CUDA-Python Reaches "GA" With NVIDIA CUDA 11.5 Release, __int128 Preview
NVIDIA has made available CUDA 11.5 today as the latest version of their popular but proprietary compute stack/platform. Notable with CUDA 11.5 is that CUDA-Python has reached general availability status.
NVIDIA CUDA 11.5 was posted today along with updated device drivers for Windows and Linux systems. Some of the CUDA 11.5 highlights include:
-CUDA-Python has reached "GA" as the Python bindings for CUDA. More details along with tests and samples of CUDA-Python can be found via GitHub.
- A preview release of the "__int128" data type. This is a limited preview right now with support for math operations, library support, and more to come later.
- Native support for signed and unsigned normalized 8-bit and 16-bit types.
- Improved interoperability with graphics frameworks.
- CUDA now supports per-process memory access policies around multi-process sharing of GPUs.
- cuBINS larger than 2GB can now be linked.
- The device-side caching behavior is now configurable.
- The CUDA compiler now supports "-arch=all" and "-arch=all-major" options for generating code for multiple architectures at the same time.
Overall the CUDA 11.5 release today is a fairly robust update to the CUDA 11 series. Windows/Linux downloads for CUDA 11.5 available from developer.nvidia.com. More details on the CUDA 11.5 changes via the release notes.