Source-linked AI summary
In-memory computing on a photonic platform
Carlos Ríos, Nathan Youngblood, Zengguang Cheng, Manuel Le Gallo, Wolfram H. P. Pernice, C David Wright, Abu Sebastian, Harish Bhaskaran
TL;DR
Separating processing and memory limits photonic in-memory computing, motivating architectures that collocate both functions. This paper demonstrates direct scalar multiplication using interacting light pulses in non-volatile photonic memory, with drift-free and linear computation.
Problem
The paper addresses the challenge of breaking the separation between processor and memory in photonic information processing.
Method
The authors use two interacting pulses and integrated phase-change photonic cells to perform direct scalar multiplication in memory.
Results
The demonstrated photonic in-memory computation is drift-free and linear, while direct scalar multiplication is achieved on the integrated cells.
Takeaways & Limitations
The work demonstrates computations with memory collocated in the same physical location using light.
Takeaways & Limitations
Optical attenuation may require a longer GST cell, while saturation can leave the GST in its most amorphous state.
Abstract
from arXiv · showhide
Collocated data processing and storage are the norm in biological systems. Indeed, the von Neumann computing architecture, that physically and temporally separates processing and memory, was born more of pragmatism based on available technology. As our ability to create better hardware improves, new computational paradigms are being explored. Integrated photonic circuits are regarded as an attractive solution for on-chip computing using only light, leveraging the increased speed and bandwidth potential of working in the optical domain, and importantly, removing the need for time and energy sapping electro-optical conversions. Here we show that we can combine the emerging area of integrated optics with collocated data storage and processing to enable all-photonic in-memory computations. By employing non-volatile photonic elements based on the phase-change material, Ge2Sb2Te5, we are able to achieve direct scalar multiplication on single devices. Featuring a novel single-shot Write/Erase and a drift-free process, such elements can multiply two scalar numbers by mapping their values to the energy of an input pulse and to the transmittance of the device, codified in the crystallographic state of the element. The output pulse, carrying the information of the light-matter interaction, is the result of the computation. Our all-optical approach is novel, easy to fabricate and operate, and sets the stage for development of entirely photonic computers.
Introduction
Photonic in-memory computing could combine integrated optical processing with memory to increase computational speed and bandwidth. This work overcomes programming and recovery challenges by demonstrating direct scalar multiplication in a single integrated photonic phase-change memory cell.
- Motivation: Photonic in-memory computing could provide increased speeds and bandwidths by processing directly in the optical domain.Its potential also leverages wavelength division multiplexing and advances in silicon photonics.
- Challenge: Integrated photonic memories face significant challenges in switching energy, speed, and single-shot programming and recovery.The paper states that it overcomes these challenges.
- Contribution: The work demonstrates the first photonic computational memory for direct scalar multiplication of two numbers.The operation multiplies a × b = c for a, b, c ∈ [0,1] using a single integrated photonic phase-change memory cell.
- Novelty: Unlike previous work, the multiplication is not achieved by sequential addition, providing a significant advance in computational efficiency.This distinguishes the demonstrated operation from the cited previous approach.
- Applications: Multiple scalar multiplying units could calculate matrix-vector multiplications in parallel for big data analytics, linear equations, machine learning, and neural networks.The paper identifies these applications as motivations for optical processing in memory and calls the work a milestone.
Results
The results demonstrate direct scalar multiplication through photonic in-memory computing, with phase-change cells providing multilevel, non-volatile transmission states. The approach combines linear, drift-free optical processing with substantially faster and lower-energy erasure than prior demonstrations, while retaining practical limitations from saturation, noise, and measurement electronics.
- Scalar multiplication: Direct scalar multiplication was demonstrated by encoding one number in input-pulse power and the other in device transmittance, with the output pulse representing their product.The output follows P_out = T × P_in, while the transmission baseline must be subtracted as an offset.
- Write/Erase operation: 577 pJ erase pulses enabled single-shot recrystallization, improving energy by more than 100-fold and speed by 25-fold over previous work.The erase pulse was sufficiently energetic and long to induce recrystallization, while the process reduced erase time from microseconds to 125 ns.
- Write/Erase operation: Approximately 200 ns was required for a stable transmission level in a Write/Erase cycle, allowing operation at 2.5 MHz despite unoptimized pulse separation.Further speed and energy improvements could result from decreasing the pulse width and optimizing the time separation.
- Multilevel programming: 13 transmission levels were programmed, with a linear attenuation response that saturates for switching energies larger than 210 pJ.The levels were limited by photodetector noise, and saturation indicates the GST reached its most amorphous state.
- Stability and limitations: No measurable transmission drift was found for up to 10^4 s with the probe continuously on, while probe-off conditions produced nearly 9% drift after reactivation.The programmed multilevel conditioning was preserved over time, although signal-to-noise ratio and noise were limited by the measurement electronics.