Python library for Cuda-X from Nvidia: fast algebraic calculations
Pyhton projects can perform fast algebraic calculations with low overhead using the new Nvidia library for Cuda-X.
(Image: sdecoret/Shutterstock.com)
Nvidia has released a new mathematical Python library specialized for Cuda-X. It offers direct, Python-like access to the mathematical core operations of Cuda-X without having to use additional C/C++ libraries. This is useful for the development of hardware-accelerated applications, libraries, frameworks or deep learning compilers.
In terms of mathematical functions, nvmath-python supports linear algebra and n-dimensional, discrete Fourier transformations in particular. Nvidia promises very good performance for these calculations, which should come close to that of the native C libraries. The overhead is minimized and allows, for example, the combination of several cores (kernal fusion) without an extra host.
The library is licensed under Apache 2 and offers stateless (function style) or stateful (class style) APIs. It integrates into the standard logger library in order to be able to track the details of the calculations. In combination with Python compilers such as Numba, device callbacks can be generated that adapt the behavior of nvmath-python itself.
According to Nvidia, nvmath cooperates seamlessly with the Python ecosystem, both with other GPU-oriented packages such as CuPy, PyTorch or RAPIDS, as well as with CPU libraries such as SciPy, scikit-learn or NumPy. A new major version of the latter has also recently been released.
Further information on nvmath-python can be found on the project's website, on GitHub or in a demo video.
(who)