UDNA is the next thing coming after the RDNA era
The next generation AMD GPUs with RDNA 4 architecture should be coming soon. The company has now confirmed the rumors that high-end models will not be released in this generation and it will cover only part of the performance and price range. But the company also discussed their long-term roadmaps for the Radeon and Instinct GPU architectures. It seems we could be in for a change as significant as the transition from GCN to RDNA.
Tom’s Hardware had the opportunity to speak with Jack Huynh, Senior Vice President and General Manager of Computing and Graphics at AMD, in Germany during IFA 2024. In this interview, it was mentioned that AMD is “deprioritizing” high-end graphics cards in favor of mainstream ones.
But this statement should probably be taken in the context that AMD is basically reflecting certain pre-existing realities it has been hit with and trying to frame them in as positive light as possible. This is probably not a permanent change in strategy, and quite possibly not so much a change in strategy as a result of internal product development. According to unofficial information, RDNA was originally supposed to have 4 high-end models, but the bigger GPU designs were canceled. And it is possible, or rather likely, that their cancellation was not something desired, but rather a step forced due to technical hurdles or problems with their complex chiplet design. That is, something that happened not because AMD decided to simply abandon the higher-end parts of the market. This is something to bear in mind when pondering these statements.
New UDNA architecture
But what’s more interesting about the Tom’s Hardware interview is the second bit of information, which is entirely new. According to Huynh, AMD is planning to unify its current line of compute GPU accelerators, such as the Instinct MI300X, and Radeon gaming graphics cards. These currently use separate architectures that don’t share the same foundation. While the Radeons have architectures coming from the RDNA line (which started in 2019 with Radeon RX 5000 graphics cards), the GPUs used for HPC markets and AI have architectures called CDNA, CDNA 2, CDNA 3, and so on, which are specifically tailored for these areas (C stands for Compute).
In the future, however, both of these separate lines are to be replaced by a single one, which AMD is now referring to as UDNA (U could probably stand for “Unified” or “Universal”). Both Radeon graphics cards (and integrated GPUs) and Instinct AI accelerators will be based on this architecture in the future.
It is not entirely clear whether this means a transition away from the current RDNA architecture (or its different generations) and whether this UDNA architecture will be something completely new. It is possible that it will be an evolutionary successor to RDNA, but with its scope extended to the compute domain and getting the new name UDNA. It would basically be a replacement of the current CDNA line with an architecture based on the RDNA foundation.
CDNA is actually the older technology in AMD’s portfolio genetically, although it was announced only after RDNA in 2020 and saw its first appearance in the Instinct MI100 accelerator the same year. However, it is an evolutionary successor to the previous GCN, Polaris and Vega architectures, which is evident in the clearly GCN-based concept of the architecture, which uses Wave64 SIMD operations on SIMD16 units (i.e. processed in multiple passes). In contrast, RDNA has primarily switched to narrower Wave32 operations over SIMD32 units, allowing them to be processed in a single pass. RDNA was apparently not as well prepared for compute use initially, so AMD chose the strategy of continuing to develop the Vega architecture to ramp up Instinct compute GPUs.
Thus, the re-unification could also reflect the maturation and scaling up of the currently “gaming-only” RDNA GPU architecture. It should also have the advantage that developers will be able to use common consumer hardware, including laptop GPUs, when developing GPU applications before deploying code in datacenters and the cloud on Instinct’s large accelerators.
Combining both lines under one architecture could have performance benefits despite losing some of the benefits given by the current specialization. This is because a shared architecture will be able to get more resources for development and optimization – improvements developed by engineers on one side will translate to the other.
This unification probably doesn’t mean that there will be no difference between the compute version and the gaming version of UDNA. Similar to Nvidia’s approach for example, ray tracing acceleration units or other fixed-function graphics blocks may be removed when UDNA is implemented in an AI accelerator. For gaming graphics cards, support for double precision computation (FP64) may be eliminated. However, this will be achieved by modifying the same underlying architecture and both variants will be able to be served by the same software “stack”.
When UDNA will be released is unknown
UDNA is possibly a hugely important news, but the real impact of it may not come until a relatively distant future. AMD has not indicated when the unification of GPUs on the UDNA architecture is expected to occur. However, at one point in the interview, hypothetical sequences of RDNA 4, RDNA 5 and RDNA 6, and against them UDNA 5 and UDNA 6 are laid out side by side. Unless these numbers are strictly for illustrative purposes (which can’t be ruled out), this could imply the transition could happen right after RDNA 4, that is, as part of the architecture we’ve been expecting to come out as RDNA 5. That architecture is rumored to be significantly innovated, so it could be a good starting point for this change in strategy. But it doesn’t rule out that this transition could occur later, for example only a generation down the road with RDNA 6 / UDNA 6. AMD itself may not be sure yet, or may not want to reveal the actual plan just yet.
It has to be said that this concept with UDNA would be closer to how Nvidia works with its architectures. The company is going to introduce the Blackwell architecture in the next generation, spanning both compute (AI) GPUs and GeForce gaming graphics. The company does split architectures in some generations (and even seems to do so in a deliberate regular pattern every other generation – see the Volta and Turing, Lovelace and Hopper generation). But that approach is not really that different from the unified generations (Pascal, Ampere and Blackwell architectures). In fact, there are differences between the compute and gaming versions of those unified architectures too too, it’s just not reflected by naming.
But the important thing here is that even the split architectures are conceptually similar and related (sharing the same predecessors and concepts), which brings compatibility advantages, whereas with AMD there is now significantly more divergence between the RDNA and CDNA lines.
However, even AMD used to have its whole GPU portfolio based on one architecture in the past, before the RDNA came along. To a large extent, UDNA will actually be a return to the previous mode of operation.
Source: Tom’s Hardware
Jan Olšan, editor @ Cnews.cz
⠀
