Researchers have developed a three-dimensional optical graphics processing unit capable of conducting image classification and planar graphic processing at an operational speed of 25 million frames per second. Conventional electronic graphics processing units face mounting physical challenges as computing power and energy efficiency demands for large artificial intelligence architectures escalate rapidly in the post-Moore era. Free-space optical diffractive neural networks have emerged as potential photonic alternatives that exploit the propagation speed of light to evaluate planar visual signals. However, standard diffractive networks traditionally depend upon coherent laser sources alongside components such as spatial light modulators or digital micromirror devices for input generation. These conventional modulation devices are constrained by operating speeds in the kilohertz range, significant physical dimensions, phase noise susceptibility, and low diffraction efficiency, which restrict overall computational frequency and miniaturization.
To overcome these processing bottlenecks, the research team designed an optical computing architecture named Gezhi that incorporates a vertically integrated addressable vertical-cavity surface-emitting laser array functioning as a high-speed planar input device. The system couples this source array with mutually incoherent diffractive neural networks engineered for spatial weight multiplexing. Unlike standard coherent systems that require external master lasers for injection phase locking and bulky setups such as 4f lens configurations, the newly developed framework eliminates the requirement for auxiliary intermediate optical lenses. By placing the mutually incoherent diffractive neural network chip directly in line with the emitter array and detector plane, the entire free-space optical computing assembly occupies a total input-to-output propagation distance of approximately 7 millimeters, compacting the functional hardware into a handheld form factor.
The optical graphics processing unit operates by fundamentally altering the optical computation mechanism using the dual coherence properties of the laser source. In the array, the physical pitch between adjacent vertical-cavity surface-emitting lasers is set to 100 micrometers, which exceeds the critical coupling threshold necessary for phase locking. Consequently, each emitter maintains individual optical coherence while remaining mutually incoherent with respect to adjacent units across the array. The architecture harnesses this physical distinction during mathematical execution: the individual coherence of each laser channel enables element-wise optical multiplication across diffractive phase layers, whereas mutual incoherence across channels allows optical power distributions arriving at downstream detection regions to undergo direct linear intensity summation without destructive wave interference, effectively executing optical addition.
To realize an operational optical neural network under mutually incoherent illumination, the authors established a specialized forward propagation numerical model. The spatial carrier distribution within the active laser medium determines the optical transverse modes, which vary with electrical injection current. Using a weak-waveguide approximation model, the team identified the presence of guided LP01 and LP11 modes. By regulating the direct injection current to approximately 2 milliamperes at a driving voltage of 1.7 volts, just above the lasing threshold, each emitter is restricted to operate in the fundamental LP01 mode, preventing helical phase structures that introduce computational inaccuracies. The numerical model calculates the individual diffraction field of each emitter independently before linearly superimposing their intensity outputs, allowing diffractive phase patterns across multiple layers to be trained synchronously.
The physical computing module integrates an 8 by 8 oxidation-confined laser array operating at a central optical wavelength of 846.5 nanometers. Addressing is managed by an external field-programmable gate array controller driving individual p-electrodes, enabling arbitrary 64-bit planar input patterns to be displayed at clock rates up to 25 megahertz. The diffractive neural network chip features two monolithically integrated phase-encoded layers etched onto opposite surfaces of a commercial quartz plate. Double-sided photolithography and inductively coupled plasma dry etching using sulfur hexafluoride gas were employed to etch binarized phase profiles with alignment tolerances between 1 and 2 micrometers and etching depth precision under 30 nanometers. Each diffractive layer contains 200 by 200 neurons with individual neuron areas of 25 square micrometers, where light modulations map to target detection zones located 5 millimeters above the chip surface.
High-speed experimental evaluation conducted with high-speed photodetectors and oscilloscopes confirmed rapid inferencing across standard machine learning benchmarks. When processing a continuous test sequence of 1,000 64-bit handwritten digit images at a 25 megahertz clock rate, the optical processor completed evaluation of the entire dataset in 40 microseconds, yielding an effective throughput of 25 million frames per second. Each processed frame exhibited a latency of 2.05 nanoseconds, matching the driving pulse duration of the laser source. Furthermore, the optical system operated under ultralow-light conditions, consuming an average optical energy per frame of 7.42 picojoules for digit zero and 3.52 picojoules for digit one, translating to energy density requirements of 7.42 and 3.52 attojoules per square micrometer per frame, respectively.
The linear intensity superposition enabled by mutually incoherent operation substantially enhanced optical throughput compared to traditional coherent networks. Experimental diffraction efficiencies reached 10.74 percent for handwritten digit zero with a signal-to-noise ratio of 13.42 decibels, and 26.02 percent for digit one with a signal-to-noise ratio of 14.02 decibels, exceeding the analog bit depth requirement of 4 bits. Numerical simulations verified that target detection areas received more than double the optical intensity of comparable coherent diffractive networks. Furthermore, the lack of optical phase crosstalk between independent emitter channels provided robust operational stability, demonstrating invariant classification accuracy when random phase noise distributions were superimposed onto the input visual patterns.
System classification capabilities were verified on downscaled handwritten datasets, achieving an experimental recognition accuracy of 98.6 percent on a binary MNIST classification task. Evaluated across multi-class datasets, the hardware attained 75.5 percent accuracy on a 4-class Extended MNIST subset and 74.6 percent on the complete 10-class MNIST dataset. The authors noted that accuracy on multi-class benchmarks was bounded by the physical 8 by 8 resolution of the current prototype laser array, which necessitated downscaling original 28 by 28 pixel benchmark images and degraded essential structural features. Comparative simulations using identical 8 by 8 input dimensions in conventional coherent networks yielded a similar accuracy ceiling of 77 percent, whereas modeled simulations expanding the mutually incoherent array to 28 by 28 emitters demonstrated accuracy exceeding 90 percent on 10-class MNIST.
Beyond standard categorical classification, the reconfigurable optical platform was implemented as a multi-channel spatial convolutional processing engine for complex graphic manipulation. By configuring 2 by 2 laser subarrays with a one-pixel stride, the processor functioned as parallel edge-extraction kernels that categorized local image segments into binary positive or negative edge decisions, successfully resolving structural contours across multi-pixel visual scenes. Additionally, the architecture demonstrated spatial image filtering by implementing a 3 by 3 Gaussian convolution kernel for noise suppression. In experimental evaluations on imagery degraded by Gaussian white noise, optical Gaussian filtering raised the initial peak signal-to-noise ratio from 6.93 decibels to 11.14 decibels, with subsequent output binarization yielding a final peak signal-to-noise ratio of 12.35 decibels.
Analysis of physical computational metrics demonstrated an operational computing power of 77.3 tera-operations per second and an energy efficiency of 950 tera-operations per watt for the 2-class MNIST evaluation. Under the 4-class Extended MNIST and 10-class MNIST configurations, measured computing capacities were 35.1 and 21.1 tera-operations per second, with corresponding energy efficiencies of 431 and 259 tera-operations per watt. These operational metrics reflect the parallel connectivity between each laser channel and approximately 4,392 diffractive neurons across the effective functional layers. The investigators noted that operational clock rates are currently restricted by external electronics, specifically the 200 megahertz internal clock limit of the field-programmable gate array and propagation limits in intermediate driver chips, rather than intrinsic optical limits of the laser components.
Engineering pathways to scale the processor include fabricating substrate-emitting laser arrays flip-chip bonded to custom complementary metal-oxide-semiconductor driving circuits capable of multi-value grayscale current modulation. Because vertical-cavity surface-emitting lasers and high-speed photodetectors possess modulation capabilities reaching gigahertz frequencies, scaling driving circuits could enable system operating speeds exceeding one billion frames per second. Furthermore, integrating dynamic materials such as liquid crystals or phase-change media into the diffractive layers could introduce physical reconfigurability for adaptive on-chip training and floating-point multi-kernel convolutional processing across optical vision workloads.
Subject of Research: High-throughput optical neuromorphic graphics processing using mutually incoherent diffractive neural networks
Article Title: High-throughput optical neuromorphic graphic processing at millions of images per second
Article References: Dong, Y., Bai, Y., Zhang, Q., Luan, H., & Gu, M. (2025). High-throughput optical neuromorphic graphic processing at millions of images per second. eLight, 5(1), Article 29. https://doi.org/10.1186/s43593-025-00106-9
Image Credits: AI Generated
DOI: 10.1186/s43593-025-00106-9
Keywords: photonic computing, optical neural networks, diffractive neural networks, neuromorphic hardware, VCSEL arrays, optical graphics processing unit, image classification, optical convolution, edge extraction, image denoising, energy efficiency, high-throughput computing, free-space optics, spatial light modulation, optoelectronics, tera-operations per second
Cite Scienmag News
Cassandra Pierce. (November 21, 2026). Optical Graphics Processor Processes 25 Million Images per Second Using Light. Scienmag. https://scienmag.com/optical-graphics-processor-processes-25-million-images-per-second-using-light/
Cassandra Pierce. "Optical Graphics Processor Processes 25 Million Images per Second Using Light." Scienmag, 21 November 2026, https://scienmag.com/optical-graphics-processor-processes-25-million-images-per-second-using-light/. Accessed 21 November 2026.
Cassandra Pierce. "Optical Graphics Processor Processes 25 Million Images per Second Using Light." Scienmag. November 21, 2026. https://scienmag.com/optical-graphics-processor-processes-25-million-images-per-second-using-light/








