AMD is rumoured abandoning the future RDNA 5 graphic architecture in favor of 'unified UDNA' tech in an effort to bring AI intelligence to gaming ASAP

AMD hasn't yet released its next-gen RDNA4 gaming GPUs, but there are already rumours that RDNA 5 was cancelled in favor of the "unified UDNA graphics architecture". If the rumours are true, AMD could be looking to add AI capabilities as soon as possible to its gaming GPUs.

AMD's current gaming GPUs are based upon the RDNA 3 architecture. Meanwhile, its enterprise, datacenter, and AI GPUs use the CDNA technology. The idea is that gaming and data center demands are sufficiently different to warrant separate architectures.

That was the story until AMD announced in September that it would combine RDNA and CDNA to create a single "UDNA architecture". AMD claimed that the new approach was easier for developers. However, this implies that software developers are creating both large language models and games at the same time. This is unlikely.

UDNA was expected to arrive on gaming PCs in the first half of next year, regardless of its merits. RDNA 4 will be released in the first months of 2019. Rumours have already circulated about its unofficial successor RDNA 5, which is expected to arrive early next year.

The assumption was that the gaming graphics would move to UDNA following RDNA 5. According to the latest rumours on Chiphell, (via PC Guide), this is not true. zhangzhonghao, a well-known leaker, claims that RDNA 5 will be a flop and AMD is preparing a UDNA family of GPUs that will succeed RDNA 4.

The poster also claims that UDNA will be derived from GCN. Say what? GCN was the AMD graphics architecture which preceded RDNA. It first appeared in Radeon HD 7700 family of GPUs in 2012. GCN didn't die when RDNA was released in 2019. CDNA was born instead. If you can forgive the acronyms, UDNA is an evolution of CDNA which, in turn, has its origins from GCN.

It's important to note that the idea of "unified" graphic architectures is not what it seems. It does not mean that AMD will sell the same chips to AI developers as well as gamers.

There are many components that go into a gaming GPU but not an AI GPU. These include raster and geometries hardware, render outputs, 2D engines and hardware video encoders and decoders.

We're really talking about a shared architecture of the shader ALUs, and the AI-accelerating matrix Cores that Nvidia refers to as Tensor Cores. These are the cores in various NPUs or neural processing units. Even a gaming GPU can have that many cores, especially with shaders so important. But it is not the entire chip.

While a move to a new architecture based on "old" GCN seems to be a retrograde step at first glance, it's difficult to draw any conclusions. The relationship with GCN could be more of a high level philosophical approach rather than carrying over the nuts-and-bolts of the GCN architectural circa 2012.

At this stage, it's impossible to know for sure what AMD's new unified strategy will mean for gamers. The most obvious benefit will be the addition of AI intelligence.

AMD does not use any AI acceleration on its graphics cards at this time to process or enhance the FSR upscaling technology. This is a stark contrast to Nvidia which heavily relies on AI with DLSS, and emphasizes the contribution of Tensor Cores built into its last three generations gaming GPUs.

RDNA 3 GPUs currently do not have matrix-cores. UDNA GPUs will also have matrix cores, as CNDA GPUs already do. UDNA gaming graphics cards will finally enter the AI era. There have been hints that AMD FSR will start to match Nvidia's FSR when it comes down to using dedicated silicon for its upscaling features and frame generation, which means future graphics cards will need to include tech the current ones don't. When you think about it that way, maybe AMD did need to move to UDNA ASAP after all.