ROCm aka Radeon Open Compute stack is several years old, but has evolved a lot lately with new features and capabilities to compete with NVIDIA CUDA and Intel oneAPI.
AMD ROCm is the first open-source software development platform for GPU computing in the HPC / Hyperscale class. AMD ROCm brings the UNIX philosophy, minimalism and modular software development for GPU computing.
AMD ROCm is built for scaling and supports multi-GPU computation in and out of server-node communication via RDMA. AMD ROCm also simplifies the stack when the driver directly incorporates RDMA peer-sync support.
For now, the main objective is based on Vega, given the future Arcturus accelerator and mainly ROCm instrumentation for workstations / SC.
ROCm 3.5.1 comes with the updated AMDKFD Driver Kernel Fusion to fix a memory access error in ROCm 3.3 for multi-GPU settings.
ROCm 3.5.1 also provides API additions for querying the priority of a stream with HIP, support for NCCL 2.7, with send / receive operations, and RCCL updates to provide network proxy profile and support for collection / spreading / all all collective operations.