Light is exceptionally good at carrying information across a chip, but it has yet to prove equally effective at holding that information while a processor works. In 2025, researchers at the USC Information Sciences Institute and the University of Wisconsin–Madison demonstrated a regenerative photonic latch, a tiny memory cell that stores a bit as light on a commercial silicon-photonics platform.
The device is not a finished optical computer, and it is unlikely to appear inside a consumer gadget soon. Instead, it addresses a major obstacle to photonic computing: processors can calculate with light, but they still rely on electronic memory, forcing data to move repeatedly between electrical and optical forms. Photonic computing memory could keep more of that information in light while it is read, processed, and written.
What the 2025 photonic memory breakthrough achieved
The team reported the first experimental demonstration of a cross-coupled differential regenerative photonic latch, or pLatch, fabricated on GlobalFoundries’ 300 mm Fotonix silicon-photonics platform. The research was presented at the 2025 IEEE International Electron Devices Meeting.
The cell demonstrated three essential jobs:
- Write: it accepted an optical signal representing a data bit.
- Hold: it retained that state reliably for several minutes.
- Read and compute: it supported optical input and output while another operation was taking place.
Researchers also demonstrated one-bit in-memory computing and a multibit first-in, first-out structure. Simulations based on the foundry process projected operation at up to 20 GHz. That figure was simulated performance, not the measured speed of a commercial chip operating continuously at 20 GHz.
Why memory is the missing piece in photonic computing
Most digital computers use a version of the von Neumann architecture: a processor performs calculations, while separate memory holds the instructions and data. The separation makes flexible computing possible, but it also forces a constant exchange of information between two physically distant components.
A peer-reviewed review of computational photonic memories identifies this data-movement problem as a central limitation. A modern system must still move data across electrical wires, then convert it into light, perform optical work, convert the result back into electricity, and store it electronically again. Every conversion adds components, energy use, and delay.
Light is not inherently unsuitable for computation. It can carry very high bandwidth through waveguides and perform many mathematical operations as signals propagate. The harder question is how to keep a value available for long enough to use it, let the system rewrite it many times, and pack enough cells into a useful processor.
This also explains why photonics has reached ordinary data center networks sooner than general-purpose computing. Moving information is already a mature job for light. Storing, addressing, and repeatedly rewriting it inside the same chip is a much more demanding task. Intelligent Living has previously examined how photonic chips are entering data center networks before they replace electronic processors.
How a photonic latch stores and regenerates data
A standard static random-access memory cell, or SRAM, uses cross-coupled transistor circuits to hold a bit. The new pLatch follows a similar logical idea but cross-couples optical elements instead.
Two light signals represent opposite states. Each signal influences the other, helping the cell preserve the chosen state rather than fading because of noise or signal loss. That feedback is the “regenerative” part of its name. A conventional passive optical loop stores a pulse, but the pulse weakens. A regenerative latch continually reinforces its output, making a more stable state.
The researchers used differential photonic signalling: information is represented by the relationship between two signals, not by a single pulse. This can make it a bit easier to distinguish from noise, offsets, and device variations. It also makes the device a useful building block rather than a complete memory chip, just as one transistor is only one component in an electronic processor.
Why the commercial foundry process matters
Many impressive optical experiments rely on exotic materials, carefully aligned laboratory equipment, or manufacturing steps that volume factories cannot easily reproduce. That can leave a technology scientifically impressive but far from commercial readiness.
Fabricating the pLatch on GlobalFoundries’ Fotonix process changes the starting point. The 300 mm platform is a commercial monolithic silicon-photonics manufacturing environment, allowing the research team to use the same broad class of processes used for established photonic components.
“Ready to scale” still requires care. The research team has demonstrated a working cell and simulated pathways toward larger pSRAM arrays. A manufacturable process does not automatically deliver competitive density, yield, endurance, or control electronics. It does, however, give the design a more credible route beyond a one-off laboratory device.
A practical maturity ladder for photonic memory
“Photonic memory” describes several technologies with very different purposes. A useful way to understand the field is to ask what job the memory performs and how close it is to a complete chip.
| Approach | What it stores | Current strength | Main limitation |
|---|---|---|---|
| Optical discs and archival storage | Long-term encoded data | High capacity and commercial maturity | Not designed for processor-speed rewriting |
| Phase-change photonic memory | Material states read optically | Potentially nonvolatile and multi-level | Endurance, switching energy, and integration |
| Programmable photonic latch | Temporary bits represented by optical signals | Demonstrated set, reset, hold, and fast response | Dedicated chips and larger arrays remain future work |
| Regenerative pLatch and pSRAM | Stable optical bits designed for logic and in-memory computing | Foundry-compatible cell, concurrent operations, simulated scaling | Density, sustained yield, and full electronic integration |
A second 2025 milestone illustrates why several tracks are advancing in parallel. Nokia Bell Labs researcher Farshid Ashtiani demonstrated a programmable photonic latch using optical logic gates and micro-ring modulators. Its reported response was in tens of picoseconds, and its wavelength selectivity was compatible with wavelength-division multiplexing.
That project and the USC–Wisconsin device use related latch concepts but different architectures. Together they show that researchers are exploring more than one route to optical volatile memory rather than relying on a single material or mechanism.
How photonic computing differs from quantum computing
Photonic computing and quantum computing both use light, but they solve different problems. Photonic computing normally manipulates classical optical signals. It can use interference, phase, amplitude, wavelength, and parallel paths to perform operations such as matrix multiplication.
Quantum computing encodes information in quantum states such as qubits. Those systems exploit entanglement, superposition, and interference in ways that classical photons do not. A photonic quantum processor may also need specialized sources, detectors, and memories for single photons or quantum states.
The pLatch is therefore an optical classical-memory device, not a quantum memory or quantum computer. It is relevant to artificial intelligence and high-performance computing because both rely heavily on large numbers of matrix operations that could pair optical computation with optical memory.
What could use photonic computing memory first?
AI accelerators and specialized optical processors are the most plausible early targets because both repeatedly read weights and intermediate data. In a conventional system, those transfers repeatedly cross the memory wall. A photonic processor that can hold some values as light could reduce the distance they travel and the number of electronic conversions they require.
Other possibilities include:
- Photonic tensor cores that multiply and accumulate optical signals near their stored data.
- Optical networking chips that temporarily buffer high-speed streams without first converting them to electronic memory.
- Neuromorphic systems that need many parallel, rapidly updated values.
- Scientific and optimization hardware built around matrix operations and iterative calculations.
Large hybrid systems already show why electronics will remain important. A 64 by 64 photonic accelerator published in Nature in 2025 integrated more than 16,000 photonic components, but it also used a separate electronic chip for control, logic, and SRAM. Its authors explicitly identified the absence of optical-domain storage as a limitation requiring electronic–photonic co-integration.
NVIDIA is already commercializing photonics for data center networking through co-packaged optical switches. Its silicon-photonics platforms place optical components close to network silicon, improving bandwidth and power efficiency. Photonic computing memory would address a different layer: temporary storage and computation within advanced processors.
What still stands between the prototype and a commercial computer?
The new pLatch is an important component, not a complete memory hierarchy. Electronic systems combine several classes of memory, from tiny, fast caches to slower, larger storage. A photonic computer would need an equivalent set of components with an electrical control layer that addresses, refreshes, checks, and protects the optical state.
Five hurdles stand out:
- Density: each optical latch may require more physical area than a compact electronic memory cell.
- Array yield: one imperfect cell is manageable; a processor needs thousands or millions working together.
- Endurance: a useful processor memory must survive a vast number of write and read cycles.
- Control and conversion: electronics still need to configure the optical circuit, verify data, and connect it to conventional processors.
- System efficiency: lasers, modulators, detectors, cooling, and supporting electronics must be counted in any energy comparison.
The minutes-long hold demonstrated by the pLatch is sufficient to prove that the optical state is stable, but it is not comparable to long-term archival storage. Nor does a simulated frequency describe the sustained performance of a fully packaged processor. These distinctions are essential when assessing claims about photonic RAM or optical caches.
Frequently asked questions
Is photonic computing real?
Yes, but most systems are hybrid. Photonic processors and accelerators are real research hardware, while light is already used widely for data center communication. The most credible near-term systems divide work between optical components that excel at bandwidth and parallel operations and electronic components that excel in control, dense memory, and general logic.
What is photonic computing memory?
Photonic computing memory is temporary or persistent digital information represented by an optical property, such as a light pulse, wavelength, phase, resonance, or material state. A practical processor memory must support controlled writing, stable holding, reading, addressing, and rewriting.
Is a photonic latch the same as photonic RAM?
A latch can be a fundamental building block inside an SRAM-style array, but it is not a complete RAM chip. A usable memory also needs many cells, addressing, control circuits, input and output paths, packaging, and methods for managing errors.
Will photonic memory replace DRAM and flash?
Not soon, and it does not have to. Electronic memory remains exceptionally dense, mature, and durable. Photonic memory is more likely to appear first beside electronic processors, storing temporary values for optical operations or buffering data inside hybrid accelerators.
Is NVIDIA building a photonic processor with this memory?
NVIDIA is publicly deploying silicon photonics in networking products and investing heavily in optical infrastructure for AI. That work is separate from the USC–Wisconsin pLatch. The research demonstrates a component that future optical processors could use, but it is not an NVIDIA product.
Could photonic memory make quantum computers faster?
Potentially, but not automatically. Quantum processors need memories tailored to quantum states and errors that differ from classical optical bit storage. The pLatch does not solve those requirements. Some future quantum-photonic systems may use related optical control techniques, but they would be separate technologies.
A missing component becomes testable
The 2025 pLatch did not make optical computers practical overnight. Its value is that it turns a long-standing systems problem into a component that can be fabricated, measured, and improved using an established foundry process.
Photonic computing already has real applications in optical networking and working research prototypes such as accelerators, specialized processors, and hybrid systems. What has been missing is a convincing path toward temporary memory that can remain optical long enough to be useful. The new photonic computing memory research supplies a building block for that path while also showing why density, endurance, control, and full-system efficiency will determine whether light eventually moves from the network into a processor’s working memory.
