This morning, AMD officially announced the launch of ROCm 5.7, the latest version of their GPU computing software stack. Along with making some significant updates to the software stack itself, particularly regarding improved support for Large Language Models (LLM) and other machine learning toolkits, the company has released a blog post outlining its future hardware development plans for the stack. also published. That means the company will officially support a limited set of his RDNA 3 architecture video cards starting this fall.
AMD’s equivalent of NVIDIA’s CUDA and Intel’s OneAPI software stacks has historically focused on a narrower range of hardware with its own GPU computing software stack. ROCm exists first and foremost to support AMD’s Instinct accelerator series (used in projects such as the Frontier supercomputer), and as a result has limited support for non-Instinct products. Formally, AMD only supports the software stack on workstation-class RNDA 2 architecture card pairs (Radeon Pro W6800 & V620), unofficial support is available for some other RDNA 2 cards and architectures, but in practice it turned out to be controversial as to how reliably it works. I’m here. Therefore, the announcement of new Radeon video card support for ROCm is worth noting, especially when it comes to consumer Radeon cards.
To wrap up the ROCm 5.6 announcement blog post, AMD announced that support for the first RDNA 3 products will be available in the fall. First, the company adds official support for AMD’s top workstation card, the Radeon Pro W7900, and for the first time consumer-grade his Radeon RX 7900 XTX. Both of these pieces are based on the same RDNA 3 GPU (Navi 31) so architecturally they are identical and it would be nice to see AMD eventually adopt it and bring it to consumer Radeon cards. It’s a sign that you should.
Broadly speaking, RDNA 3’s computing core differs significantly from RDNA 2 (and CDNA 2) due to the introduction of dual-issue SIMD execution and the consequent need to extract the ILP from the instruction stream. So adding proper RDNA 3 support to the ROCm stack would be a no-brainer for AMD’s software team, especially if they are also working to support the launch of his MI300 (CDNA 3) accelerator family later this year. it’s not work.
In addition to the first two Navi 31 cards, AMD is also working to support “additional cards and enhancements to be released in the future”. So far, AMD’s official video card support has never extended beyond a single GPU (e.g. Navi 21) within a given generation. So, does this mean AMD will finally expand its range to include more Navi 3x GPUs, or does this mean officially supporting more Navi 31 cards (such as the W7800)? increase. AMD’s statement also seems to suggest that full ROCm feature set support may not be available in the first iteration of his RDNA 3 support, but I read too much into it. There is a possibility that
Meanwhile, although by no means official, AMD’s blog post states that the company is also improving its unofficial support for Radeon products. A number of issues with his ROCm on unsupported GPUs have been fixed in his ROCm 5.6 release, making the software stack routinely usable on a wider range of hardware.
All in all, the Navi 3x’s potential as a compute offering has been underutilized since its launch over half a year ago, with AMD finally readying support for the latest desktop GPU architectures within its compute stack. What has happened is a welcome development. AMD has received some well-deserved criticism over the years for ROCm’s limited support for their underlying GPU products. Therefore, this announcement CEO Dr. Lisa Su comments earlier this month AMD’s commitment to improving ROCm support marks AMD finally making much-needed (and much-anticipated) progress in improving its ROCm product stack.
However, while AMD is preparing to add more hardware support for ROCm, it is also preparing to remove some. Support for products based on AMD’s Vega 20 GPUs such as Instinct MI50 and Radeon Pro VII Sunset is expected to begin later this year. His ROCm support for these products will enter maintenance mode in Q3 with the release of ROCm 5.7. No features or performance optimizations will be added to that hardware branch at that time. Bug fixes and security updates are released for about another year. Ultimately, AMD has officially warned that it plans to fully end support for its hardware after the second quarter of 2024.
Finally, despite some premature rumors, for those who were hoping for full ROCm support on Windows, that didn’t happen with ROCm 5.6. AMD currently has very limited support for Windows in their ROCm toolchain (ROCm is used for AMD back-in in both the Linux and Windows versions of Blender). ROCm’s development logs indicate that they are still working on this issue. However, the software stack does not yet have full Windows support. AMD has generally remained completely silent on the matter, avoiding anything to raise hopes for the release of ROCm-on-Windows. That said, I’d expect Windows to be properly supported at some point in the future, but there’s nothing to indicate that it will happen any time soon. Especially since he has an MI300 on the horizon, AMD seems to be able to catch bigger fish.
