TOKYO – December 9, 2021 – Preferred Networks, Inc. (PFN) today released CuPy™ v10, the new major update to the open-source library for general-purpose matrix calculation.

CuPy v10 provides the following new features:

1. Multi-node and multi-GPU communication improvements
The addition of the cupyx.distributed module allows collective/peer-to-peer communication of arrays between processes using different GPUs in either shared memory or distributed environments. Also, when possible, peer-access is enabled by default to transparently access arrays residing in other GPU’s memory. These features make it easy to perform parallel computation and utilize multiple GPUs in CuPy.

2. Arm architecture support
Arm platform (aarch64) is now officially supported. Newly provided binary packages for NVIDIA Jetson will enable the use of CuPy in edge environments.

3. Array API Standard support
CuPy v10 adds cupy.array_api module conforming with the draft specification of the Array API Standard. This improves the interoperability with other array libraries such as NumPy or PyTorch, and allows code written in terms of this API to be executed in CuPy.

4. Enhanced JIT capabilities
The provided transpiler from Python to CUDA now supports several Python built-in constructions and CUDA atomic primitives. These improvements now allow writing highly performant kernels with more expressiveness.

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