Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Nat Commun . 2026 Apr 9;17:3396. doi: 10.1038/s41467-026-71599-2 Search in PMC Search in PubMed View in NLM Catalog Add to search Deep neural network inference on an integrated, reconfigurable photonic tensor processor Lennart Meyer Lennart Meyer 1 Kirchhoff-Institute for Physics, University of Heidelberg, Heidelberg, Germany Find articles by Lennart Meyer 1 , Jelle Dijkstra Jelle Dijkstra 1 Kirchhoff-Institute for Physics, University of Heidelberg, Heidelberg, Germany Find articles by Jelle Dijkstra 1 , Simon Tebeck Simon Tebeck 1 Kirchhoff-Institute for Physics, University of Heidelberg, Heidelberg, Germany Find articles by Simon Tebeck 1 , Liam McRae Liam McRae 1 Kirchhoff-Institute for Physics, University of Heidelberg, Heidelberg, Germany Find articles by Liam McRae 1 , Niklas Bahr Niklas Bahr 1 Kirchhoff-Institute for Physics, University of Heidelberg, Heidelberg, Germany 2 Volkswagen AG, Wolfsburg, Germany Find articles by Niklas Bahr 1, 2 , Daniel Steinmeyer Daniel Steinmeyer 2 Volkswagen AG, Wolfsburg, Germany Find articles by Daniel Steinmeyer 2 , Sergey Koptyaev Sergey Koptyaev 3 Enlightra, Renens, Switzerland Find articles by Sergey Koptyaev 3 , Johana Bernasconi Johana Bernasconi 3 Enlightra, Renens, Switzerland Find articles by Johana Bernasconi 3 , Nikolay G Pavlov Nikolay G Pavlov 3 Enlightra, Renens, Switzerland Find articles by Nikolay G Pavlov 3 , Maxim Karpov Maxim Karpov 3 Enlightra, Renens, Switzerland Find articles by Maxim Karpov 3 , John D Jost John D Jost 3 Enlightra, Renens, Switzerland Find articles by John D Jost 3 , Wolfram Pernice Wolfram Pernice 1 Kirchhoff-Institute for Physics, University of Heidelberg, Heidelberg, Germany Find articles by Wolfram Pernice 1 , Frank Brückerhoff-Plückelmann Frank Brückerhoff-Plückelmann 1 Kirchhoff-Institute for Physics, University of Heidelberg, Heidelberg, Germany Find articles by Frank Brückerhoff-Plückelmann 1, ✉ Author information Article notes Copyright and License information 1 Kirchhoff-Institute for Physics, University of Heidelberg, Heidelberg, Germany 2 Volkswagen AG, Wolfsburg, Germany 3 Enlightra, Renens, Switzerland ✉ Corresponding author. Received 2025 Nov 17; Accepted 2026 Mar 24; Collection date 2026. © The Author(s) 2026 Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ . PMC Copyright notice PMCID: PMC13066431 PMID: 41957042 Abstract Artificial neural networks set the pace in machine vision, natural language processing, and scientific discovery, but their performance depends on fast and efficient tensor computations. Analog photonic systems are a promising alternative to digital electronics because they enable ultra-fast, low-latency computing while avoiding capacitive charging losses and electrical crosstalk. Here we present a photonic tensor processor for deep neural network inference, integrated into a standard 19-inch rack unit with a high-speed electronic interface to PyTorch for seamless hardware deployment. The processor implements an all-optical crossbar with nine inputs and three outputs for parallel intensity-based accumulation of weighted signals. Fabricated in imec’s iSiPP50G silicon photonics platform, the chip integrates electro-absorption modulators and photodiodes for scalability and compatibility with high-volume manufacturing. An integrated self-injection-locked microcomb provides stable multi-wavelength carriers. We demonstrate inference on MNIST and CIFAR-10 with 98.1% and 72.0% accuracy, highlighting a compact, reprogrammable platform toward scalable high-speed optical AI accelerators. Subject terms: Laser material processing, Optoelectronic devices and components, Frequency combs Sustainable artificial intelligence needs faster and more efficient hardware. Here, authors demonstrate a fully integrated silicon-photonic tensor processor for deep neural-network inference. They benchmark the processor on the MNIST and CIFAR-10 dataset. Introduction Artificial intelligence is becoming increasingly central to science, industry, and society. As models grow in size and complexity, they place ever greater demands on computational infrastructure 1 , 2 . A major driver of this demand is tensor operations, a fundamental operation underpinning nearly all layers of modern neural networks. At the scale of modern AI models, inference requires hundreds of billions to trillions of multiply-accumulate (MAC) operations, making tensor processing the dominant contributor to latency and energy consumption 3 – 8 . Analog computing is gaining renewed interest for accelerating AI by performing linear algebra operations directly in the physical domain as signals propagate 9 , 10 . In the electronic domain, mature crossbar arrays with memory cells which store weights at each junction exemplify this approach 11 – 13 , with IBM recently showcasing a 64-core in-memory chip capable of deep network inference 14 . Alternatively, Lightmatter’s and Lightelligence’s processors employ hybrid photonic–electronic architectures, leveraging light’s intrinsic parallelism and speed to push computational performance beyond conventional limits 15 , 16 . By combining optical input modulation with electronic accumulation, these hybrid systems achieve near-digital precision across demanding AI workloads and ultra-low latency in optimization tasks. Photonic approaches bring exciting advantages well-known from telecommunication, such as low latency and low propagation loss. Thus, the possibility to fully leverage these properties through purely optical systems has inspired many different approaches to all-optical computing, including phase-change-material-based in-memory computing 17 , 18 , time-wavelength interleaving 19 , multiplexing across multiple degrees of freedom 20 – 23 , diffraction 24 , 25 , coherent MZI meshes 26 , 27 , and utilizing partial or incoherent light sources 28 – 30 . While several demonstrations achieve striking (theoretical) computational powers, many are hard-wired for specific operations, require abstract problem mappings, or face integration and system-level constraints. Here, we present an integrated photonic tensor processor (PTP) for deep neural network inference, where the linear tensor operations are performed all-optically. The PTP adds optical intensities from incoherent inputs and thus requires only an output photodiode array sized to the output dimension. Electro-absorption modulators provide high-speed input and weight modulation and a one-to-one mapping from model weights to drive levels, enabling arbitrary tensor operations without auxiliary transformations except scaling. Packaged as a rack-mounted system with electronic I/O, calibration procedures, and PyTorch integration, the PTP executes pretrained networks on photonic hardware without chip-specific retraining. Results System architecture and operation Our PTP is based on a silicon on insulator (SOI) photonic integrated circuit (PIC). The PIC is fabricated on imec’s iSiPP50G silicon photonics platform and implements an incoherent optical crossbar array with integrated electro-absorption modulators (EAMs) and photodetectors. The SOI chip contains EAMs to encode input vectors and to set matrix weights through changes in transmission, while on-chip photodiodes convert the results of the optical matrix-vector multiplication into electrical current. A schematic of the PTP architecture is shown in Supplementary Fig. 3 . In order to embed the photonic system within a computer-addressable electronic framework, we employ a ZCU216 RFSoC with a high-speed field programmable gate array (FPGA) to drive and read out the PIC. The integrated on-board RF-DACs run at 4 GS/s for EAM modulation and the RF-ADCs at 2 GS/s for readout. An external host connects over Ethernet to a Jupyter server on the RFSoC’s CPU for interactive control, while the CPU generates the vectors and weight sequences. The CPU then hands these values to the FPGA, which conditions and routes them. High-speed streams are directed to the RF-DACs (EAM inputs) and slower settings to multi-channel DACs that program the crossbar. Reprogramming the full weight array takes 62 ms, whereas streaming inputs and digitizing outputs operate continuously at the RF-DAC/RF-ADC sample rates. Transimpedance amplifiers (TIAs) convert the PIC outputs to voltages, and the RF-ADCs digitize the voltages. The FPGA can decimate the samples before the CPU returns results and metadata to the host. TIAs and weight DACs sit off-board and connect directly to the RFSoC. We wire-bond the photonic chip to a custom carrier printed circuit board (PCB, Fig. 1a ), which we insert into a system board using high-speed RF cables. This allows straightforward physical integration and provides electrical connectivity. The entire system is housed in a standard 19-inch rack. Slow control signals, such as bias voltages ranging from −10 to 10 V, are available via dedicated DAC channels. Figure 1b shows a schematic overview of the system. Fig. 1. Photonic tensor processor architecture. Open in a new tab a Photonic Hardware accelerator in a 19-inch rack housing, RFSoC evaluation board, chip on carrier PCB, TIAs, weight DACs, and optical splitters. Close-up of the wire-bonded PIC with glued fiber array, and a top-view image of the PIC. b Schematic overview of the setup and data flow. In order to optically drive the system, we employ a fully packaged and low-noise self-injection-locked (SIL) microcomb based on a high-Q Si 3 N 4 microresonator 31 – 35 as the input light source. We utilize individual comb lines as input carriers from a single comb source with a fixed spectral spacing, avoiding the need to manually tune and stabilize multiple laser sources. The microcomb provides a 485 GHz free-spectral range (FSR) and delivers about ~11 mW total output, of which we tap 5% for monitoring (Fig. 2c ). We demultiplex the comb lines and route the individual carriers to the input grating couplers. EAMs encode vector entries at the inputs onto the different wavelengths using a four-samples-per-symbol scheme with differential DAC drive, and EAMs at each cell store the matrix weights. We compensate for the nonlinear response and wavelength-dependent extinction ratios (Fig. 2d ) during weight calibration, which stabilizes effective weight accuracy across carriers. PIC-integrated SiGe photodiodes convert the accumulated optical outputs. We operate the photodiodes at 3 V reverse bias, chosen after observing diminishing responsivity gains beyond 3 V (Fig. 2e ). Optical packaging influences the practical power budget. Careful four-axis alignment yields about 2.5 dB insertion loss, and adhesive bonding introduces roughly an additional 1.5 dB along with a wavelength shift (Fig. 2f ). We account for these effects in per-channel/per-carrier power allocation and during calibration. Together, carrier power, calibrated EAM nonlinearity, photodiode biasing, and packaging-aware power allocation enable stable, parallel optical multiply-accumulate execution on chip and define the operating envelope for further system scaling. Fig. 2. Characterization of individual components. Open in a new tab a , b Chip-level pictures of the Photonic Tensor Processor and high-Q Si 3 N 4 microresonator used for microcomb generation. c Optical spectrum of the self-injection-locked microcomb with a FSR of 485 GHz. d Wavelength- and bias-voltage-dependent transmission of the EAMs. We operate the modulators around a bias voltage of 2 V in a linear regime. e Photocurrent vs optical input power for different bias voltages with the corresponding responsivity curve. We operate the photodiodes with a bias voltage of 3 V. f We measure transmission of a test structure before and after gluing of the fiber array. Due to the glueing and shrinkage, the peak transmission drops from −2.7 to −3.8 dB. Weight programming We employ a dedicated calibration routine to set the system accuracy by mapping target weights to modulator drive voltages using measured transfer functions. We further apply crosstalk compensation and a global output rescale to correct inter-channel coupling and restore a consistent gain. The Supplementary Note 9 details the complete procedures. We realize signed weights with a balanced readout scheme in which each logical weight is the transmission difference between a main row and a reference row. This symmetric encoding centers the operating point and implements signed multiply–accumulate behavior. The primary performance metric is the statistical output error for matrix vector multiplications, ϵ MVM , defined as the normalized difference between the optically processed results y k ¯ and ideal digital references y k . ϵ MVM = ∣∣ y k − ȳ k ∣∣ 2 ∣∣ y k ∣∣ 2 1 In order to improve overall accuracy, we optionally implement averaged measurements at the cost of decreasing throughput and increasing latency. The MVM error initially decreases with the root of the number of averages, corresponding to a dominantly stochastic error, e.g., from the thermal noise of the transimpedance amplifier. The single-shot, low-latency mode exhibits an error of (19.4 ± 0.5)%, while averaging four times reduces the error to (10.9 ± 0.3)% in a higher-precision mode. Additional averaging yields diminishing returns and approaches a systematic noise floor near 3% as shown in Fig. 3a . This noise floor is mostly attributed to a non-perfect weight programming, also see Supplementary Note 4 . Figure 3b compares measured and ideal outputs. The data points cluster tightly around the diagonal, and the error histogram follows the anticipated one over the square root of the averaging factor scaling, indicating that noise is predominantly stochastic. Figure 3c reports absolute weight error versus target weight. We infer effective analog weights with a least-squares fit from measured inputs and outputs. The mean absolute error is below 5%, with larger deviations at higher magnitudes that reflect the asymmetric nonlinear response of the modulators. Finally, we evaluate the system stability over a measurement period of 120 min. For both the low-latency and precision modes, the MVM error stays within 1.1% of the respective mean value, enabling reliable analog computation over long timescales. Fig. 3. System evaluation of the PTP. Open in a new tab a MVM error vs. averaging. The decreasing error suggests mainly stochastic error sources down to a noise floor of ~3%. b Measured MVM results vs. target MVMs with error distribution (inset). The measured data points lie close to the ideal diagonal. Lower errors for the precision setting of the PTP vs. the fast mode are visible. c Weight error vs. set weight. We estimate the actual weight of the analog system using a least-squares regression from the measured MVM output and input matrix. d MVM error of the two PTP settings over time. Deep neural network inference To benchmark the system, we develop and evaluate two convolutional neural networks on the photonic tensor processor platform. The first model is a compact baseline network with two convolutional layers followed by a fully connected output layer for 10-class classification. We use this model exclusively for MNIST. The second model is a slightly deeper architecture optimized for CIFAR-10, featuring four convolutional layers before the classification layer. Tables 1 and 2 detail both networks. During inference, the photonic hardware performs the convolution and fully connected layers, and the remaining operations run digitally, translating to 97.5% and 99% optical workload, respectively. Table 1. Network architecture for MNIST classification Type/Stride Filter shape Input size Conv/s1 3 × 3 × 1 → 16 28 × 28 × 1 ReLU - 28 × 28 × 16 MaxPool/s2 2 × 2 28 × 28 × 16 Conv/s1 3 × 3 × 16 → 32 14 × 14 × 16 ReLU - 14 × 14 × 32 MaxPool/s2 2 × 2 14 × 14 × 32 FC 7 × 7 × 32 (flattened) → 10 1568 Open in a new tab Table 2. Network architecture for CIFAR-10 classification Type/Stride Filter shape Input size Conv/s1 3 × 3 × 3 → 32 32 × 32 × 3 BN + ReLU - 32 × 32 × 32 Conv/s1 3 × 3 × 32 → 64 32 × 32 × 32 BN + ReLU - 32 × 32 × 64 MaxPool/s2 2 × 2 32 × 32 × 64 Conv/s1 3 × 3 × 64 → 64 16 × 16 × 64 BN + ReLU - 16 × 16 × 64 Conv/s1 3 × 3 × 64 → 128 16 × 16 × 64 BN + ReLU - 16 × 16 × 128 MaxPool/s2 2 × 2 16 × 16 × 128 Dropout p = 0.2 8 × 8 × 128 FC 8 × 8 × 128 (flattened) → 10 8192 Open in a new tab We pretrain both models digitally using PyTorch, employing stochastic gradient descent with momentum for initial training, followed by Adam for hardware-aware fine-tuning. To improve model robustness, we apply cross-entropy loss and data augmentation. After training, we perform fine-tuning over 50 epochs using AIHWKit-Lightning 36 , incorporating hardware-aware weight noise (fixed at 5%) and output noise (10 and 20%) to reflect the system’s measured behavior. After fine-tuning, we deploy both models on the photonic system for inference. As shown in Fig. 4 , the smaller network achieved a classification accuracy of approximately 98% on the full MNIST test dataset in precision mode, and 91% in low-latency mode. For CIFAR-10, we use the more expressive model in precision mode, since CIFAR-10 classification is a much more complex task with less noise tolerance. Photonic inference on a subset of 400 images reaches an accuracy of 72%. Fig. 4. Optical inference classification results. Open in a new tab a Measured test accuracy of both network architectures for different settings. The “mnist-net” on the left and the “cifar10-net” on the right. The accuracy of the digital networks in gray. Low-latency, precision mode of the PTP in orange and blue. b Confusion matrices for both datasets in the precision mode. The larger drop in accuracy during photonic inference for CIFAR-10 is mainly driven by the larger layer sizes and, consequently, the higher dimensionality of the underlying matrix-vector multiplications. Because the PTP has a finite matrix size, we implement these MVMs via tiling, i.e., we decompose an MVM into K partial MVMs and accumulate their outputs digitally. If the partial outputs are uncorrelated and the MVM error is uncorrelated and stochastic, both the signal and the absolute error scale ~ K , such that the relative error remains approximately constant. In contrast, correlated, systematic errors scale as ~ K , causing the relative error to increase with the amount of tiling. The PTP system is dominated by stochastic noise, as shown in Fig. 3a , and thus supports tiling. However, a small systematic error component is present and can become dominant for a large amount of tiling. In the CIFAR-10 network, the fully connected classifier has the largest input dimension. Executing only this layer digitally increases the accuracy by 6%, also see Supplementary Note 3 . Discussion The fully programmable analog PTP executes tensor operations with trained weights and stable photonic I/O. A calibrated mapping from target weights to modulator voltages, including inter-channel crosstalk compensation, enables accurate and repeatable tensor operations. End-to-end inference on MNIST and CIFAR-10 validates practical utility under real-world conditions. With hardware-aware fine-tuning and moderate averaging, MNIST remains near a digital baseline, while CIFAR-10 reflects tighter noise tolerance yet maintains useful task performance. Beyond component- or core-level demonstrations 20 , 22 , 28 , 37 , this work shows a deployable inference system, integrating calibration, electronic I/O, and a PyTorch interface that runs pretrained networks without chip-specific retraining. By explicitly measuring the hardware noise statistics but using a Gaussian error abstraction, training is decoupled from a particular chip instance, supporting portability across devices and enabling scalable deployment. This establishes a functioning optical hardware accelerator that realizes the same operation class as electronic accelerators with programmable linear layers. Because the system is analog, it is subject to noise, distortions, and device variability. Deep networks are robust to reduced precision, often around four bits 38 – 40 , and even lower with mixed precision 40 – 43 , which motivates our tolerance targets. We measure an accuracy-latency trade-off via temporal averaging. In a single-shot and lowest latency pass, the mean MVM error is about 20%. Averaging by four reduces this to about 11%. The error is mostly stochastic and decreases approximately as the inverse square root of the number of averages. Therefore, improving the signal-to-noise ratio, for example, by reducing loss in individual components, will further boost system performance. A key benefit of photonics is that it sidesteps capacitive-charging limits that dominate large electronic crossbars and interconnects. Because signals propagate optically within the tensor processor, we avoid tensor-size-dependent charge/discharge penalties that bound electronic latency 44 – 46 . In our incoherent intensity-accumulating architecture, a single optical wavefront executes the full tensor operation, so optical transit and I/O bandwidth limits set the latency rather than the tensor dimensions themselves. One modulation interval can realize a single-cycle tensor operation, assuming loss and bandwidth remain in budget. Static-weight photonic systems, such as diffractive or fixed-weight designs, can deliver remarkable throughput and efficiency, yet they typically implement a fixed transform with a limited set of operations 21 , 22 , 24 , 47 . Here, we program arbitrary trained weight matrices electrically, bringing photonic MVM close to the flexibility of electronics while preserving the latency advantage. To achieve this flexibility, we separate the time scales of weights and activations. We operate the processor in a weight-stationary inference regime, keeping the weights constant over long sequences of input vectors to minimize digital data transfer, while inputs and outputs are streamed continuously. The reported 62 ms corresponds to a full array update of all weight control voltages and is dominated by control path overhead rather than by intrinsic device limits. In contrast, input activations are modulated at 4 GSPS and outputs are sampled at 2 GSPS, so sustained throughput is set by system-level input–output bandwidth rather than weight updates. We choose these sampling rates as a trade-off because higher ADC sampling rates reduce conversion efficiency 48 – 50 . In the present prototype, energy efficiency is dominated by system-level electronics, in particular DACs, ADCs, and TIAs, rather than by the optical core itself. Using a projected total power consumption of 2.5 W for the 9 × 3 configuration under continuous streaming at an effective 1 GHz symbol rate, the system achieves 27 GMAC/s, corresponding to 0.022 TOPS/W. For comparison, NVIDIA reports an INT8 throughput of 3958 TOPS (with sparsity) at 700 W (H200 SXM) 51 corresponding to ~5.65 TOPS/W. Analog electronic accelerators achieve 1.74–9.34 TOPS/W depending on task and operating point 52 , and analog electro-optic hybrid accelerators report about 0.82 TOPS/W 15 . Importantly, the efficiency is expected to increase with throughput. For an n × n tensor core, the number of MAC operations per symbol scales as n 2 , whereas the dominant converter and amplifier overhead scales with the number of analog channels, ~ n . Scaling core size, parallelism, and bandwidth increases the compute performance of the PTP. Crossbar arrays up to 32 × 32, without an electro-optic interface, have been demonstrated 22 , and multiplexing across additional degrees of freedom, such as wavelength division multiplexing, enables processing several MVMs with a single core in parallel 53 . Increasing electronic interface bandwidth enables proportionally higher streaming rates, leveraging the large bandwidth of photodiodes and electro-absorption modulators 54 . However, efficiency eventually decreases because ADC conversion efficiency degrades at higher sampling rates. As one perspective, a 32 × 32 core with 4 wavelength channels at 1 GHz computes 8.2 TOPS. In addition to scaling the PTP, hardware-software co-design provides a complementary lever on the model side, for example, via grouped and depthwise-separable convolutions that reduce effective matrix sizes and map naturally onto smaller photonic tiles. Latency is a principal strength of our PTP, which performs linear matrix vector multiplications in the optical domain by the propagation of optical fields within the integrated circuit. Unlike systolic arrays, the full matrix vector multiplication is executed within a single symbol period, without deep pipelining, reducing the time to first result. For the sub-millimeter scale photonic circuit used here, the optical propagation time is negligible compared to the symbol duration. Deploying additional degrees of freedom for data encoding can increase parallelism and throughput without increasing the single-cycle latency 20 , 22 , 23 , 28 , 55 . This low-latency regime is most attractive when one large or recurring layer sets the step time, including recurrent architectures such as LSTMs and Hopfield nets. Methods System setup We deploy an Enlightra SIL-microcomb as a light source and split the optical intensity using an Agiltron splitter. An Ando AQ6330 optical spectrum analyzer monitors the low power output. We amplify the high power output using a PriTel LNHPFA-33. An FS FMU-D402160M Multiplexer filters the relevant wavelength channel, C23/C28/C33/C37/C42 of the ITU grid. We amplify C23 and C42 with a PriTel LNHPFA-33-Pre-Amp each and route all optical carriers through Thorlabs FPC562 Polarization controller to the PIC. We glue an SQS fiber array to the PIC with the photoresin Nanoscribe IP-S. Femto HSA-Y-1-60 TIAs interface the chip and convert the photocurrent. Two Analog Devices DC2025A-A boards provide voltage for the weight array of PTP. An RFSoC ZCU216 Evaluation Kit orchestrates the electronic interface. An Arroyo 6305 combo source and two Keithley SourceMeters (2450 and 2400) control the TEC, the laser diode, and two phase heaters of the comb, respectively. Device fabrication We fabricate the PTP on imec’s iSiPP50G silicon photonics platform during a multi-project wafer run ( https://www.imeciclink.com/en/asic-fabrication/silicon-photonics-foundry-services ). The design kit provides the monolithically integrated electro-optic devices we used. Directional Couplers for arbitrary splitting ratios were designed based on the coupling parameters of the design kit. The diced PIC is mounted on a custom printed circuit board. The PCB is fabricated using Eurocircuits defined impedance pool with 4 layers ( https://www.eurocircuits.com/services/defined-impedance-pool/ ). Photonic matrix We encode each input vector element ( x 1 , … , x M ) onto a different wavelength channel using the on-chip modulators. The matrix is implemented as a waveguide crossbar array 22 equipped with directional couplers that evenly distribute the optical power to all EAM cells. For an M × N matrix, the horizontal couplers have splitting ratios of 1 / ( N − j + 1 ) for column index j , and the vertical couplers have splitting ratios of 1 / i for row index i . The matrix elements themselves are encoded in the EAM transmission. A frequency comb with a line spacing larger than the electrical bandwidth ensures interference-free detection of the summed optical intensities along the individual columns. The output power at each column photodiode corresponds to the inner product between the input vector and the respective kernel. Input encoding We are using a zero-mean, four-sample, return-to-zero, alternating encoding, such that the TIAs see only changes around the optical bias. Our desired input symbol is k n ∈ − 1 , 1 We encode it into 4 DAC samples without a DC part, setting the operation/symbol rate of the system to a quarter of the DACs speed: s n = s 4 n + 1 , s 4 n + 2 , s 4 n + 3 , s 4 n + 4 = k n , 0 , − k n , 0 2 We drive the ADC at half the speed of the DAC and sample at the non-zero time slots: k ~ n = C s 4 n + 1 − s 4 n + 3 3 With C being a constant of all involved device parameters (details in the Supplementary Note 7 ). The two nonzero samples produce opposite-signed fluctuations around the bias, so the AC-coupled output carries a signed value even though the optical power itself is non-negative. Normalizing with 1 2 C recovers the encoded symbol 1 2 C * k ~ n = 1 2 C * C s 4 n + 1 − s 4 n + 3 = 1 2 C * C k n − − k n = k n 4 Weight error The weight Error is defined as: ϵ weight = ∣ ∣ w − w ~ ∣ ∣ 2 Δ w 5 with Δ w = max w − min w With w being the desired weight and w ~ the weight set by the PTP. For each target weight value and averaging setting, 1000 random input vectors were processed using the photonic hardware. The photonic output was used to reconstruct the actual implemented weights w ~ via regression. From this, the weight error represents the deviation between the reconstructed and target weights. Repeating this procedure for all weight values and averaging settings yields the datasets of weight errors and reconstructed weights, which is shown in Fig. 3c . MVM Error For the MVM vs averaging measurements in Fig. 3a , 20 independent runs were performed to assess how averaging affects the accuracy of photonic matrix–vector multiplications. In each run, 1000 random input vectors of dimension 10 were processed through a 10 × 10 photonic weight matrix. The averaging parameter was varied over 10 discrete levels. For every run and averaging setting, the photonic MVM outputs were compared with digitally computed reference results to obtain the total MVM error from Eq. 1 . All measured and reference output matrices (1000 × 10 per averaging level) were recorded to enable detailed statistical analysis of accuracy scaling with averaging. To examine the statistical distribution of individual MVM output errors of Fig. 3b , the complete measured and reference datasets from 100 runs were evaluated at two averaging settings (low-latency and precision). Each run processed 1000 random input vectors of dimension 10 through a 10 × 10 photonic weight matrix, producing 1000 output vectors with 10 elements each. Every individual value within these output vectors—that is, each input–output product corresponding to one row–column multiplication result of the MVM—was compared to its digitally computed reference. The measured values were plotted against their targets to visualize overall fidelity, and the deviations were analyzed via histograms and Gaussian fits. Interfacing Optical: We glue a fiber array to the PIC using the Nanoscribe IP-S photoresin. We align the fiber array with a 4-axis optimization stage before applying the glue. After retracking the array, we apply the glue and re-align for optimal transmission before curing the resin. To prevent the resin from flowing into the coupling region, we print a barrier onto the bottom of the fiber array using a Nanoscirbe Quantum X and the photoresin IP-S. Electrical: A carrier PCB was designed to establish an electrical connection to the chip. The chip is glued to the bare copper area in the center and connected to the PCB via bond wires. The input modulation voltages from the RF-DACs enter the board differentially via Samtec ARC6 connectors and are transitioned to single-ended signals using local baluns (Mini-Circuits’ TCM2-43X+). All input modulators share a common connection to the bias voltage, which is decoupled with capacitors located directly underneath. The photodiode output currents are routed to SMA connectors at the left edge to be connected to external TIAs. The TIA inputs provide a 50 R path to ground, while the other end of the photodiodes shares a common, locally decoupled bias voltage like the input EAMs. The TIA outputs are connected to a second set of SMA connectors, transitioned to differential signals using the same baluns as for the inputs, and then routed to ARC6 connectors to be connected to the RF-ADCs. Although not high-speed, the bias and weight voltages use the same type of connector. All traces are impedance-matched to 50 R (single-ended) or 100 R (differential) and groupwise length-matched. Supplementary information Supplementary Information (1.5MB, pdf) Transparent Peer Review file (799.4KB, pdf) Acknowledgements The research is funded by: European Union’s Horizon 2020 research and innovation programme (grant no. 101017237, PHOENICS project, WP, MK) and the European Union’s Innovation Council Pathfinder programme (grant no. 101046878, HYBRAIN project, WP). We acknowledge financial support from Heidelberg University for the publication fee. Author contributions Conceptualization: L.M., J.D., D.S., W.P. and F.P. Methodology: L.M., J.D., S.T., L.M.R., N.B., S.K., J.B., N.G.P. and F.P. Investigation: L.M., J.D., S.T., F.P. and N.B. Visualization: L.M. and J.D. Funding acquisition: D.S., M.K., J.D.J. and W.P. Project administration: W.P., F.P. Supervision: W.P., F.P. Writing—original draft: L.M., F.B., W.P. Writing—review and editing: All authors Peer review Peer review information Nature Communications thanks the anonymous reviewers for their contribution to the peer review of this work. A peer review file is available. Funding Open Access funding enabled and organized by Projekt DEAL. Data availability All data is available in the main text and the supplementary materials. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Supplementary information The online version contains supplementary material available at 10.1038/s41467-026-71599-2. References 1. Sevilla, J. et al. Compute Trends Across Three Eras of Machine Learning. In 2022 International Joint Conference on Neural Networks (IJCNN) 1–8. 10.1109/IJCNN55064.2022.9891914 (IEEE, 2022). 2. Ben-Nun, T. & Hoefler, T. Demystifying parallel and distributed deep learning: an in-depth concurrency analysis. ACM Comput. Surv. 52 , 1–43 (2020). [ Google Scholar ] 3. Mahmoud, M. et al. TensorDash: Exploiting Sparsity to Accelerate Deep Neural Network Training. In 2020 53rd Annual IEEE/ACM International Symposium on Microarchitecture (MICRO) 781–795. 10.1109/MICRO50266.2020.00069 (IEEE, 2020). 4. Sharify, S., Lascorz, A. D., Siu, K., Judd, P. & Moshovos, A. Loom: Exploiting Weight and Activation Precisions to Accelerate Convolutional Neural Networks. In 2018 55th ACM/ESDA/IEEE Design Automation Conference (DAC) 1–6. 10.1109/DAC.2018.8465915 (IEEE, 2018). 5. Cheng, H., Zhang, M. & Shi, J. Q. A survey on deep neural network pruning: taxonomy, comparison, analysis, and recommendations. IEEE Trans. Pattern Anal. Mach. Intell. 46 , 10558–10578 (2024). [ DOI ] [ PubMed ] [ Google Scholar ] 6. Chen, Y., Xie, Y., Song, L., Chen, F. & Tang, T. A survey of accelerator architectures for deep neural networks. Engineering 6 , 264–274 (2020). [ Google Scholar ] 7. Talib, M. A., Majzoub, S., Nasir, Q. & Jamal, D. A systematic literature review on hardware implementation of artificial intelligence algorithms. J. Supercomput. 77 , 1897–1938 (2021). [ Google Scholar ] 8. Misra, J. & Saha, I. Artificial neural networks in hardware: a survey of two decades of progress. Neurocomputing 74 , 239–255 (2010). [ Google Scholar ] 9. Zhou, C., Kadambi, P., Mattina, M. & Whatmough, P. N. Noisy machines: understanding noisy neural networks and enhancing robustness to analog hardware errors using distillation. Preprint at 10.48550/arXiv.2001.04974 (2020). 10. Joshi, V. et al. Accurate deep neural network inference using computational phase-change memory. Nat. Commun. 11 , 2473 (2020). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 11. Ielmini, D. & Wong, H.-S. P. In-memory computing with resistive switching devices. Nat. Electron 1 , 333–343 (2018). [ Google Scholar ] 12. Alibart, F., Zamanidoost, E. & Strukov, D. B. Pattern classification by memristive crossbar circuits using ex situ and in situ training. Nat. Commun. 4 , 2072 (2013). [ DOI ] [ PubMed ] [ Google Scholar ] 13. Boybat, I. et al. Neuromorphic computing with multi-memristive synapses. Nat. Commun. 9 , 2514 (2018). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 14. Gallo, M. L. et al. A 64-core mixed-signal in-memory compute chip based on phase-change memory for deep neural network inference. Nat. Electron 6 , 680–693 (2023). [ Google Scholar ] 15. Ahmed, S. R. et al. Universal photonic artificial intelligence acceleration. Nature 640 , 368–374 (2025). [ DOI ] [ PubMed ] [ Google Scholar ] 16. Hua, S. et al. An integrated large-scale photonic accelerator with ultralow latency. Nature 640 , 361–367 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 17. Ríos, C. et al. In-memory computing on a photonic platform. Sci. Adv . 5 , eaau5759 (2019). [ DOI ] [ PMC free article ] [ PubMed ] 18. Brückerhoff-Plückelmann, F., Feldmann, J., Wright, C. D., Bhaskaran, H. & Pernice, W. H. P. Chalcogenide phase-change devices for neuromorphic photonic computing. J. Appl. Phys. 129 , 151103 (2021). [ Google Scholar ] 19. Xu, X. et al. 11 TOPS photonic convolutional accelerator for optical neural networks. Nature 589 , 44–51 (2021). [ DOI ] [ PubMed ] [ Google Scholar ] 20. Ou, S. et al. Hypermultiplexed integrated photonics–based optical tensor processor. Sci. Adv. 11 , eadu0228 (2025). [ DOI ] [ PMC free article ] [ PubMed ] 21. Dong, B. et al. Higher-dimensional processing using a photonic tensor core with continuous-time data. Nat. Photon. 17 , 1080–1088 (2023). [ Google Scholar ] 22. Feldmann, J. et al. Parallel convolutional processing using an integrated photonic tensor core. Nature 589 , 52–58 (2021). [ DOI ] [ PubMed ] [ Google Scholar ] 23. Bente, I. et al. The potential of multidimensional photonic computing. Nat. Rev. Phys. 7 , 439–450 (2025). [ Google Scholar ] 24. Xu, Z. et al. Large-scale photonic chiplet Taichi empowers 160-TOPS/W artificial general intelligence. Science 384 , 202–209 (2024). [ DOI ] [ PubMed ] [ Google Scholar ] 25. Hu, J. et al. Diffractive optical computing in free space. Nat. Commun. 15 , 1525 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 26. Shen, Y. et al. Deep learning with coherent nanophotonic circuits. Nat. Photon 11 , 441–446 (2017). [ Google Scholar ] 27. Bogaerts, W. et al. Programmable photonic circuits. Nature 586 , 207–216 (2020). [ DOI ] [ PubMed ] [ Google Scholar ] 28. Dong, B. et al. Partial coherence enhances parallelized photonic computing. Nature 632 , 55–62 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 29. Brückerhoff-Plückelmann, F. et al. Probabilistic photonic computing with chaotic light. Nat. Commun. 15 , 10445 (2024). [ DOI ] [ PMC free article ] [ PubMed ] 30. Brückerhoff-Plückelmann, F. et al. Probabilistic photonic computing for AI. Nat. Comput. Sci. 5 , 377–387 (2025). [ DOI ] [ PubMed ] [ Google Scholar ] 31. Marin-Palomo, P. et al. Microresonator-based solitons for massively parallel coherent optical communications. Nature 546 , 274–279 (2017). [ DOI ] [ PubMed ] [ Google Scholar ] 32. Raja, A. S. et al. Electrically pumped photonic integrated soliton microcomb. Nat. Commun. 10 , 680 (2019). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 33. Ulanov, A. E. et al. Synthetic reflection self-injection-locked microcombs. Nat. Photon. 18 , 294–299 (2024). [ Google Scholar ] 34. Voloshin, A. S. et al. Dynamics of soliton self-injection locking in optical microresonators. Nat. Commun. 12 , 235 (2021). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 35. Shen, B. et al. Integrated turnkey soliton microcombs. Nature 582 , 365–369 (2020). [ DOI ] [ PubMed ] [ Google Scholar ] 36. Büchel, J. et al. AIHWKIT-Lightning: A Scalable HW-Aware Training Toolkit for Analog In-Memory Computing. In NeurIPS 2024 Workshop Machine Learning with new Compute Paradigm s (2024). 37. Moralis-Pegios, M., Giamougiannis, G., Tsakyridis, A., Lazovsky, D. & Pleros, N. Perfect linear optics using silicon photonics. Nat. Commun. 15 , 5468 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 38. McKinstry, J. L. et al. Discovering Low-Precision Networks Close to Full-Precision Networks for Efficient Inference. In 2019 Fifth Workshop on Energy Efficient Machine Learning and Cognitive Computing - NeurIPS Edition (EMC2-NIPS) 6–9 (IEEE, Vancouver, BC, Canada, 2019). 10.1109/EMC2-NIPS53020.2019.00009. 39. Jacob, B. et al. Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition 2704–2713 (IEEE, Salt Lake City, UT, 2018). 10.1109/CVPR.2018.00286. 40. Banner, R., Nahshan, Y. & Soudry, D. Post training 4-bit quantization of convolutional networks for rapid-deployment. In Advances in Neural Information Processing Systems (eds Wallach, H. et al.) vol. 32 (Curran Associates, Inc., 2019). 41. Wu, B. et al. Mixed precision quantization of convnets via differentiable neural architecture search. Preprint at 10.48550/arXiv.1812.00090 (2018). 42. Dong, Z., Yao, Z., Gholami, A., Mahoney, M. W. & Keutzer, K. HAWQ: Hessian AWare Quantization of Neural Networks With Mixed-Precision. In Proc. IEEE/CVF International Conference on Computer Vision (ICCV) (2019). 43. Wang, K., Liu, Z., Lin, Y., Lin, J. & Han, S. HAQ: Hardware-Aware Automated Quantization With Mixed Precision. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019). 44. Nahmias, M. A. et al. Photonic multiply-accumulate operations for neural networks. IEEE J. Sel. Top. Quantum Electron 26 , 1–18 (2020). [ Google Scholar ] 45. Miller, D. A. B. Rationale and challenges for optical interconnects to electronic chips. Proc. IEEE 88 , 728–749 (2000). [ Google Scholar ] 46. Gunn, C. CMOS photonics for high-speed interconnects. IEEE Micro. 26 , 58–66 (2006). [ Google Scholar ] 47. Ghazi Sarwat, S. et al. An integrated photonics engine for unsupervised correlation detection. Sci. Adv. 8 , eabn3243 (2022). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 48. B. Murmann. ADC Performance Survey 1997-2025. 49. Luo, L., Chen, S., Zhou, M. & Ye, T. A 0.014mm 2 10-bit 2GS/s time-interleaved SAR ADC with low-complexity background timing skew calibration. In 2017 Symposium on VLSI Circuits C278–C279. 10.23919/VLSIC.2017.8008507 (IEEE, 2017). 50. Li, R. et al. 10GS/s 10 bit Time-interleaved SAR ADC in 28 nm CMOS. In 2023 8th International Conference on Integrated Circuits and Microsystems (ICICM) 16–20. 10.1109/ICICM59499.2023.10365895 (IEEE, 2023). 51. NVidia H200 GPU. NVidia H200 GPU https://www.nvidia.com/de-de/data-center/h200/ (2026). 52. Le Gallo, M. et al. A 64-core mixed-signal in-memory compute chip based on phase-change memory for deep neural network inference. Nat. Electron 6 , 680–693 (2023). [ Google Scholar ] 53. Brückerhoff-Plückelmann, F. et al. Broadband photonic tensor core with integrated ultra-low crosstalk wavelength multiplexers. Nanophotonics 11 , 4063–4072 (2022). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 54. Pantouvaki, M. et al. Active Components for 50 Gb/s NRZ-OOK Optical Interconnects in a Silicon Photonics Platform. J. Lightwave Technol. 35 , 631–638 (2017). 55. Xu, R., Taheriniya, S., Tang, Z. & Pernice, W. Focused Ion Beam Deposition-Assisted Higher-Order Mode Conversion in Asymmetric Directional Couplers. In 2025 25th Anniversary International Conference on Transparent Optical Networks (ICTON) 1–3. 10.1109/ICTON67126.2025.11125451 (IEEE, 2025). Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials Supplementary Information (1.5MB, pdf) Transparent Peer Review file (799.4KB, pdf) Data Availability Statement All data is available in the main text and the supplementary materials. Articles from Nature Communications are provided here courtesy of Nature Publishing Group ACTIONS View on publisher site PDF (2.4 MB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top