AMD Releases ROCm 5.7.1 With rocBLAS-GEMM-Tune & Bug Fixes

AMD on Friday released ROCm 5.7.1 as their latest tagged release to this open-source GPU compute stack that continues progressing to better take on NVIDIA CUDA and Intel oneAPI.

Following the ROCm 5.7 release from September, the ROCm 5.7.1 is predominantly about delivering various bug fixes to users of this open-source GPU compute solution. Most notable with ROCm 5.7.1 is introducing rocblas-gemm-tune to rocBLAS as a new utility to find the best-performing GEMM kernel for each GEMM problem set. This command-line tool for finding the best performing GEMM kernel can then be used with the ROCBLAS_TENSILE_GEMM_OVERRIDE_PATH for subsequently overriding the default kernel selection.

ROCm 5.7.1 also has several bug fixes to the HIP run-time, a few fixes to the hipSOLVER, and other random fixes throughout.

When it comes to the supported GPUs, officially with the documentation the only supported Radeon GPU remains the Radeon VII. In the Radeon Pro series is the Radeon PRO W6800 / PRO V620 / PRO V11. With the AMD Instinct hardware it's the Instinct MI250X / MI250 / MI210 / MI100 / MI50 and then the MII25 support is on the way to being retired. Sadly there remains no officially advertised RDNA3 support whether it be the Radeon PRO W7000 series or the consumer Radeon RX 7000 series. AMD is working on expanding their supported hardware and should be offering broader RDNA3 GPU support. But at least as of the documentation in ROCm 5.7.1, there's still no support mentioned though some users have reported it working to varying degrees. Hopefully in time for the next major ROCm release the RDNA3 GPU support and the compatibility expectations will be cleared up.

For downloading ROCm 5.7.1 or to learn more about the changes with this point release, visit rocm.docs.amd.com.