The chips operate using a form of AI known as a photonic spiking neural system. These systems mimic biological neurons, which communicate through rapid pulses. In the photonic version, those signals travel as light through optical circuits rather than electrical currents.
According to the research team led by Shuiying Xiang at Xidian University in China, the new architecture removes a key limitation that previously slowed the development of photonic AI.
Photonic neural computing
Photonic spiking neural systems represent neural activity using brief optical pulses known as spikes. These optical signals can move through photonic circuits significantly faster than electrical signals, offering potential advantages in both speed and energy efficiency.
However, earlier photonic systems were limited because they could only handle linear computations using light. The nonlinear operations necessary for learning and decision-making still required conversion into electronic signals, introducing delays and reducing efficiency.
Xiang explained that previous systems needed electronic hardware to perform these nonlinear steps, which undermined the speed and energy benefits of photonics.
The new design removes this bottleneck by enabling both linear and nonlinear neural computations to be performed entirely in the optical domain.
Two-chip photonic platform
To demonstrate the concept, the researchers created a programmable photonic neuromorphic platform consisting of two chips working together.
The first chip houses a 16-channel photonic neuromorphic processor with 272 trainable parameters. It can process multiple optical signals simultaneously.
The second chip contains a distributed feedback laser array combined with a saturable absorber. This component generates nonlinear optical spiking with a low activation threshold, enabling the system to perform learning operations directly with light.
Reinforcement learning experiments
The team tested the system using reinforcement learning, an AI method where models learn through trial and error.
Engineers initially trained the neural model in software. The photonic chips then carried out the hardware training and execution, while additional software adjustments fine-tuned results to compensate for small hardware variations.
Researchers evaluated the system using two classic control problems widely used in AI research.
One was the CartPole task, which involves balancing a pole on a moving cart. The other required stabilizing an inverted pendulum.
Results showed that hardware decisions closely matched the software model. Accuracy decreased by only about 1.5 percent for the CartPole task and roughly 2 percent for the pendulum test.
Performance results
The system demonstrated strong computational performance. Photonic linear processing reached about 1.39 tera operations per second per watt, while nonlinear computation achieved nearly 988 giga operations per second per watt.
On-chip computing latency was measured at just 320 picoseconds, highlighting the potential speed advantages of photonic AI systems.
Toward future photonic AI hardware
Researchers believe the technology could help build AI systems that require rapid learning and extremely low energy consumption.
Potential applications include autonomous driving, robotics, and embodied intelligence systems capable of adapting to real-world environments.
The current prototype operates with 16 optical channels, but the research team plans to scale the design further. Future versions may include a 128-channel photonic spiking neural chip capable of supporting more complex reinforcement learning tasks.
The team is also exploring compact hybrid photonic systems designed for edge computing devices.
If the technology continues to advance, photonic processors could eventually offer a powerful alternative to traditional electronic AI hardware in next-generation intelligent machines.
