ZLUDA is an experimental project that offers a drop-in replacement for CUDA on Intel GPU. It allows running unmodified CUDA applications on Intel GPUs without modification with near-native performance. While ZLUDA is slower in the Geekbench by approximately 2%, it uses Level 0, a low-level, high-performance API, and does not support asynchronous execution, giving it an unfair advantage in benchmark tests. ZLUDA is a private project and is not an Intel or NVIDIA project. It supports Gen9 and newer Intel GPUs and may work with AMD GPUs. The project is not complete, and users should not rely on it for anything serious.
https://github.com/vosen/ZLUDA