Building off the ROCm 6.3 release from earlier this month, there’s been a Friday night drop of ROCm 6.3.1 with some rather exciting end-of-year improvements.
ROCm 6.3.1 introduces initial support for the new AMD Instinct MI325X accelerators! This evolution of the MI300X is now supported with the ROCm 6.3.1 release. The AMD Instinct MI325X was announced in October with 256GB of HBM3E and various evolutionary improvements over the MI300X. Production shipments of the MI325X began this quarter.
ROCm 6.3.1 introduces per-queue resiliency support for Instinct MI300 accelerators to better protect against misbehaving applications on MI300 class hardware to ensure that other workloads are better protected/unaffected in their separate queues.
ROCm 6.3.1 also introduces the ROCm Runfile Installer as a new easy installer for setting up the ROCm software enviornment without relying on a native Linux package management system like Debian/APT packages. This Runfile Installer also works regardless of internet connection state. But for this initial rollout only Ubuntu 22.04 is initially supported. Hopefully this ROCm Runfile Installer will end up working well with time across other non-officially supported Linux distributions like outside the confines of the enterprise Linux distributions. This would address a major pain point of ROCm right now of the limited official Linux OS/distribution selection where AMD provides official packages. So fingers crossed for more ROCm Runfile Installer improvements in the new year.
ROCm 6.3.1 also brings documentation improvements and initial support for Debian 12…. It’s long past time AMD support ROCm on Debian 12 that debuted back in mid-2023 with the Linux 6.1 kernel usage. Alas the official support is here now and is another reason hopefully this ROCm Runfile Installer support will hopefully bear fruit in 2025 for expanding Linux distribution coverage.
More details on setting up ROCm 6.3.1 for Linux systems and all of the changes in this new point release via rocm.docs.amd.com.