TOKYO – April 22, 2021 – Preferred Networks, Inc. (PFN) today released CuPy™ v9, the new major update to the open-source library for general-purpose matrix calculation.

CuPy v9 provides the following new features:

1. JIT API
CuPy provides a JIT API to write CUDA kernels in Python. Kernels compiled through this API achieve the same performance level as CUDA kernels directly written in C/C++. 

2. NVIDIA cuSPARSElt Support
CuPy now integrates the Ampere architecture support for sparse matrix multiplications by using the cuSPARSElt library.

3. Windows Support
Windows support is now officially provided as CI configurations and testing configurations have been deployed.

4. Support for AMD GPUs (ROCm)
ROCm Platform can now be used in CuPy and both binary packages and Docker images are provided.

5. Improved NumPy/SciPy Compatibility
The CuPy API has been extended to support Generalized Universal Functions and the new Random Generator API introduced in NumPy 1.17.

CuPy v9 incorporates numerous improvements contributed by the CuPy community, including many from NVIDIA Corporation and U.S. Department of Energy Brookhaven National Laboratory’s CSI-HPC group.

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