Photonic computing chips, enabled by on chip light propagation and multidimensional multiplexing such as wavelength division, time division, and space division multiplexing, offer high bandwidth, low latency, and inherent parallelism. These features make them promising for high throughput and energy efficient computing. However, their advantages are still difficult to fully realize in complex intelligent tasks. In all optical computing systems, keeping the main computation in the optical domain can help the system approach the physical speed limit. Yet complex tasks require powerful nonlinear operations, programmability, scalability, and low noise, which remain difficult to achieve purely in optics. Adding electronic modules can improve task capability, but optical to electrical conversion, electrical to optical conversion, sampling, quantization, and electronic control loops introduce extra latency and power consumption. This forms the intrinsic speed-complexity trade-off in photonic computing chips. In a Research Highlights article published in Journal of Semiconductors, the team from Shanghai Jiao Tong University reviews recent progress in photonic computing chips, including architectures based on on chip interference, on chip diffraction, and microring resonators, as well as photonic electronic chips for high speed visual tasks and general computing. The article further highlights LightGen, an all-optical synthesis chip for large-scale intelligent semantic vision generation. By integrating millions of photonic neurons on a chip, implementing all-optical dimension variation via the proposed optical latent space (OLS), and developing BOGT, a Bayes-based algorithm for Optical Generative model Training that enables ground-truth-independent training, LightGen experimentally implemented complex generative tasks, including high-resolution semantic image generation, 3D generation and manipulation comparable to NeRF, high-resolution video generation and semantic manipulation, denoising, and style transfer. LightGen shows that, through the joint innovation of optical representation mechanisms and chip integration, all-optical chips may move beyond acceleration of individual operators and develop into physical computing platforms for complex generative intelligence.