AMD ROCm 6.0 Now Available To Download With MI300 Support, PyTorch FP8 & More AI
ROCm 6.0 was released a short time ago on GitHub. AMD's documentation portal has also been now been updated for reflecting the changes with ROCm 6.0.
ROCm 6.8 introduces support for the AMD Instinct MI300A/MIX300X accelerators but with the exception that the support isn't currently in place for Red Hat Enterprise Linux 9. It's rather surprising that the MI300X/MI300A aren't yet working on RHEL9 (EL9 Linux distros) but presumably they will get that working soon.
ROCm 6.0 also brings support for more AI libraries including DeepSpeed, ONNX-RT, Jax, CuPy and others. There is also now FP8 support in PyTorch, hipSPARSELt to provide sparse matrix core technique support on RDNA/CDNA GPUs for speeding up AI workloads and inferencing, and improved documentation. RHEL 9.3 and RHEL 8.9 point releases are also now supported with ROCm 6.0.
For easing the deployment of ROCm 6.0, a new rocm meta package is also now present in the ROCm repositories for easily installing all the ROCm packages. ROCm 6.0 additionally now puts the AMD instinct MI50 on end-of-support notice.
There are also many improvements/optimizations and fixes to the numerous contained libraries within ROCm 6.0. When it comes to Linux-supported GPUs by ROCm 6.0, the only Radeon cards listed are the Radeon RX 7900 XTX and Radeon VII... A bit odd when a month ago AMD was talking up ROCm for the RX 7900 XT. We also don't have any better information to communicate right now on when the rest of the AMD Radeon RX 7000 series line-up may see official ROCm support.
For Radeon Pro hardware support with ROCm 6.0 on Linux, the Pro W7900, Pro W6800, Pro V620, and Pro VII are the listed cards. Lastly the Instinct MI300, MI250, MI210, and MI100 series are fully supported by ROCm 6.0 while MI50 is reaching end-of-support.
More details on the many ROCm 6.0 software changes via rocm.docs.amd.com.
32 Comments
