Researchers from Xidian University have successfully developed a two-chip photonic computing system that overcomes critical barriers to real-time learning in spiking neural networks. Led by Shuiying Xiang, the team created a system capable of performing both linear and nonlinear computations entirely within the optical domain, eliminating the need for energy-intensive and slow conversions between light and electricity. Traditional photonic spiking neural systems have been limited by their inability to handle nonlinear calculations optically. This constraint typically required signals to be converted back to electronic formats for learning and decision-making steps, introducing delays and negating the speed and energy advantages of photonics. The new system, described in the journal Optica, utilizes a 16-channel photonic neuromorphic chip featuring 272 trainable parameters. This architecture allows for the simultaneous processing of multiple optical signal streams and the dynamic adjustment of connections through learning. The research team addressed three primary challenges: the lack of large-scale, low-threshold nonlinear photonic spiking neuron arrays, the absence of fully programmable photonic spiking neural network chips, and the uncertainty regarding hardware implementation of photonic spiking reinforcement learning. To solve these issues, they engineered a 16x16 Mach-Zehnder interferometer mesh chip tailored for spiking networks and a chip containing a distributed feedback laser array with a saturable absorber, optimized for low-threshold nonlinear activation. To ensure accurate performance, the team developed a collaborative hardware-software framework. This approach allows models to be pre-trained in software, fine-tuned directly on the chips to account for physical variations, and further optimized via software. The system was tested using standard reinforcement learning benchmarks: balancing a pole on a moving cart (CartPole) and swinging a pendulum from a hanging to an upright position (Pendulum). The results demonstrated remarkable fidelity between software simulations and hardware execution. Hardware decisions matched software accuracy with only a 1.5% drop in the CartPole task and a 2% drop in the Pendulum task. In practice, the system achieved perfect performance on the CartPole task and robust results on the more complex Pendulum scenario, proving its capability to handle both simple and demanding learning tasks in real time. In terms of efficiency, the chips deliver performance comparable to high-end GPUs. The system achieved an energy efficiency of 1.39 tera operations per second per watt for linear computation and maintained high density across nonlinear tasks. Most notably, the on-chip computing latency was measured at just 320 picoseconds, enabling ultra-fast decision-making essential for applications like autonomous driving and embodied robotics where milliseconds matter. Looking ahead, the researchers aim to scale this technology further by designing a 128-channel fully functional photonic spiking neural network chip. This next generation of hardware is intended to tackle even more complex tasks, such as neuromorphic autonomous navigation. Additionally, the team plans to demonstrate compact, hybrid-integrated large-scale chips to pave the way for practical deployment in edge computing scenarios, bringing advanced, light-based artificial intelligence closer to real-world implementation.