Aiming at NVIDIA CUDA, AMD will release ROCm 6.1 update: giving full play to the performance of AI accelerator cards

According to news from this site on February 29, after AMD released ROCm 6.0 in December last year, it is non-stop promoting the 6.1 version update to further enhance its competitiveness with NVIDIA CUDA.

AMD has recently submitted a large number of commits on ROCm 6.1 to the GitHub public repository, which shows that AMD is expected to release a new version in the near future.

According to the commits currently submitted, ROCm 6.1 will be upgraded to the MIPOpen 3.1 version, introducing an AI-based parameter prediction model for the conv_hip_igemm_group_fwd_xdlops solver (solver) , fixed many bugs in the existing version and brought a new solver.

ROCm 6.1 also focuses on improving MIGraphX, adding support for FP8 precision, more operators, Whisper / Llama-2 / Stable Diffusion 2.1 document examples, new ONNX examples, and GEMM's BLAS automatic adjustment in version 2.9 , and initial code to run MIGraphX ​​on Microsoft Windows.

ROCm 6.1 also introduces a new "hipother" repository for HIP, which holds HIP backend implementations for non-AMD platforms. Note from this site: HIP is AMD's C runtime API and kernel language for portable AMD/NVIDIA GPU applications, and the HIPIFY tool helps convert CUDA sources to HIP.

The above is the detailed content of Aiming at NVIDIA CUDA, AMD will release ROCm 6.1 update: giving full play to the performance of AI accelerator cards. For more information, please follow other related articles on the PHP Chinese website!