New AMD ROCm 5.5 Includes RDNA3 Enhancements, Other Changes

ROCm 5.5, the latest version of AMD’s open-source GPU computing stack and their alternative to NVIDIA’s CUDA and Intel’s oneAPI / Level Zero, was formally released overnight as expected following yesterday’s AMD Git activity. ROCm 5.5 is AMD’s latest version of their open-source GPU compute stack. The most notable change in ROCm 5.5 is the increased support for Radeon GPUs from the RX 7000 series and RDNA3. If we look back at older Docker test builds, we can see that ROCm 5.5 has been successful for owners of Radeon RX 7900 series graphics cards. This is particularly essential since graphics cards based on RDNA3 will soon be available for the Radeon PRO W7800 and W7900 series. Although ROCm 5.5 has been tested to have improved support for RDNA3, the fact that this is not mentioned at all in the recently published version 5.5 release notes is somewhat surprising…

On the other hand, AMD has a tendency to “officially” simply focus on the accelerator and graphics card support for their professional and workstation products. The GPU support matrix for ROCm 5.5 is still somewhat depressing, as it only includes a small number of GFX9, CDNA, and RDNA GPUs. But everything that has been gathered up to this point suggests that ROCm 5.5 ought to be working out better now for Linux users who have Radeon RX 7000 series graphics cards. The ROCm 5.5 release notes do note a ~128k stack size limit now rather than 16k, various HIP compiler changes as well as new APIs, a new Linux file-system, hierarchy, and hipBLAS / hipCUB / hipSOLVER / hipSPARSE / RCCL / rocALUTION / rocBLAS / rocFFT / rocRAND / rocSOLVER / rocSPARSE / rocWMMA / Tensile library updates. Some of the individual libraries now list support for GFX11 (RDNA3), and many of the updates to the libraries are focused on providing performance optimizations, new interfaces, and other improvements.