Optical Interconnect Breakthrough Accelerates AI Inference
China’s answer to AI’s growing appetite for computing power may involve moving data with light instead of adding more GPUs. Researchers at Peking University have developed an optical interconnect system that reportedly made distributed AI inference more than 100 times faster while using one-ninth of the usual computing resources, according to a report in the South China Morning Post. This early-stage research could give China and the wider APAC region another route to faster, more energy-efficient AI infrastructure as data center operators confront rising power demands, hardware costs, and processor supply constraints.
The breakthrough addresses one of the most persistent bottlenecks in modern AI: the time and energy spent transferring data between processors. As neural networks grow larger and more complex, the communication overhead between chips has become a critical factor limiting overall performance. Traditional electrical interconnects, while reliable, are constrained by physical limitations such as resistance and capacitance, which cause signal degradation and heat generation at high speeds. Optical interconnects, by contrast, use light to transmit data, offering dramatically higher bandwidth and lower latency with significantly less energy consumption.
How the Optical Interconnect System Works
According to the SCMP report, the Peking University team connected standard electronic chips using custom optical hardware and algorithms. The system used field-programmable gate arrays (FPGAs), programmable chips commonly used in data centers, autonomous vehicles, and other applications that require high levels of parallel processing. A silicon photonic transceiver handled the conversion between electrical and optical signals at 400 gigabits per second. A second component managed communication among the chips, allowing data to travel over optical links rather than relying solely on slower electrical connections.
Inside AI reported that the system maintained 99.5% of the accuracy achieved by a single-chip setup while processing data at about 100 times the rate of a comparable electrical system. The approach targets one of AI infrastructure’s most stubborn problems. As models grow larger, processors spend more time and energy transferring data among chips. Adding GPUs can increase capacity, but it also raises power use, cooling requirements, and equipment costs.
China’s Drive for Alternative AI Scaling Methods
The research is particularly relevant to China, as its technology companies and data center operators seek additional computing capacity without relying solely on larger GPU clusters. With export restrictions limiting access to advanced processors from companies like Nvidia and AMD, Chinese developers have been forced to innovate. Optical interconnects could eventually help Chinese AI developers move data between processors faster while reducing the amount of computing hardware required for inference. Greater efficiency would be valuable as operators manage electricity use, cooling demands, equipment costs, and access to advanced processors.
Compatibility with commonly used FPGAs also gives the design a practical advantage over optical computing systems built around entirely custom processors. Existing data centers would still need significant hardware changes before they could use the technology at scale. However, the ability to integrate with widely available FPGAs lowers the barrier to adoption compared to fully custom silicon photonic solutions.
Technical Foundations: Silicon Photonics and FPGAs
Silicon photonics is a technology that uses silicon as an optical medium, enabling the fabrication of photonic components using existing CMOS manufacturing processes. This compatibility with standard semiconductor fabrication could reduce costs and accelerate commercialization. The transceiver developed by the Peking University team operates at 400 Gbps, which is competitive with the fastest electrical interconnects currently in use. By using light for inter-chip communication, the system avoids the signal integrity issues that plague electrical links at high data rates.
FPGAs are well-suited for this application because they can be reprogrammed to implement custom communication protocols and data processing pipelines. The Peking University researchers developed algorithms that optimized the distribution of inference tasks across multiple FPGAs, minimizing the amount of data that needed to be transferred and maximizing the utilization of each chip. This software-hardware co-design was critical to achieving the reported 100x speedup.
Comparison with Electrical Interconnects
Traditional electrical interconnects rely on copper traces on circuit boards or within chips. As data rates increase, these traces suffer from skin effect, dielectric losses, and crosstalk, all of which degrade signal quality. To compensate, designers often use equalization and error correction, which add latency and power overhead. Optical interconnects, on the other hand, are immune to electromagnetic interference and can carry multiple wavelengths of light simultaneously, enabling wavelength-division multiplexing to increase bandwidth further.
In the Peking University demonstration, the optical interconnect achieved a 100x improvement in inference throughput while using one-ninth the computing resources. This means that the same workload that would require nine conventional clusters could be performed by a single cluster with optical interconnects. The energy savings are also significant: because optical links consume less power per bit transmitted, the overall power demand for data transfer drops sharply, reducing the thermal load on data centers.
Current Limitations and Future Prospects
The system remains a laboratory demonstration rather than a production-ready platform. Researchers must still miniaturize and package the optical components, test their reliability over long periods, and prove that the performance gains hold across much larger clusters. Silicon photonic hardware may also be more difficult and expensive to manufacture than conventional electrical connections, though economies of scale could bring costs down as the technology matures.
Inside AI reported that the system still needs to prove it can maintain the same performance across much larger clusters with thousands of connected nodes. Chinese cloud providers and enterprises should view the findings as a possible direction for future infrastructure, rather than as equipment they can purchase today. The study suggests China could increase AI inference capacity through more efficient chip-to-chip communication, but commercial deployment will depend on whether the design can preserve its latency, accuracy, and resource savings at scale.
Broader Implications for AI Infrastructure
The potential impact of this research extends beyond China. Data centers worldwide are straining to meet the demands of AI training and inference. In 2025, the International Energy Agency reported that data centers accounted for about 2% of global electricity consumption, a figure that could rise sharply as AI adoption accelerates. Technologies that improve the efficiency of inter-chip communication could be as important as advancements in processor performance.
Optical interconnects are already used in some high-performance computing systems, such as the Fugaku supercomputer in Japan and the Summit system in the United States. However, these implementations are typically custom and expensive. The Peking University work demonstrates that optical interconnects can be integrated with commodity FPGAs, potentially making the technology more accessible.
If the technique can be scaled, it could enable distributed AI inference across many nodes without the communication overhead that currently limits performance. This would be particularly beneficial for large language models and multimodal AI systems that require massive parallel processing.
Potential Competition and Collaboration
Other research groups and companies are also exploring optical interconnects for AI. For instance, the French company CEA-Leti has developed silicon photonic transceivers for data centers, and the University of California, Berkeley has worked on optical networks for AI accelerators. The Peking University team’s approach stands out because of the emphasis on combining optical hardware with FPGA-based processing, which could enable faster adoption in existing infrastructure.
Collaboration between Chinese research institutions and global partners could accelerate progress, but geopolitical tensions may complicate such cooperation. Nonetheless, the fundamental science of optical interconnects is well understood, and the main challenges are in engineering and manufacturing. With continued investment, the technology could reach commercial viability within the next three to five years.
Next Steps for the Peking University Team
The researchers are now focused on improving the packaging and reliability of the optical components. They also plan to test the system with larger clusters, potentially using cloud-scale testbeds. Another area of interest is developing more efficient algorithms for load balancing and data distribution that can fully exploit the potential of optical interconnects. Additionally, they are exploring ways to integrate the optical transceiver directly into the FPGA package, reducing the distance between the chip and the optical interface and further improving performance.
Given the rapid pace of AI development, the demand for faster and more efficient interconnects will only grow. The Peking University optical interconnect system represents a promising step toward a future where AI systems are not limited by the speed of electrical data transfer.
Source: eWeek News