PyTorch Day India 2026 Exploring Tile based Programming Abstractions for KLA’s Image Processing Work

KLA's process control tools use image processing algorithms to identify critical defects during semiconductor manufacturing ensuring that the manufacturing process achieves and maintains high yield. Deploying these algorithms on NVIDIA GPUs at near speed-of-light performance often requires hand-tuning the kernels using CUDA C++ given the unique nature of loads. This, however, poses several challenges with development and maintenance owing to the complexity of the developed kernels. NVIDIA's cuTile is a higher-level abstraction to author GPU kernels, developed to address these challenges and presents an interesting trade-off between performance, productivity, and maintainability. In this talk, we will discuss our work on implementing some commonly used image-processing kernels on NVIDIA's cuTile and share our experience on development efforts, and achieved performance.