UDNA aka GFX13: Entry in the kernel code points to AMD’s next GPU generation – and we know almost nothing

Alright, there it is again – that infamous moment when a single line of code raises more questions than answers. GFX13 is the new fun term AMD is apparently using internally for its upcoming GPU architecture. The source? An entry in a kernel-related codebase. And as always, that’s enough to spin an entire narrative out of a few bits and bytes. You know the drill.

RDNA was yesterday – now comes… UDNA?

The hottest contender for the name of this ominous new architecture: UDNA. Of course, nobody really knows what the U stands for. “Unified” is the most popular guess, but it could just as easily stand for “Unconfirmed”, “Undecided” or “Unpublished”. The fact is that the current RDNA family has not even completed its cycle with RDNA 4, and the next generation is already being worked on in the code. According to the ever-active leaker Kepler_L2 – a source that has been more or less accurate in the past – GFX13 is the internal designation of UDNA, possibly also RDNA 5. AMD itself remains silent on this, which is quite common in the industry. So there is nothing official – just lines of code, speculation and a fair amount of guesswork.

“Wavegroup”: more buzzword than feature?

The discovery of a new feature called “Wavegroup” is exciting. At first glance, this sounds like a musical self-help group, but it is more likely an extension of the previous “Wavefront” concept. The latter is known at AMD to describe a group of threads (typically 32 or 64) that march together through the shader pipeline. The “wavegroup” could represent a kind of meta-level above this – i.e. several wavefronts that are managed or coordinated together. Whether this is a new scheduling approach, a load-balancing mechanism or just another marketing buzzword without substance – who knows? In any case, the kernel code reveals very little.

TSMC N3E: The next step in the shrinking cure

According to the rumors, UDNA (or whatever it is called) will rely on TSMC’s N3E manufacturing process – i.e. 3 nm technology with improved parameters for high-performance designs. This is understandable, as the rest of the industry (Apple, Intel, Nvidia) are also tightening their belts in terms of production sizes. It is interesting to note that the new architecture is apparently not only intended for dedicated graphics cards, but could also appear in APUs – i.e. in the integrated GPUs of future Zen 6 processors. At least in theory. According to the current roadmaps, a large part of the Zen 6 family will still work with RDNA 3.5 – see Strix Point, Strix Halo, Krackan Point. So far, there is no trace of UDNA-based iGPUs.

https://x.com/Kepler_L2/status/1922109005389840884

RDNA 4: Mid-range instead of a demonstration of power

While speculation about GFX13 is increasing, RDNA 4 remains the concrete product for the time being. And here, AMD has deliberately left out the high-end segment. Instead of a new RX 7900 XTX, there is currently more solid mid-range fare – such as the RX 9700 XT and 9600 XT. The RX 9060 XT is set to follow at Computex, if the leaks are to be trusted. Apparently AMD is saving the really big buns for UDNA. The launch period? 2026 at the earliest – provided no new delays or course changes intervene. Until then, RDNA 4 will remain the last chapter of the classic GPU architecture before the major standardization is possibly due.

A code name does not make an architecture

Anyone who believes that the discovery of “GFX13” is the big breakthrough in AMD’s GPU future is probably mistaken. The entry in the kernel is a piece of the mosaic – no more, no less. It is an indication that AMD has long been working internally on the next generation. But what form it will take, how it will position itself in terms of performance, what the “Wavegroup” will really bring in the end – all this remains in the dark at the moment. And as is so often the case, if you expect too much too soon, you risk being disappointed in the end. Until AMD itself puts its cards on the table, we will have to observe, interpret and (hopefully) soberly analyze. The rest is speculation – and there is already enough of that.

Source: Kepler_L2