SPECIALTY GRAND CHALLENGE article

Front. Photonics, 29 January 2024

Sec. Neuromorphic Photonics and Photonic Computing

Volume 4 - 2023 | https://doi.org/10.3389/fphot.2023.1336510

Grand challenges in neuromorphic photonics and photonic computing

  • Dipartimento di Elettronica, Informazione e Bioingegneria, Politecnico di Milano, Milano, Italy

Introduction

For decades, information and communication technologies have exploited the properties of light to transport an ever-increasing amount of data to provide massive connectivity and broadband services. Optical fibers have replaced copper cables in long-haul transmission links, inter- and intra-datacenter communications, and fiber-to-the-home access networks. Although photonics is well-established for high-capacity transmission, its use in computing is still largely unexploited.

This is because Moore’s scaling of digital electronic processors and the use of parallel and distributed computing architectures have satisfied the need for increasingly better computer performance. However, with the advent of the deep learning era, after around 2010, computing power started doubling approximately every 4 months, which is much faster than Moore’s law Zhang and Nauman (2020); Sevilla et al. (2022) (see Figure 1). Supercomputers are rapidly approaching the exascale era, in which requirements regarding latency, bandwidth, and energy consumption are challenging for digital electronics. In addition to this, the explosive growth of artificial intelligence (AI) and machine learning, and their penetration into everyday life, is forcing us to reconsider the traditional way computers work Xu and Jin (2023), and post-Moore paradigms and computer architectures need to be seriously considered Shalf (2020). Centralized architectures are indeed inefficient in implementing models used for artificial neural networks (ANN), which are inherently distributed and require massive parallel interconnections between a multitude of elementary computing units Jain et al. (1996).

FIGURE 1

Furthermore, we should consider that, even though we presume that all information is digital, “real” data, such as images, sounds, and objects, are inherently analog in their nature. If, on the one hand, digital computing does allow very complex processing, on the other hand, as the volume of data increases, digital operations become increasingly less sustainable. Today, for applications requiring massive computing, there is great interest in reconsidering analog technologies, which use dedicated circuits to efficiently process large amounts of data at very high speed Ambrogio et al. (2023).

In this scenario, photonics is emerging as a promising technology to enable the sustainable development of high-performance computing (HPC) Kitayama et al. (2019). Not only can photonics provide fast and high-capacity interconnections among distributed processors Thraskias et al. (2018), it can also perform calculations efficiently inside them Stroev and Berloff (2023). Mapping mathematical operations onto photonic hardware enables the implementation of accelerators for neuromorphic computing Shastri et al. (2021), analog optical processors Giamougiannis et al. (2023); Wu et al. (2022), and quantum integrated processors Harris et al. (2017); Corrielli et al. (2021). Several examples of programmable photonic circuits for advanced optical computing and quantum processing have been recently proposed Bogaerts et al. (2020). Multiplications and accumulations (MACs) and matrix-vector multiplications (MVMs) Zhou et al. (2022), which are the most time- and energy-consuming operations in ANNs, can be efficiently performed directly in the optical domain by exploiting diffractive free-space devices (Figure 2A) Lin et al. (2018); Mengu et al. (2020) and multi-plane light converters (MPLCs), or coherent feed-forward meshes of interferometers (Figure 2B) Shen et al. (2017); Zhang et al. (2021); Pai et al. (2023b) or non-coherent filter banks based on wavelength division multiplexing (Figure 2C) Tait et al. (2018); de Lima et al. (2019). Such accelerators are expected to enable processing speeds of petaMAC operations per second per mm2 and energy efficiencies of attojoule per MAC Nahmias et al. (2020); Nozaki et al. (2019). Recursive programmable photonic circuits with optical feedback structures have been proposed as mathematical co-processors and reconfigurable solvers of several classes of integral/differential equations (Figure 2D) Tzarouchis et al. (2022).

FIGURE 2

To make these proof-of-concept photonic devices evolve toward mature large-scale computing systems, research in “neuromorphic photonics and photonic computing” needs to tackle several open challenges that vertically span from the investigation of novel material properties up to the optimization of computing networks:

  • • Energy sustainability. To dramatically reduce power dissipation in photonic computers that can be programmed and trained with “quasi-zero-energy-consumption”, technology breakthroughs are required to implement fast and energy-efficient optical actuators. Recently, several advances have been demonstrated by exploiting fast electro-optic and plasmonic weighting banks in photonic ANNs Dabos et al. (2022) as well as non-volatile switchable elements based on phase-change materials Figure 3AMiscuglio and Sorger (2020); Feldmann et al. (2019); Farmakidis et al. (2022), optomechanical devices Quack et al. (2023), optical memristors Farmakidis et al. (2022); Youngblood et al. (2023), and liquid crystal devices Yin et al. (2022).

  • • More functionalities. Novel devices are required to enlarge the portfolio of signal processing operations that can be operated directly in the optical domain, both in free-space optics and guided wave devices. Here, promising device concepts are emerging, which are enabled by the realization of tuneable meta-surfaces, subwavelength meta-structures, and meta-devices engineered at the nanoscale Figure 3BCui et al. (2014); Halir et al. (2015); Li et al. (2022).

  • • Scalability. Today, photonic integrated circuits (PICs) can reach very high-integration densities enabled by complementary-metal-oxide-semiconductor (CMOS) manufacturing, but scalability to complex on-chip architectures requires sophisticated modeling and design techniques as well as advanced algorithms, control techniques, and training tools for automated calibration and adaptive reconfiguration Figure 3CHughes et al. (2018); Milanizadeh et al. (2020); Filipovich et al. (2022); Pai et al. (2023b).

  • • Non-linear processing and data storage. Such operations, which are essential in neuromorphic computing and AI, are still inefficient when operated in optics with respect to the performance offered by electronics; in this view, the integration of photonics and electronic circuits may offer a viable route to profit from the main advantages offered by the two technological platforms. However, this approach requires reconsideration of the way integrated circuits are realized, developing suitable procedures for co-designing and co-packaging photonics and electronics chips Peserico et al. (2022).

  • • Electronic-photonic co-integration. The integration of analog photonic elements into analog/digital electronic circuits could herald new classes of computing devices combining the best of both worlds. However, such interplay opens new issues related to the co-design, co-integration, and co-packaging of photonic and electronic circuits. Monolithic integration Figure 3DAtabaki et al. (2018); Zanetto et al. (2023), heterogeneous integration, and chip-to-chip bonding Chang et al. (2023); Nezami et al. (2023) are different approaches that are being explored to leverage the potential of photonic computing.

FIGURE 3

Applications for photonic computing are envisioned in mathematical accelerators for deep learning Kitayama et al. (2019); Zhou et al. (2022), photonic reservoir computing Nakajima et al. (2021), and spiking neural networks Feldmann et al. (2019); cognitive radio Zhu and Pan (2020) and compressive sensing Kilic et al. (2023); the classification of unstructured signals, high-energy particle collision Duarte et al. (2018), and qubit systems Magesan et al. (2015); the non-linear control and optimization in robotics and autonomous vehicles Kuutti et al. (2021); intelligent signal processing in wireless and fibre-optic communication Huang et al. (2022); Milanizadeh et al. (2022); and optical quantum key distribution Kwek et al. (2021) and chaotic systems and encryption Pai et al. (2023a).

Although photonic computing is still in its infancy, in recent years, we have assisted an impressive increase in fundamental and applied research in this field, which is driven by large industrial investment and the birth of small- and medium-scale enterprises. This means that photonic computing is not just a fascinating scientific area but its potential is critically strategic for stakeholders to drive investments in those areas that might have the highest impact, such as AI and quantum information. Photonic computing will probably never replace conventional computers but it can offer real opportunities for a sustainable scalability of performance in distributed and edge-computing systems. Indeed, without photonics, we may have no other alternative technologies to address the bottleneck that electronics are bound to face in meeting the requirements of future computing. Miller (2017).

Statements

Author contributions

FM: Writing–original draft.

Funding

The author(s) declare financial support was received for the research, authorship, and/or publication of this article. FM acknowledges financial support from the Italian National Recovery and Resilience Plan (NRRP) of NextGenerationEU, partnership on ‘Telecommunications of the Future’ (PE00000001—program ‘RESTART’, Structural Project ‘Rigoletto’ and Focused Project ‘HePIC’).

Conflict of interest

The author declares that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

The author(s) declared that they were an editorial board member of Frontiers, at the time of submission. This had no impact on the peer review process and the final decision.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

References

Summary

Keywords

photonic computing, photonic integrated circuits (PICs), photonic neural networks, photonic quantum computing, analog computing, photonics, neuromorphic photonics

Citation

Morichetti F (2024) Grand challenges in neuromorphic photonics and photonic computing. Front. Photonics 4:1336510. doi: 10.3389/fphot.2023.1336510

Received

10 November 2023

Accepted

30 December 2023

Published

29 January 2024

Volume

4 - 2023

Edited and reviewed by

Marco Peccianti, Loughborough University, United Kingdom

Updates

Copyright

*Correspondence: Francesco Morichetti,

Disclaimer

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher.

Outline

Figures

Cite article

Copy to clipboard


Export citation file


Share article

Article metrics