TOKYO – March 30, 2023 – Preferred Networks, Inc. (PFN) today released CuPy™ v12, the new major update to the open-source library for general-purpose matrix calculation.

CuPy v12 provides the following new features:

• SciPy API-compatible interpolation module
  cupyx.scipy.interpolate, a new module providing interpolation algorithms that can be used in the same way as SciPy, has been implemented. This will ease GPU application in various data analysis tasks and preprocessing, including missing data prediction and smoothing. This module was developed under the support of the Essential Open Source Software for Science program of the Chan Zuckerberg Initiative.

• Thrust/CUB support in JIT kernels
  APIs for Thrust/CUB libraries to be used in JIT kernels have been added. This feature will enable users to quickly define performant CUDA kernels using Python code. The implementation is based on the work done in the PFN Summer Internship 2022 Program.

• Support for the latest NVIDIA platforms
  The latest NVIDIA platforms, including CUDA 12, NVIDIA H100 Tensor Core GPU, and Jetson Orin, are now supported.

Please check the announcement in the CuPy blog for more details.
https://medium.com/cupy-team/released-cupy-v12-4497315e811e

PFN will continue to swiftly incorporate the latest research outcomes into CuPy while collaborating with supporting companies and open source communities.