An intelligent adaptive optics processing platform based on GPU and multi-core CPU architectures

Adaptive optics (AO) enables real-time control of optical wavefronts, but traditional algorithms face limits in computation and noise resistance. Recently, deep learning has shown advantages in nonlinear fitting, yet challenges remain in largescale data processing and real-time deployment. Most studies stop at simulation or offline validation, with few addressing practical system integration. To advance deployment in real AO systems, this paper presents an intelligent AO platform using GPU and multi-core CPU. The platform features: (a) CUDA-based acceleration for Hartmann image preprocessing; (b) a lightweight neural network with TensorRT inference and INT8 quantization; and (c) a real-time controller (RTC) supporting both traditional and intelligent correction. Closed-loop experiments were performed with indoor point targets and a rotating phase screen. For Hartmann images of 528×519 pixels and 396 sub-apertures, combined with a 451-actuator deformable mirror, the system achieved an average latency of 436 μs and stable closed-loop operation at 2000 Hz. Under a phase screen with average aberration RMSE of 402.69 nm, the intelligent algorithm reduced residual RMSE to 28.85 nm, outperforming the centroid-based algorithm’s 42.06 nm. These results confirm that the platform meets real-time performance and correction accuracy requirements, demonstrating engineering feasibility of intelligent AO.