ROCm is an open platform for high-performance computing on GPUs that embodies the 'UNIX philosophy in terms of choice, minimalism, and modularity of software development in a GPU usage environment' [1]. ROCm supports integration with multiple programming languages for the convenience of both developers using ROCm in their projects and users employing ROCm for personal purposes.
Key changes in ROCm 3.3.0 release:
- Users can now install and use different versions of the toolkit simultaneously (previously, only one version could be installed and used at a time).
- A feature has been added to provide information about the GPU process. Information can now be accessed through APIs and CLI.
- Support for 3D Pooling Layers has been added, allowing AMD Radeon Instinct to run 3D convolutional networks, such as ResNext3D, on GPUs.
- Improvements have been made to the ONNX neural network exchange format. Support has been added for pretrained models in the formats: ONNX, NNEF, and Caffe.
- Several previously used features of the code object manager (Comgr) have been deprecated.
As of today, ROCm officially still does not support AMD APU (integrated AMD GPUs), although they are included in upstream driver versions and in the ROCm OpenCL runtime. Expected support for GPUs based on the Navi architecture was also not included in this release.
Source: linux.org.ru
