Radeon ROCm 1.9.2 Released - Brings SDMA/RDMA Support For Vega 20, HIP/HCC Improvements

While we know ROCm 2.0 is coming out before year's end and that will have many improvements like complete OpenCL 2.0 support, ROCm 1.9.2 is out today as the latest stable release for this Radeon Open eCosystem stack.

ROCm 1.9.2 brings some notable changes for just being a point release ahead of the big ROCm 2.0 milestone. Vega 20 remains one of the big areas for AMD's driver/software developers for what will begin shipping next year as the Radeon Instinct MI50 / MI60 accelerators.

With ROCm 1.9.2 there is now SDMA/RDMA MPI support for this 7nm Vega. This includes support for ROCnRDMA on Mellanox InfiniBand and support for GPU's data interaction in different nodes via MPI with HIP and OpenCL applications. For Vega 20 there are also async 2D memory copy optimizations.

ROCm 1.9.2 also brings improved linker performance for the HIP/HCC code for getting CUDA code-bases running on AMD Radeon GPUs. It also now disables ThinLTO when early finalization is used. The loader has also been updated to support LLVM Code Object v3.

Rounding out today's update are improvements to the ROCProfiler for HSA run-time changes, fixes, and other additions.

More details on the ROCm 1.9.2 changes via the project's GitHub.