Partial RDNA 3 Video Card Assist Coming to Future ROCm Releases
AMD this morning is formally saying the launch of the most recent model of its GPU compute software program stack, ROCm 5.7. Together with making a number of vital updates to the software program stack itself – significantly round enhancing help for giant language fashions (LLMs) and different machine studying toolkits – the corporate has additionally printed a weblog submit outlining the long run {hardware} growth plans for the stack. Briefly, the corporate shall be bringing official help to a restricted set of RDNA 3 structure video playing cards beginning this fall.
AMD’s counterpart to NVIDIA’s CUDA and Intel’s OneAPI software program stacks, AMD has traditionally pursued a narrower {hardware} focus with their very own GPU compute software program stack. ROCm exists initially to help AMD’s Intuition line of accelerators (utilized in initiatives such because the Frontier supercomputer), and consequently, help for non-Intuition merchandise has been restricted. Formally, AMD only supports the software stack on a pair of workstation-class RNDA 2 architecture cards (Radeon Professional W6800 & V620), whereas unofficial help is on the market for another RDNA 2 playing cards and architectures – although in apply this has confirmed to be a combined bag as to how reliably it really works. Consequently, any announcement of recent Radeon video card help for ROCm is notable, particularly when it entails a client Radeon card.
Closing out their ROCm 5.6 announcement weblog submit, AMD is saying that help for the primary RDNA 3 merchandise shall be arriving within the fall. Kicking issues off, the corporate shall be including official help for the Radeon Pro W7900 – AMD’s high workstation card – and, for the primary time, the consumer-grade Radeon RX 7900 XTX. Each of those components are primarily based on the identical RDNA 3 GPU (Navi 31), so architecturally they’re equivalent, and it’s a welcome signal to see AMD lastly embracing that and bringing a client Radeon card into the fold.
Broadly talking, RDNA 3’s compute core differs considerably from RDNA 2 (and CDNA 2) because of the introduction of twin challenge SIMD execution, and the ensuing have to extract ILP from an instruction stream. So the addition of correct RDNA 3 help to the ROCm stack is just not a small enterprise for AMD’s software program crew, particularly when they’re additionally working to help the launch of the MI300 (CDNA 3) accelerator family later this yr.
Together with the primary two Navi 31 playing cards, AMD can be committing to bringing help for “further playing cards and expanded capabilities to be launched over time.” So far, AMD’s official video card help has by no means prolonged past a single GPU inside a given era (e.g. Navi 21), so it is going to be attention-grabbing to see whether or not this implies AMD is lastly increasing their breadth to incorporate extra Navi 3x GPUs, or if this simply means formally supporting extra Navi 31 playing cards (e.g. W7800). AMD’s assertion additionally appears to indicate that help for the complete ROCm function set is probably not accessible within the first iteration of RDNA 3 help, however I could also be studying an excessive amount of into that.
In the meantime, although it’s not by any means official, AMD’s weblog submit additionally notes that the corporate is enhancing on their unofficial help for Radeon product, as nicely. Quite a few points with ROCm on unsupported GPUs have been mounted within the ROCm 5.6 launch, which ought to make the software program stack extra usable on a day-to-day foundation on a wider vary of {hardware}.
General, it is a welcome growth to see that AMD is lastly lining up help for his or her newest desktop GPU structure inside their compute stack, as Navi 3x’s potential as a compute product has remained lower than totally tapped because it launched over half a yr in the past. AMD has taken some not-undeserved flak through the years for ROCm’s restricted help for his or her bread-and-butter GPU merchandise, so this announcement, together with CEO Dr. Lisa Su’s comments earlier this month that AMD is working to enhance their ROCm help, point out that AMD is lastly making some a lot wanted (and vastly awaited) progress with enhancing the ROCm product stack.
Although as AMD prepares so as to add additional {hardware} help for ROCm, they’re additionally getting ready to take some away, as nicely. Assist for merchandise primarily based on AMD’s Vega 20 GPU, such because the Intuition MI50 and Radeon Professional VII, is set to begin sunsetting later this year. ROCm help for these merchandise shall be getting into upkeep mode in Q3, with the discharge of ROCm 5.7, at which era no additional options or efficiency optimizations shall be added for that department of {hardware}. Bug fixes and safety updates will nonetheless be launched for roughly one other yr. Finally, AMD is giving a proper heads up that they’re seeking to drop help for that {hardware} totally after Q2 of 2024.
Lastly, for anybody who hoped to see full Home windows help for ROCm, regardless of some untimely rumors, that has not occurred with ROCm 5.6. At the moment, AMD has a really restricted diploma of Home windows help within the ROCm toolchain (ROCm is used for the AMD backin in each Linux and Home windows editions of Blender) and ROCm growth logs point out that they’re persevering with to work on the matter; however full Home windows help stays elusive for the software program stack. AMD has remained fairly mum on the matter general, with the corporate avoiding doing something that might set any expectations for a ROCm-on-Home windows launch. That stated, I do nonetheless count on to see correct Home windows help sooner or later sooner or later, however there’s nothing to point it is occurring any time quickly. Particularly with MI300 on the horizon, AMD would appear to have greater fish to fry.
