Researchers at the University of Sydney have created an optical computing chip that performs neural network calculations using photons instead of electrons, demonstrating processing speeds more than 1000 times faster than electronic chips for certain machine learning operations.

The photonic processor performs matrix multiplications, the fundamental mathematical operation in neural networks, by encoding data in properties of laser light and using optical components to manipulate those properties. The computations happen at light speed without generating the heat that limits electronic processor performance.

While still a research prototype, the technology points toward future AI systems that could process vast amounts of data orders of magnitude faster than today’s chips while consuming far less energy. Applications could include real-time video analysis, massive language models, and scientific simulations.

Why Optical Computing

Electronic computing faces fundamental limits. As transistors shrink and processor speeds increase, heat generation and power consumption become constraining factors. Moving data between processors and memory consumes energy and creates bottlenecks.

Optical computing sidesteps some of these limits. Photons can pass through each other without interfering, allowing parallel processing that’s difficult with electrons. Optical operations occur at light speed, about 100,000 times faster than electrical signals in silicon.

Light also doesn’t generate heat the way electrical current does. Heat dissipation limits how densely electronic components can be packed and how fast they can operate. Optical components can potentially operate faster and at higher densities.

However, optical computing has faced challenges that have prevented practical systems despite decades of research. Building optical components with the precision needed for computing is difficult. Interfacing optical processors with electronic memory and input/output systems adds complexity.

Professor David Miller, who leads the Sydney research group, said recent advances in photonics manufacturing have finally made optical computing practical for specialized applications like neural network inference, even if general-purpose optical computers remain distant.

The Sydney Chip

The processor uses silicon photonics, where optical components are fabricated on silicon wafers using processes similar to electronic chip manufacturing. This allows precise optical structures and integration with electronic controls.

Input data is encoded in the amplitude and phase of laser light at different wavelengths. Optical filters and interference structures perform mathematical operations on this light. The result is detected by photodetectors and converted to electronic signals.

Matrix multiplication, which dominates neural network computation, maps naturally to optical operations. A matrix can be encoded in an optical interference pattern, and multiplying by that matrix involves passing light through the pattern. This happens nearly instantaneously regardless of matrix size.

The prototype chip performs 8-bit integer matrix multiplications on 256×256 matrices at rates exceeding 1 trillion operations per second. Electronic GPUs perform similar operations, but the optical chip does so with 100 times lower energy consumption and no cooling required.

Accuracy is currently limited compared to electronic processors. The optical operations are analog, meaning they’re subject to noise and imprecision. The researchers achieve effective 8-bit precision, adequate for many neural network applications but less than the 16-bit or 32-bit precision electronic systems typically use.

Performance Characteristics

The speed advantage is largest for certain types of neural network architectures. Convolutional neural networks used for image processing benefit greatly because they involve many matrix multiplications with relatively small matrices. Large language models that use enormous matrices also benefit.

For simpler computations that don’t involve matrix multiplications, the optical chip offers little advantage over electronics. It’s specialized for a specific type of calculation rather than being a general-purpose processor.

Scaling to larger problems faces challenges. The prototype handles 256×256 matrices, but real neural networks often involve matrices with thousands or millions of elements. Optical systems for such large matrices would be physically very large or require breaking calculations into smaller pieces.

Integration with electronic systems creates bottlenecks. The optical chip performs calculations quickly, but getting data in and out still uses electronic interfaces operating at conventional speeds. For small problems where data transfer dominates, the optical advantage disappears.

The researchers achieved the best results for “edge AI” applications where neural networks process sensor data in real-time. Autonomous vehicles, augmented reality, and industrial inspection all need fast neural network inference on modest-sized problems, precisely where the optical chip excels.

Path to Practical Systems

Commercialising optical computing requires solving numerous engineering challenges. Manufacturing yields for photonic components are lower than for electronic chips, increasing costs. Testing and calibration is more complex. Integration with standard electronic systems requires interface chips.

The Sydney team is working with Australian and international manufacturing partners on production-ready designs. Moving from laboratory prototypes to manufactured products typically takes 3-5 years and requires substantial capital investment.

Several commercial applications are being targeted. Data centres running AI workloads could use optical processors for specific tasks where the speed and energy advantages justify the cost. Edge devices like autonomous vehicles or industrial robots could integrate optical processors for real-time perception.

Military and aerospace applications have shown particular interest because the low heat generation makes optical processors attractive for power-constrained platforms like satellites or UAVs. These applications can justify higher costs for performance advantages.

The research group has spun out a startup company, Photon Intelligence, to commercialise the technology. The company has raised initial venture funding and is developing engineering prototypes for customer trials. Full commercial products are targeted for 2027-2028.

Australian Photonics Ecosystem

Australia has substantial photonics research capability spread across universities and CSIRO. Sydney, ANU, RMIT, and Macquarie University all have strong photonics groups working on telecommunications, sensing, quantum technologies, and computing applications.

The country has been less successful at commercialising photonics research. Several companies including Oclaro and Finisar operated Australian facilities but eventually closed or moved operations offshore. Building sustainable photonics manufacturing in Australia has proven difficult.

Recent government initiatives aim to strengthen photonics commercialisation. The Adelaide-based Australian Photonics Centre provides facilities and support for photonics startups. The IMCRC (Innovative Manufacturing Cooperative Research Centre) supports photonics manufacturing research and industry partnerships.

Whether these efforts succeed depends partly on whether applications like optical computing create sufficient market pull to justify Australian manufacturing investment. The global photonics market is large and growing, but most manufacturing happens in Asia and North America.

Comparison to Electronic AI Accelerators

Electronic AI chips have improved dramatically over the past decade. Companies including NVIDIA, Google, and numerous startups have developed specialised processors that perform neural network calculations far more efficiently than general-purpose CPUs.

These electronic accelerators use similar concepts to optical computing, including specialized matrix multiplication units and aggressive parallelism. They’ve improved energy efficiency by factors of 100-1000 compared to earlier electronic approaches.

Optical computing must compete not against CPUs but against these specialised electronic AI chips. The bar for competitive advantage is high. Optical systems need to offer substantial benefits to justify different manufacturing and integration challenges.

The Sydney research suggests optical approaches can win for specific applications by achieving better energy efficiency and eliminating cooling requirements. This is valuable even if raw performance is similar to electronic alternatives.

There’s also potential for hybrid systems combining electronic and optical processors. Electronics handle control, memory, and general computation, while optical components accelerate specific operations. Several research groups including the Sydney team are exploring hybrid approaches.

Fundamental Research Questions

Understanding the theoretical limits of optical computing requires ongoing research. What’s the minimum energy required for optical computation? How small can optical components become before quantum effects dominate? What precision is fundamentally achievable?

These questions matter because they determine whether optical computing offers lasting advantages over electronics or just temporary benefits that advancing electronic technology will eliminate.

Current results suggest that for matrix multiplications specifically, optical approaches have fundamental advantages in speed and energy efficiency. Whether this extends to other computations or system-level performance remains active research.

Machine learning researchers are also investigating whether neural networks can be redesigned to better exploit optical computing’s strengths and work around its limitations. Network architectures co-designed with hardware could extract more performance than adapting existing architectures.

Global Context

Optical computing research is active internationally. MIT, Stanford, Oxford, and several Asian universities have substantial programs. Companies including Lightmatter and Luminous Computing are pursuing commercial optical AI chips.

Different groups emphasize different approaches. Some focus on extreme miniaturisation using integrated photonics. Others use free-space optics with physically larger but more flexible systems. Each approach involves different trade-offs.

The Sydney group’s silicon photonics approach benefits from compatibility with existing semiconductor manufacturing, potentially enabling lower costs at scale. But it faces limits in how many optical components can be integrated.

International collaboration is common in optical computing research, with frequent researcher exchanges and shared publications. The field is small enough that most researchers know each other, fostering cooperation alongside competition.

Future Directions

Extending optical computing beyond matrix multiplication to other operations would expand applicability. Some neural network operations like activation functions and normalisation don’t map naturally to optical implementations. Research addresses how to handle these operations efficiently.

Improving precision while maintaining speed advantages is another priority. Many applications need higher precision than current optical systems provide. Hybrid analog-digital approaches might provide better precision without sacrificing all speed advantages.

Scaling to much larger problems requires new architectures. The researchers are investigating modular approaches where multiple optical chips work together, and hierarchical designs where different operations occur at different scales.

For Australian research and industry, optical computing represents an opportunity in an emerging technology sector where Australian capabilities could compete globally. Whether that potential translates to commercial success depends on sustained research funding, entrepreneurial skill, and some luck in navigating the difficult path from laboratory to market.

The University of Sydney demonstration shows that optical neural network processors can deliver dramatic speed and efficiency improvements for specific tasks. Whether this laboratory success leads to practical systems depends on engineering development over coming years. The fundamental principles are proven; turning them into products is the next challenge.