CuPy 9.0 Brings AMD GPU Support To This Numpy-Compatible Library

In recent months there has finally been more open-source projects traditionally focused on NVIDIA GPU compute beginning to offer mainline Radeon support using the open-source ROCm stack. Following the recent PyTorch 1.8 with ROCm support, CuPy 9.0 was released last week with that traditionally CUDA focused library now supporting AMD's ROCm stack.

CuPy 9.0 was released last week as the NumPy compatible open-source array library that has been focused on NVIDIA's CUDA interface, hence the CuPy name. With CuPy 9.0 there is support for AMD graphics processors with ROCm both via the official binary packages and Docker images. CuPy 9.0 also brings Windows support, a JIT API for writing kernels in Python, NVIDIA cuSPARSEit support, and NumPy/SciPy compatibility improvements.

CuPy 9.0 also features some performance improvements and improved documentation. The new AMD GPU support has been tested against ROCm 4.0.

More details on the CuPy 9.0 changes via this blog post. The release notes and source downloads for the new CuPy release over on GitHub.