For those wanting to tap the power of CUDA from the Python programming language, CUDA 13.3 marks the CUDA Python 1.0 milestone as a stable, supported means of being able to leverage CUDA in Python apps for AI, data science, scientific computing, and related uses.
For C++ fans, CUDA 13.3 brings CUDA Tile for C++ in bringing the CUDA Tile programming model to the C++ world.
In addition to these programming enhancements, CUDA 13.3 also introduces the CompileIQ compiler auto-tuning framework that can provide up to 15% speed-ups on kernels like GEMM and attention.
CUDA 13.3 also brings a Numba CUDA MLIR back-end , various math library updates, C++23 support in the NVCC and NVRTC code, mmap() support, and other improvements.
More details on this CUDA 13.3 feature update via the NVIDIA Developer Blog .