplication: 784 measured multiplications using 16 weights and 49 input amplitudes closely matched exact results, with residual-error standard deviation of 0.0034.The inputs were sequentially modulated by a variable optical attenuator.
- Multilevel GST cells for scalar multiplication: Independent Write/Erase calibration compensates for fabrication variability among microheater devices.The reported multiplication measurement also used digital offset correction for nonzero baseline transmission.
In-memory parallel multiplication operations for image processing
Wavelength-division multiplexing enables four-channel parallel multiplication for image processing, while larger GST switching contrast improves filtering accuracy by reducing error.
- In-memory parallel multiplication operations for image processing: Four wavelength channels process an image in parallel through one electronically programmed GST cell, reducing whole-image processing time by a factor of four.RGB pixels were flattened into a matrix and encoded as optical amplitudes for simultaneous detection.
- In-memory parallel multiplication operations for image processing: The setup applies a shared GST weight to four separated optical channels and demultiplexes the outputs for simultaneous photodetection.The four channels exhibited nearly identical switching contrast after programming.
- In-memory parallel multiplication operations for image processing: Image blurring error SD decreased from 0.071 to 0.008 as switching contrast increased from 4% to 64%.The comparison used advanced image filtering with blurring and edge-detection kernels.
An in-memory photonic–electronic dot-product engine for image recognition
The photonic–electronic dot-product engine performs parallel optical convolution with electronically programmable GST weights and supports CNN recognition on MNIST fashion products and handwritten digits.
- Recognition performance: 87% and 86% inferencing accuracies were achieved experimentally for MNIST handwritten-digit and fashion-product recognition, respectively, comparable with software calculations.The corresponding calculated accuracies were 88% for digits and 87% for fashion products.
- Optical convolution: Four GST cells encode a 2 × 2 kernel, while four wavelength channels process each image patch’s scalar multiplications in parallel.The system computes 169 sequential dot products over 13 × 13 valid-padded patches.
- CNN architecture: The CNN processes 14 × 14 images using four 2 × 2 kernels with ±1 elements to generate four 13 × 13 activation maps.The kernels detect four edge features before subsequent recognition stages.
Discussion
The demonstrated system combines electrically reprogrammable non-volatile GST memory with high-throughput optical computing and achieves accurate image-processing and CNN inference.
- Recognition results: 87% and 86% inferencing accuracies were obtained for handwritten-digit and fashion-product recognition, respectively, comparable with software calculation.The system used an in-memory photonic–electronic dot-product engine for CNN convolutional layers.
- Device performance: 4-bit encoding, 1.7 nJ/dB crystallization energy consumption, and 158.5% switching contrast characterize the electrically programmable GST cells.These properties were demonstrated in non-resonant SOI waveguide microheater devices.
Device fabrication
The hybrid photonic–electronic system was fabricated on a silicon-on-insulator wafer with patterned and etched silicon device layers.
- SOI fabrication: The system was fabricated on an SOI wafer with a 220-nm silicon device layer and a 2-μm buried oxide layer.Electron-beam lithography patterned the top silicon layer before shallow etching.
Measurement setup
Optical and electrical measurements used fibre-chip coupling, continuous-wave and pulsed electrical sources, and probe-based device characterization.
- Optical and electrical instrumentation: Optical measurements coupled continuous-wave laser light into the chip through an apodized waveguide grating coupler.Electrical pulses came from a pulse generator and DC signals from a source meter, combined through a bias tee.
- Programming and computation: A 50-ns, 7-V pulse amorphized GST, whereas a 200-ns, 3-V pulse crystallized it during device operation.Parallel multiplication used a broadband light source filtered before optical processing.
CNN model
The CNN processes compressed MNIST images using a convolutional layer with four predefined 2×2 edge-detection kernels.
- 500 MNIST images were compressed to 14×14 pixels, with 400 used for training and 100 for testing.
- The CNN receives 13×13-pixel valid-padded inputs and applies four predefined 2×2 kernels for edge detection.