AMD is actively promoting the ROCm 6.1 version update to strengthen its competitiveness with NVIDIA CUDA. Since the release of ROCm 6.0 in December last year, AMD has submitted a large number of ROCm 6.1 commitments to the GitHub public warehouse, indicating the upcoming release of a new version.
The latest commits show that ROCm 6.1 will be upgraded to MIPOpen 3.1, introducing an AI based parameter prediction model, fixing multiple bugs in the existing version, and introducing a new solver (conv_hip_igimm_group_fwdxdlops).
In addition, ROCm 6.1 focuses on improving MIGraphX, adding support for FP8 precision, more operators, Whisper/Lama-2/Table Diffusion 2.1 documentation examples, new ONNX examples, and GEMM’s BLAS automatic adjustment on top of version 2.9. The initial code for running MIGraphX on Microsoft Windows has also received attention.
ROCm 6.1 has launched a new HIP resource library called “hip ther”, aimed at storing HIP backend implementations for non AMD platforms. It is worth noting that HIP is AMD’s C++runtime API and kernel language, which can be used for portable AMD/NVIDIA GPU applications. The HIPIFY tool helps convert CUDA source code to HIP format. Through these updates, AMD continuously optimizes its software ecosystem to provide users with a better GPU computing experience.
