ROCm is an open platform for high-performance computing on the GPU, which carries "the philosophy of UNIX in terms of choice, minimalism and modularity of software development in the GPU environment" [1]. ROCm supports the integration of many programming languages ββfor the convenience of both developers using ROCm in their projects and users using ROCm for personal purposes.
Main changes in ROCm 3.3.0 release:
- Users can install and use different versions of the toolkit at the same time (previously only one version was available for installation and use).
- Added the function to provide information about the GPU process. API and CLI can be used to get information.
- Added support for 3D Pooling Layers, which allows you to run convolutional (convolutional) networks, such as ResNext3D, on AMD Radeon Instinct 3D GPUs.
- Improvements have been made to the ONNX neural network exchange format. Added support for pre-trained models in the following formats: ONNX, NNEF and Caffe.
- A number of previously used Code Object Manager (Comgr) functions have been declared unsupported.
To date, ROCm still does not officially support AMD APUs (AMD Integrated GPUs), although they are included in upstream drivers and in the ROCm OpenCL runtime. Expected GPU support on the Navi architecture was also not included in the release.
[1] ROCM documentationSource: linux.org.ru