Optical computing is back in the headlines. On 9 October 2026 Koç University in Istanbul announced that assistant professor Uğur Teğin and his students had built two systems that “use light itself, rather than conventional electronic circuitry, to perform computations for artificial intelligence“. Tech Xplore carried the release the next morning under the headline “A new approach to sustainable AI: Offloading part of the computational burden onto light”. The research behind it is real, peer reviewed and published in two Nature Portfolio journals, Communications Engineering and Communications Physics.
The press release is short on numbers, so we read the research itself: both papers, their earlier arXiv preprints, the supplementary files and the transparent peer-review record that Nature Portfolio publishes with each article. That record holds the most useful fact in the whole story. Asked by reviewers to show the energy savings, the authors produced a power budget for the fibre-laser system. It draws 32.5 watts and spends about 0.54 joules on each image, which on their own comparison is less efficient than an Nvidia A100 today. The large efficiency gains they describe depend on a faster modulator that the current rig does not have.
None of that makes the work less interesting. Computing with light has a genuine case as AI’s appetite for power grows, which we covered in the growing environmental threat of data centres. But “sustainable” is a claim about energy, and this optical computing research is careful to say what it has and has not measured. This article sets out what was built, what the results show, what the reviewers challenged and what it would take for light to cut an AI power bill.
Table of contents
- What Koç University’s Optical Computing Announcement Says
- Why Sustainable AI Is Turning to Optical Computing
- Paper One: Multichannel Optical Computing for Colour Images
- Self-Optimising Optical Computing Without a Digital Twin
- Paper Two: A Fibre Laser as an Optical Computing Processor
- The Energy Numbers Behind This Optical Computing Claim
- What Peer Reviewers Asked of the Optical Computing Papers
- Limits Before Optical Computing Can Cut AI’s Power Bill
- What Optical Computing Means for Businesses Planning AI Infrastructure
- Optical Computing FAQ
- References and Further Reading
What Koç University's Optical Computing Announcement Says
The release describes two separate optical computing experiments from the same lab, both aimed at one idea: let the physics of light do part of the work that a neural network would otherwise do in software, then hand a simpler problem to a small digital model.
The two papers behind the release
Both papers name Teğin as the corresponding author, and both were funded by the Scientific and Technological Research Council of Türkiye (TÜBİTAK) under grant 122C150. Fatma Nur Kılınç, a master’s student, appears on both. Both were published open access under a Creative Commons non-commercial, no-derivatives licence.
| Item | Paper one | Paper two |
|---|---|---|
| Title | Self-optimizing multichannel optical computing | Multimode fiber laser cavities as nonlinear optical processors |
| Journal | Communications Engineering | Communications Physics |
| Authors | Fatma Nur Kılınç, Uğur Teğin | Dilem Eşlik, Bahadır Utku Kesgin, Fatma Nur Kılınç, Uğur Teğin |
| arXiv preprint | 12 January 2026 | 10 February 2026 |
| Received by journal | 30 June 2026 | 10 February 2026 |
| Accepted | 10 August 2026 | 5 August 2026 |
| Published | 21 August 2026 | 3 September 2026 |
| Light source | 633 nm continuous-wave laser, 5 mW | Ytterbium-doped fibre laser, pumped with 1.2 W at 976 nm |
| Headline result | Skin-lesion accuracy from 68% to 98% (release) | 85% to 99% across four image benchmarks |
The release went out 49 days after the first paper appeared and 36 days after the second. That gap is normal for a university press office, and it explains why the coverage arrived as a single “sustainable AI” story about two papers that were written and reviewed separately.
What “offloading part of the burden” means
Neither system is an all-optical computer. In each, an image is written onto a spatial light modulator (SLM), a liquid-crystal panel that shifts the phase of light pixel by pixel. Light then passes through an optical system that scrambles and recombines it, a camera records the resulting pattern, and a plain linear classifier on an ordinary computer reads the answer off that pattern.
The optical computing step replaces the feature-extraction layers of a deep network, the expensive part, with physics. The digital step that remains is tiny: a Ridge classifier with a few thousand trainable weights. That split, heavy transform in light and light decision in silicon, is what the release means by offloading “part of the computational burden”.
What the release leaves out
The release says the systems “could contribute to the development of faster and more energy-efficient AI systems”, and it is honest that both “still rely on electronic components such as cameras and control units”. What it does not report is any measured energy figure, the speed at which either system processes images, or the fact that the second system was tested only on greyscale images. All three are in the papers or their review files, and they shape how far the sustainability framing reaches.
Why Sustainable AI Is Turning to Optical Computing
The pressure behind optical computing research is electricity. Moving data between memory and processors, and multiplying large matrices, is what modern AI accelerators spend most of their power on, and demand is rising faster than efficiency improves.
The energy problem in numbers
The International Energy Agency’s Energy and AI report puts global data centre electricity use at around 415 terawatt-hours in 2024, about 1.5% of the world’s consumption, and expects it to more than double to around 945 TWh by 2030, with AI the main driver. It adds that a typical AI-focused data centre uses as much electricity as 100,000 households, and that the largest now under construction will use 20 times as much.
Chip designers are attacking the same problem from inside silicon, for example with AI hardware that filters out irrelevant visual data before it is processed, and with stacked chips that split fast and slow layers. Optical computing takes a different route: it changes the medium that carries the calculation.
Where light helps and where it does not
Light is good at the linear algebra that dominates neural networks. A lens performs a Fourier transform as light passes through it; interference adds many signals at once; and propagation itself costs no energy beyond the source. Many beams can share the same space without interfering electrically, which is the parallelism the release refers to.
Light is bad at three things. It does not easily produce the nonlinear functions neural networks need between layers. It is hard to store and buffer. And every optical computing system must convert data into light and back again, through modulators, lasers and cameras that run on electricity at electronic speeds. Most of the engineering in optical computing, including both Koç papers, is about the first and third of those problems.
Earlier milestones in optical computing for AI
The Koç work builds on a decade of optical computing results, several from the same research family. Teğin did his PhD in photonics at EPFL and a postdoc at Caltech, and joined Koç’s electrical and electronics engineering department in 2023, according to the release.
| Year | Work | What it showed |
|---|---|---|
| 2018 | Diffractive deep neural networks (Lin et al., Science) | Printed layers that classify images as light passes through them |
| 2021 | Scalable optical learning operator (Teğin et al., Nature Computational Science) | Nonlinear pulse propagation in multimode fibre used as a learning layer |
| 2024 | Nonlinear processing with linear optics (Yıldırım et al., Nature Photonics) | Repeated linear passes that produce an effective nonlinearity |
| 2025 | Analog optical computer (Microsoft Research, Nature) | MicroLEDs, an SLM and a camera sensor for inference and optimisation |
| 2025 | Optical generative models (UCLA, Nature) | A digital encoder plus an optical decoder that generates images in one step |
| 2026 | The two Koç papers | Colour-preserving optical processing; a laser cavity as the processor |
The common thread in optical computing for AI is a hybrid: optics for the bulk transform, electronics at the edges. Microsoft’s analog optical computer showed the same pattern, and its most quoted figure, roughly 100 times a GPU’s efficiency, was a projection for a scaled-up future machine rather than a measurement of the prototype. That distinction between projected and measured efficiency matters again below.
Paper One: Multichannel Optical Computing for Colour Images
The first paper, by Kılınç and Teğin, tackles a habit in the field: most optical computing experiments convert colour images to greyscale before they start, which throws away two of the three colour channels.
How the multi-plane light conversion rig works
The optical computing setup is compact. A 5 mW red laser at 633 nanometres lights a Holoeye PLUTO-2.1 SLM with 1,920 by 1,080 pixels. The beam bounces four times between the SLM and a flat mirror, so four separate regions of the same panel act as four cascaded phase masks. A FLIR Blackfly S camera records the output, which is cropped and pooled to 2,700 numbers per image and passed to a Ridge classifier.
In the colour version, the red, green and blue channels are written side by side onto the SLM, separated by 20 pixels to limit crosstalk. As they propagate together, interference mixes them, so the camera sees a pattern that encodes each channel and the relationships between them. The phase-only design loses almost no light, and the masks can be reprogrammed for each task.
Why keeping colour matters
Converting to greyscale discards 67% of an RGB image’s raw information, as the preprint puts it. For skin-lesion classification, where small differences in hue carry diagnostic weight, that loss shows up directly in accuracy. On the HAM10000 dermatoscopy dataset (10,015 images, seven diagnostic categories), optical processing of colour images reached 98%, against 92% for the same optical computing pipeline fed greyscale.
The supplementary file sets that 98% against published deep-learning results. The best digital methods it lists reach 95% to 98%, but they rely on data augmentation; one deep ensemble scored 96% with augmentation and 90% without it. The optical computing result used no augmentation.
The results, and how they changed between preprint and journal
The preprint and the published version report different numbers for the harder benchmarks. The published abstract gives gains of 25 to 45 percentage points over raw pixels, a colour advantage of 6 to 11 points and a further 7 to 12 points from self-optimisation. The January preprint said 26 to 58, 5 to 6 and 6 to 7.
| Benchmark | Preprint (January) | Published supplement (August) |
|---|---|---|
| HAM10000, greyscale, optical | 92% | 92% |
| HAM10000, colour, optical | 98% (raw pixels 67%) | 98% (raw pixels 68%, per release) |
| STL-10, greyscale, optical | 77% | 68% |
| STL-10, colour, optical | 83% | 79% |
| STL-10, colour, optimised mixing | 90% | 86% |
| Flowers-17, before optimisation | 74% | 64% |
| Flowers-17, after optimisation | 80% | 76% |
The direction is consistent: the headline skin-lesion result held, while the natural-image scores came down by 4 to 10 points. The published version also notes that on Flowers-17 only half the labelled images were used for training, with the other half reserved for optimising the system. The Communications Engineering review file runs to a single page and says the manuscript “has been previously reviewed at another journal”, so the reports that drove the revisions are not public.
Not only images: a regression test
The same rig also handled tabular data. Nine features from the Abalone dataset, which predicts a shellfish’s age from its measurements, were laid out as a three-by-three grid of phase values. The optical features gave a prediction error of 1.99 years, against 2.26 to 2.29 years for six digital models, from linear regression to a small neural network, using the same features and split. It is a small dataset, but it shows the method is not limited to pictures.
Self-Optimising Optical Computing Without a Digital Twin
The second contribution of paper one is how the optical computing system tunes itself. Most optical computing designs are trained in simulation, then printed or programmed into hardware, and lose accuracy because the real optics never match the model: misalignment, dust, heat drift.
Tuning the input with six numbers
The first method leaves the optics alone and tunes how the colour channels are blended before they reach the SLM. Six mixing coefficients, each between 0 and 1, define linear combinations of red, green and blue. A Bayesian optimiser searched those six numbers using 10% of the data. On STL-10 it added 7 points in both versions of the paper, from 83% to 90% in the preprint and from 79% to 86% in the published supplement.
The appeal for anyone building optical computing hardware is that this costs nothing physical. The optical system and its masks stay fixed; only a software step before the modulator changes.
Tuning the masks with a sandpile
The second method adjusts the phase masks themselves, 22,500 values, while the hardware is running. Rather than computing gradients, which needs an accurate model of the optics, it borrows from self-organised criticality, the physics of avalanches. A simulated Abelian sandpile receives grains at random; when a site exceeds four grains it topples onto its neighbours, and the cascade picks which mask pixels to nudge.
Each candidate change is uploaded to the SLM and measured. If accuracy improves on half the data, the rest is checked; if it still improves, the change is kept. On Flowers-17 the preprint reports a rise from 74% to 80% within about 75 iterations, “within hours” of hardware-in-the-loop training.
What the robustness checks show
The published supplement adds two simulation studies. Across 50 random starting masks, simulated Flowers-17 accuracy averaged 41.51% with a standard deviation of 2.60 points. Across 10 independent sandpile runs, simulated accuracy rose from 0.36 to about 0.42 after 100 iterations. The authors note that the simulation does not capture every experimental imperfection, which is visible in the gap: 41.5% simulated against 64% measured before optimisation. The real optical computing hardware outperforms its own model, the very mismatch the self-tuning approach is meant to absorb.
Paper Two: A Fibre Laser as an Optical Computing Processor
The second paper, published in Communications Physics, goes further. Instead of passive optics, it uses an active laser cavity as the processor, so the laser’s own physics supplies the nonlinear step that passive optical computing struggles to provide.
How the cavity is built
The light source and the processor are the same device. A one-metre ytterbium-doped fibre, pumped with 1.2 W of light at 976 nm, provides gain. Four metres of graded-index multimode fibre carry many spatial modes, two metres of which form a Sagnac loop that acts as a mirror and output. The SLM closes the cavity as a programmable end mirror, and a camera records the output.
An image written onto the SLM as a phase pattern changes which combinations of modes the cavity supports. According to the authors’ response to reviewers, the field settles within about 100 round trips of 50 nanoseconds each, roughly 5 microseconds, into a stable intensity pattern that depends on the input. That pattern is the feature vector.
Where the nonlinearity comes from
The paper attributes the nonlinear behaviour to gain saturation: all the modes draw on the same pool of excited ytterbium ions, so a mode carrying more power suppresses the gain available to its competitors. The authors estimated the competing Kerr effect at about 0.00004 radians per pass, four to five orders of magnitude below the phase changes the SLM imposes, so they ignore it.
Reviewers asked for a control with the pump switched off. The authors explained that this is impossible here: the pumped fibre is the only light source, so switching it off leaves the cavity dark. Instead they disabled gain in their simulation, and accuracy fell to about 32.56%. The supplement does not say which benchmark that figure comes from.
Results across four datasets
The fibre-laser optical computing system was tested on four image classification tasks, all converted to greyscale and resized before encoding. In each case the same Ridge classifier was trained on raw pixels and on the laser’s output patterns.
| Dataset | Task | Raw pixels | Random projection | Laser cavity |
|---|---|---|---|---|
| TrashNet | Recyclable waste, 6 classes | 24.31% | 33.60% | 97.23% |
| RSSCN7 | Aerial scenes, 7 classes | 18.75% | 28.57% | 95.00% |
| OCT MNIST | Retinal scans, 4 classes | 43.12% | Not reported | 98.78% |
| HAM10000 | Skin lesions, 7 classes | 49.77% | Not reported | 85.22% |
The random-projection column was added at a reviewer’s request. It matters because it answers the obvious objection: a fixed random mixing of pixels, done digitally with the same output size, reaches only 28.57% and 33.60%, far short of the cavity. Whatever the laser does, it is more than scrambling. The OCT MNIST figure comes from a balanced 4,000-image subset, 800 of them held out for testing.
Thousands of weights instead of millions
The paper’s headline comparison is trainable parameters. The cavity’s linear readout uses 2,500 to 10,000 weights; the deep networks it is compared with use tens of millions.
| Model | Trainable parameters | RSSCN7 accuracy | HAM10000 accuracy |
|---|---|---|---|
| Laser cavity + Ridge | 2,500 to 10,000 | 94.21% | 85.22% |
| ResNet-50 | 26,000,000 | 93.64% | 85.30% |
| AlexNet | 60,000,000 | 91.85% | 95.29% |
| VGG16 | 138,000,000 | 93.57% | 96.05% |
On the aerial scenes the cavity edges ahead of every network listed, using 2,600 times fewer weights than ResNet-50 at the top of its range. On skin lesions it trails VGG16 by about 11 points. The authors say so plainly in their response to reviewers: “we do not claim accuracy parity there”. One reviewer called the parameter comparison “not fully fair”, because it sets a readout layer against whole networks, and asked for energy and throughput instead. That request produced the most important numbers in this optical computing story.
The Energy Numbers Behind This Optical Computing Claim
The release’s case is energy. Neither paper measured energy in its first version; the fibre-laser paper added a power budget during review, after two of its three reviewers raised it.
The power budget the reviewers asked for
The proof-of-concept optical computing system draws about 32.5 watts. The figure that dominates is not the laser.
The SLM can change its pattern only 60 times a second, so the system classifies at most 60 images a second. Dividing 32.5 W by 60 gives the paper’s own figure of about 0.54 joules per image. Counting the optical transform as equivalent operations, the authors arrive at 12.9 trillion operations per second, or 0.40 TOPS per watt.
Today’s rig against an Nvidia A100
The authors compare that figure with an Nvidia A100, which they list at 1.56 TOPS per watt. On that comparison, the light-based system today is about 3.9 times less efficient than a GPU released in 2020.
Checking the A100 figure against Nvidia’s datasheet
The paper’s table labels the A100 row “INT16” with 312 TOPS and 1.56 TOPS per watt. Those numbers do not fit together. Nvidia’s datasheet gives the 400 W SXM card 312 TFLOPS at FP16 and 624 TOPS at INT8, both without sparsity. 1.56 TOPS per watt matches 624 ÷ 400, the INT8 figure; 312 ÷ 400 is 0.78. Using the FP16 number, the A100’s lead over today’s optical computing rig shrinks from 3.9 times to about 1.95 times. Either way, the GPU is ahead today.
What the 230 TOPS per watt projection assumes
The large efficiency claim is a projection. If a fast modulator replaced the liquid-crystal SLM and ran at 32 kHz, the authors estimate about 30 W of power, 6,900 TOPS of throughput and roughly 230 TOPS per watt, “exceeding the NVIDIA A100 … by more than two orders of magnitude”.
| Measure | Today (60 Hz SLM) | Projected (32 kHz modulator) | Change |
|---|---|---|---|
| Power | 32.5 W | About 30 W | About the same |
| Images per second | 60 | 32,000 | 533 times more |
| Energy per image | 0.54 J | 30 W ÷ 32,000 = 0.94 mJ | About 576 times less |
| TOPS per watt | 0.40 | About 230 | About 575 times more |
| Against A100 at 1.56 | 3.9 times worse | 147 times better | Depends on the modulator |
The projection is plausible physics: the cavity settles in about 5 microseconds, which caps it at around 200,000 images a second, so 32,000 leaves headroom. But no part of it has been built. Two further caveats apply. The TOPS figure counts the optical transform as equivalent digital operations, a convention borrowed from diffractive-network research, not work a GPU would actually have to do. And the 30 W is an assumption: a modulator fast enough to run at 32 kHz would draw its own power, and the paper does not specify one. The sustainability case for this kind of optical computing lives entirely in the modulator that does not yet exist.
What Peer Reviewers Asked of the Optical Computing Papers
Nature Portfolio publishes the reviewer reports and author responses for the fibre-laser paper. They are worth reading, because the reviewers raised exactly the questions a sceptical reader of any optical computing claim would ask, and the paper improved as a result.
One reviewer called it incremental
Reviewer 1 found the experiments “very good” but the contribution “incremental”, arguing that using nonlinear cavity dynamics as a feature map “is well known” and that the work belonged in “a specialized optics journal rather than a high-impact physics journal”. The same reviewer noted that “system advantages such as energy efficiency, latency, and scalability are not presented”. The authors answered with new theory on how the modes compete and with the energy note discussed above.
Reviewer 2 called the demonstration “novel and convincing” and asked detailed questions about the laser: how many modes take part, and whether it lases on one mode or many. The authors’ answer is that “of order ten to a few tens” of modes participate, far fewer than the hundreds the passive fibre could support, because the gain fibre carries only a few.
Missing controls and weak baselines
Reviewer 3 asked for major revisions on nine points. Among them: a stronger baseline than raw pixels, which produced the random-projection column; statistics over more than one data split, which produced a TrashNet result of 95.55% plus or minus 0.94 points across four splits; and a fix to the LDA cluster plots, which had possibly been fitted on test data. The authors refitted them on training data only.
The reviewer also asked for class-balanced scores on HAM10000, whose classes are very uneven. Those show the weakness that headline accuracy hides. Balanced accuracy is 0.88 and macro-F1 is 0.75, and for vascular lesions, the rarest class with 28 test images, precision is 0.24: about three in four images the system labels vascular are something else. The authors attribute that to the unweighted readout.
A promised small-model baseline
Reviewer 3 also asked for “a small digital model with comparable trainable complexity”, the comparison that would show whether the laser beats a compact network rather than a giant one. The authors’ response said a LeNet-class model “will be trained”. The published supplementary file we read contains no such result, so the question of how a 10,000-parameter digital network would fare on these tasks remains open.
Disclosures on the record
The fibre-laser paper declares that Teğin is an editorial board member of Communications Physics, the journal that published it, and that he was not involved in its review or the decision to publish. That is standard and properly disclosed. In the second review round, one reviewer said no point-by-point response had been provided; the authors explained it had been filed as a cover letter and resubmitted it.
Limits Before Optical Computing Can Cut AI's Power Bill
Both optical computing papers list their own limitations, and they line up with what the energy numbers already suggest.
The 60 Hz bottleneck
Both systems use the same model of liquid-crystal SLM, which the fibre-laser paper says refreshes 30 to 60 times a second, and both process one image at a time. At 60 images a second, a system running all day handles 216,000 images an hour. That is fine for a laboratory, but a data centre inference card handles many thousands of images a second. Faster devices, such as digital micromirror arrays, exist, and the authors name them as the next step.
Electronics stay in the loop
The camera, the SLM driver, the pump diode and the host computer are all electronic, and in the fibre-laser system they account for the entire 32.5 W. Optical computing saves energy only if the optical step replaces enough digital work to pay for the conversions into and out of light. Today it does not; with a fast modulator, the authors’ arithmetic says it would by a wide margin.
Drift, calibration and scale
The fibre-laser paper says the system “requires calibration to account for environmental perturbations to the fiber”, and its theory section describes a trade-off: more modes give a richer feature space but make the output more sensitive to disturbance. The colour paper lists SLM fill factor, alignment stability and noise as barriers to larger images and deeper cascades.
Small benchmarks, small images
The datasets are modest: TrashNet has 2,520 images, RSSCN7 2,400, and the OCT result uses a 4,000-image subset. Images were resized to fit the modulator. The fibre-laser paper itself says that harder tasks, including “fine-grained recognition or adversarial robustness”, remain to be evaluated. Nothing here touches language models, which is where most of the AI electricity in the IEA’s forecast will go.
What Optical Computing Means for Businesses Planning AI Infrastructure
For organisations budgeting for AI compute, this optical computing research changes nothing this year. It is laboratory work with no product attached, and the release does not mention commercial plans.
Where optical computing could matter first
The likeliest early uses are narrow: image classification close to a sensor, where the light is already present and the task is fixed. Medical imaging, waste sorting and remote sensing, three of the domains Koç tested, all fit that description. A camera that performs part of the recognition in optics before digitising could cut both power and data volume at the edge, which is a different proposition from replacing a data centre GPU.
Questions to ask any optical AI vendor
Optical computing start-ups and research spin-outs will keep citing projected efficiency, and much optical computing coverage repeats it. The Koç papers are a useful template for the questions to ask, because the authors answered most of them in public.
| Question | Why it matters | Koç fibre-laser answer |
|---|---|---|
| Is the efficiency measured or projected? | Projections assume parts that do not exist yet | 0.40 TOPS/W measured; 230 projected |
| Does the power figure include the electronics? | Modulators and cameras dominate | Yes: 32.5 W, of which the SLM is 24 W |
| What is the energy per inference? | TOPS can count equivalent operations | About 0.54 J per image today |
| What is the throughput? | Speed sets the cost per task | 60 images a second |
| What is the baseline? | Raw pixels make any method look good | Random projection added in review |
| How does it handle rare classes? | Headline accuracy hides them | Vascular-lesion precision 0.24 |
What to watch next
Three developments would move optical computing from interesting to useful. First, a demonstration with a fast modulator that measures energy per image end to end. Second, a comparison against a small digital network of similar size, the test reviewer 3 requested. Third, integration onto a chip, which both papers name as the route to compact hardware. Until those arrive, the sustainable-AI framing describes a destination, not the current state of the hardware.
Optical Computing FAQ
What did Koç University announce?
Two peer-reviewed papers from Uğur Teğin’s lab, published in August and September 2026, in which light performs part of an image-classification task before a small digital classifier finishes it. The university’s release went out on 9 October 2026.
Does this optical computing system use less energy than a GPU?
Not yet. The fibre-laser system draws 32.5 W and uses about 0.54 joules per image, which the authors put at 0.40 TOPS per watt, against 1.56 for an Nvidia A100. Their 230 TOPS per watt figure assumes a 32 kHz modulator that has not been built into the system.
How accurate are the optical computing systems?
The colour system reached 98% on the HAM10000 skin-lesion dataset. The fibre laser reached 97.23% on recyclable waste, 95.00% on aerial scenes, 98.78% on retinal scans and 85.22% on skin lesions, using a linear classifier with 2,500 to 10,000 weights.
Is optical computing ready for data centres?
No. The fibre-laser system runs at up to 60 images a second, both process one image at a time, rely on electronic cameras and modulators, and have been tested only on small image datasets. Nothing in this work addresses language models.
Why do the numbers differ between versions of the first paper?
The January preprint and the August journal version report different scores for the natural-image benchmarks; STL-10 with optimised mixing went from 90% to 86%. The journal’s review file says the paper was reviewed elsewhere first and does not reproduce those reports.
What is a spatial light modulator?
A panel, usually liquid crystal, that changes the phase or brightness of light pixel by pixel under computer control. Both Koç optical computing systems use one to write each image into the light beam, and its refresh rate is the main speed and energy bottleneck.
References and Further Reading
Self-optimizing multichannel optical computing (Communications Engineering)
Multimode fiber laser cavities as nonlinear optical processors (Communications Physics)
Self-optimizing multichannel optical computing, preprint (arXiv)
Multimode fiber laser cavities as nonlinear optical processors, preprint (arXiv)
Transparent peer review file, fibre-laser paper (Communications Physics)
Supplementary material, fibre-laser paper (Communications Physics)
Supplementary information, multichannel paper (Communications Engineering)
Energy and AI: executive summary (International Energy Agency)
NVIDIA A100 Tensor Core GPU datasheet (Nvidia)
Analog optical computer for AI inference and combinatorial optimization (Microsoft Research)
Sustainable generative AI: UCLA develops novel light-based system (UCLA Samueli)
All-optical machine learning using diffractive deep neural networks (Science)
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
