As deep learning models scale to trillions of parameters, traditional silicon-based digital accelerators are hitting a physical wall. Moving data between memory and compute units consumes orders of magnitude more energy than the computation itself, while charging the metallic wires of electronic circuits introduces RC delays and heat dissipation bottlenecks. To break through this "memory wall," researchers are turning to optics.
As Florent Krzakala et al. observe in Artificial intelligence: From electronics to optics, optical systems can process massive parallel calculations at the speed of light, achieving peta-scale performance with a fraction of the power required by conventional GPU clusters. The core mechanism enabling this leap is Wavelength-Division Multiplexing (WDM)—a technique that exploits the fundamental quantum nature of light to perform matrix algebra simultaneously across the optical spectrum.
The Bosonic Advantage and Spectral Parallelism
To understand why optical computing scales so efficiently, we must look at the difference between electrons and photons. Electrons are fermions; they carry a charge and repel one another, meaning two distinct electrical signals cannot occupy the same wire simultaneously without interfering and short-circuiting.
Photons, however, are bosons. As explained by Yuanhao Gong et al. in Photonics for Neuromorphic Computing: Fundamentals, Devices, and Opportunities, photons can occupy the exact same spatial path without any mutual interaction. This property allows a single physical channel—an optical waveguide—to carry multiple independent data streams simultaneously, provided each stream is encoded onto a different wavelength (or color) of light.
By utilizing WDM, a photonic accelerator can achieve massive spatial and spectral parallelism. Instead of sending a vector through a circuit element-by-element over time, an entire vector can be transmitted through a single waveguide at once, with each vector element mapped to a distinct optical frequency. Tzu-Chien Hsueh et al. demonstrate in Monolithic Silicon-Photonics Linear-Algebra Accelerators Enabling Next-Gen Massive MIMO that this approach drastically increases computation density, allowing simultaneous multiplications to occur across the spectrum within a single hardware footprint.
Computing at the Speed of Light: Broadcast-and-Weight
The backbone of neural networks is Matrix-Vector Multiplication (MVM), which relies on parallel Multiply-Accumulate (MAC) operations. WDM architectures implement MVM through a "broadcast-and-weight" mechanism.
Input Vector (X) Weight Mapping (W) Summation (Y)
[x1] -> λ1 (Red) ---\
[x2] -> λ2 (Green) ----+--> [ Microring Resonators ] --> [ Photodetector ]
[x3] -> λ3 (Blue) ---/ (Attenuate by w1,w2,w3) (Natural Power Sum)
- Multiplication via Attenuation: The input vector elements are modulated onto the intensities of different wavelengths ($\lambda_1, \lambda_2, \dots, \lambda_n$) and broadcast into a shared waveguide. This light then passes through an array of Microring Resonators (MRRs). An MRR is a tiny circular waveguide designed to couple with and attenuate specific wavelengths based on its refractive index. As Haoran Zhang et al. describe in Integrated platforms and techniques for photonic neural networks, by tuning these rings, specific neural weights are assigned to distinct spectral channels. The multiplication occurs instantly as the light passes the ring and is attenuated by the weight value.
- Addition via Non-Coherent Summation: In electronics, adding numbers requires complex routing grids and logic gates. In a WDM optical system, addition is practically free. When the modulated wavelengths are routed to a single photodetector, the detector measures the total optical power. As Mehrdad Morsali et al. point out in OISA: Architecting an Optical In-Sensor Accelerator for Efficient Visual Computing, this non-coherent integration sums the data points automatically at the speed of light.
This combination eliminates the routing overhead that plagues digital chips. Shurui Fei et al. show in An Efficient General-Purpose Optical Accelerator for Neural Networks that aggregating partial results onto a single waveguide via MRRs avoids the massive physical area and signal degradation associated with traditional interleaved electronic grids.
Scaling the Spectrum: Microcombs and Multi-FSR Resonance
To compute larger matrices, the system needs more wavelengths. Generating dozens of distinct wavelengths using individual on-chip lasers would be prohibitively bulky and power-hungry. Instead, modern architectures rely on two elegant optical phenomena: soliton microcombs and multi-mode laser cavities.
A soliton microcomb is a single light source that, when passed through a nonlinear cavity, converts a continuous laser beam into a stable "comb" of dozens of evenly spaced optical frequencies. Bo Xu et al. highlight in Recent Progress of Neuromorphic Computing Based on Silicon Photonics: Electronic–Photonic Co-Design, Device, and Architecture that these microcombs provide the precise multi-wavelength channels necessary to scale up high-throughput MVM cores.
Similarly, the multiple longitudinal modes of a single semiconductor laser cavity can be exploited. Rui-Qian Li et al. demonstrate in Scalable wavelength-multiplexing photonic reservoir computing that these longitudinal modes can act as parallel physical neurons. Because each channel shares the same physical medium while processing data independently, the system clock rate and processing speed scale linearly with the number of active channels.
To squeeze even more computational density out of the physical hardware, researchers exploit the periodic nature of microring resonators. An MRR does not just resonate at a single wavelength; it resonates at multiple repeating intervals across the spectrum, known as Free Spectral Ranges (FSRs).
MRR Transmission Spectrum
| Resonance 1 Resonance 2 Resonance 3
| (FSR 1) (FSR 2) (FSR 3)
| || || ||
| || || ||
---+----------||---------------------||---------------------||-----> Wavelength
Weight A Weight B Weight C
(Mapped to λ1) (Mapped to λ2) (Mapped to λ3)
In Reconfigurable versatile integrated photonic computing chip, Kun Liao et al. demonstrate that by utilizing multiple FSRs of a single microring, independent computational weights can be mapped to different parts of a computation within the exact same physical component. This multi-FSR multiplexing multiplies the information density without increasing the physical footprint of the photonic device.
In-Memory Photonic Computing
The ultimate realization of this architecture involves embedding WDM directly into storage mediums to eliminate data movement entirely. Sasindu Wijeratne et al. propose this in Predictive Performance of Photonic SRAM-based In-Memory Computing for Tensor Decomposition, introducing a photonic SRAM (pSRAM) array. By routing multi-wavelength light directly through memory cells, multiplications and dot products are executed as the light traverses the storage medium itself.
Because these operations bypass electronic charge accumulation and rely entirely on the propagation of light, the latency is extraordinarily low. As Fabio Pavanello et al. report in EUROPULS: NEUROmorphic energy-efficient secure accelerators based on Phase change materials aUgmented siLicon photonicS, photonic accelerators leveraging wavelength multiplexing can operate at frequencies exceeding 50 GHz with inference latencies under 100 picoseconds. By shifting from the spatial constraints of copper wires to the spectral freedom of the optical domain, WDM computing provides a scalable, light-speed alternative for next-generation artificial intelligence.
Key Foundational Work
Metasurface-Enabled On-Chip Multiplexed Diffractive Neural Networks in the Visible (2021) — contributed 1 concept elements to this direction.
Artificial intelligence: From electronics to optics (2020) — contributed 1 concept elements to this direction.