CUDA thirteen point three arrives with stable Python one point zero for GPU

Published on May 29, 2026 | Translated from Spanish

NVIDIA has released CUDA 13.3, an update that marks a milestone with the arrival of CUDA Python 1.0. This version offers stable support for running GPU code from Python, targeting artificial intelligence, data science, and scientific computing. The unified GPU programming platform becomes more accessible for developers seeking performance without leaving their favorite ecosystem.

GPU architecture cross-section during CUDA 13.3 execution, Python code blocks transforming into parallel processing streams, multiple CUDA cores illuminated in blue and orange, data flowing through memory hierarchy, AI tensor cores actively processing, glowing circuit pathways connecting Python interpreter to GPU hardware, technical illustration style, photorealistic engineering visualization, dramatic volumetric lighting, complex chip die pattern with visible transistor structures, ultra-detailed semiconductor textures, clean industrial aesthetic, high contrast metallic surfaces

Stable Python 1.0: fewer wrappers, more cores 🚀

CUDA Python 1.0 eliminates the need for external wrappers or intermediate libraries like PyCUDA. Developers can now invoke CUDA kernels directly from Python scripts, with native typing and optimized memory management. The API exposes functions for controlling threads, blocks, and streams, enabling fine-tuning of parallelism. Additionally, it integrates with libraries like NumPy and cuDF, making it easier to migrate prototypes to production without losing performance compared to C++.

And to think we used to use for loops like cavemen 😅

Now any data scientist can feel like a GPU engineer without switching editors. Just import cuda and launch kernels as if they were normal functions. Of course, then comes the debugging moment and you realize your Python script is still Python: slow to start, but fast when CUDA kicks in. At least you no longer have to make excuses that Python is slow; now you can blame your graphics card for not having enough cores.