After several ROCm 6.3 point releases, AMD today rolled out ROCm 6.4 as the next update to their open-source GPU/accelerator compute stack and ahead of their big Advancing AI event in June where they will talk about future ROCm work.
ROCm 6.4 is out today with a number of changes over the ROCm 6.3 series. With this new version they are not acknowledging any official support still for AMD RDNA4 / GFX12 GPUs. Unofficially it seems to work but still not part of their official support matrix and I still haven’t received any communication on how they intend to officially position the Radeon RX 9000 series / RDNA4 ROCm support or on any official support for the likes of Strix Halo. But they are now officially supporting the Radeon PRO W7800 48GB graphics card with ROCm 6.4.
Some of the other ROCm 6.4 changes include:
– Compatibility between the ROCm user-space software and the AMDKFD kernel mode drivers has been improved to allow for better compatibility across newer/older kernel versions. AMD has expanded their internal testing to allow for more user/kernel combinations.
– PyTorch 2.5 and PyTorch 2.6 support added.
– The Megatron-LM Framework for ROCm has added support for new fused kernels: Fused Attention (QKV), Fused Layer Norm, and Fused ROPE.
– VP9 support is added to rocDecode and rocPyDecode. Bitstream reader support is also added for rocDecode.
– New modules for the ROCm Data Center Tool.
– Official support for Oracle Linux 9 as Oracle’s RHEL9 derivative.
– Official support for the Radeon PRO W7800 48GB GPU.
More details on AMD ROCm 6.4.0 via rocm.docs.amd.com.
