News AMD to disclose roadmaps in mid-November — The future of Zen 6, RDNA, CDNA, and UDNA expected

AMD going all in on Coal to power AI.

383e3e2034c7c0ef02594d84f89fe5d9cb249497.gifv


Nah, I'm just teasing.... We can trust the big corporations to do the right thing to power AI... 😏
 
It's pre-news!

So much has been leaked already that I'd like to hear about something like XDNA3. Or that Sonoma Valley that never appeared.

I thought the future was clear they would move to UDNA for all GPU accelerators?
RDNA and CDNA should be merged into UDNA, but it isn't completely clear when that will happen. Looks like Anton is leaning on the existence of "CDNA5" to say we'll get "RDNA5" and then "UDNA6" (or "UDNA1") following that.
 
  • Like
Reactions: Makaveli
Thought it’s planned for next generation of consumer GPUs or are you saying that you don’t know when they will release the next after RDNA4? RDNA4 is the last generation before UDNA combine? I could be wrong but that’s what I thought AMD said?
 
Thought it’s planned for next generation of consumer GPUs or are you saying that you don’t know when they will release the next after RDNA4? RDNA4 is the last generation before UDNA combine? I could be wrong but that’s what I thought AMD said?
To me it sounds like RDNA 5 was already being worked on and UDNA merge will happen on the gen after that.
 
Thought it’s planned for next generation of consumer GPUs or are you saying that you don’t know when they will release the next after RDNA4? RDNA4 is the last generation before UDNA combine? I could be wrong but that’s what I thought AMD said?
It isn't completely confirmed. If you go back to the interview when UDNA was announced, this was said:
So, going forward, we’re thinking about not just RDNA 5, RDNA 6, RDNA 7, but UDNA 6 and UDNA 7.
Which seems to indicate RDNA5 and then UDNA6 taking over. But he also said:
PA: So, this merging back together, how long will that take? How many more product generations before we see that?

JH: We haven’t disclosed that yet. It’s a strategy. Strategy is very important to me. I think it’s the right strategy. We’ve got to make sure we’re doing the right thing.
So we don't know 100% when it will happen. Maybe that will be announced in November.
 
Thought it’s planned for next generation of consumer GPUs or are you saying that you don’t know when they will release the next after RDNA4? RDNA4 is the last generation before UDNA combine? I could be wrong but that’s what I thought AMD said?
Nah, there will be some milking of evolutionary current tech in the form of RDNA5 and CDNA5.... My guess is RDNA5 is a mild evolution of RDNA4 to provide something that tops out to a Radeon R9-X095XTX (made up name is mine, X=10), a class leading discrete graphics card. UDNA1 (or 6) is being punted a bit... I mean... think how long AMD milked GCN before moving to RDNA ?
 
Nah, there will be some milking of evolutionary current tech in the form of RDNA5 and CDNA5.... My guess is RDNA5 is a mild evolution of RDNA4 to provide something that tops out to a Radeon R9-X095XTX (made up name is mine, X=10), a class leading discrete graphics card. UDNA1 (or 6) is being punted a bit... I mean... think how long AMD milked GCN before moving to RDNA ?
Yeah but times are different they are that much farther behind Nvidia. Not that UDNA solves that problem as there CDNA cards are also behind Nvidia, but I understand UDNA should be their first top tier cards since rdna3 and I can’t imagine they’ll wait another 2 years to be competitive at the top end of the spectrum?
 
Nah, there will be some milking of evolutionary current tech in the form of RDNA5 and CDNA5.... My guess is RDNA5 is a mild evolution of RDNA4 to provide something that tops out to a Radeon R9-X095XTX (made up name is mine, X=10), a class leading discrete graphics card.
RDNA5 could be +10-20% raster, +100% raytracing per compute unit, but with more compute units and a move to 3GB GDDR7 modules for higher capacities.
The news is that in 2 months there will be some news... Countdown to the in-depth analysis of the future roadmap to be part of Tom's Premium... 😉
From what I can tell, anyone can comment in the forum threads for premium articles, so we can give our own hot takes no problem. Just need to find the thread.
 
RDNA5 could be +10-20% raster, +100% raytracing per compute unit, but with more compute units and a move to 3GB GDDR7 modules for higher capacities.

From what I can tell, anyone can comment in the forum threads for premium articles, so we can give our own hot takes no problem. Just need to find the thread.
That's what I've seen touted. RDNA5 should be a massive step up in RTing and there will be a top tier as was always stated even back when top tier RDNA4 was cancelled. AMD had separate teams working on RDNA5 even while RDNA4 was under development. Top tier RDNA4 was cancelled for a number of reasons, but one was to ensure RDNA5 would not be delayed in trying to get a 9090XTX class card working as they wanted.
 
  • Like
Reactions: usertests
It isn't completely confirmed. If you go back to the interview when UDNA was announced, this was said:

Which seems to indicate RDNA5 and then UDNA6 taking over. But he also said:

So we don't know 100% when it will happen. Maybe that will be announced in November.
Thanks for explaining this so I didn't have to. I was about to cause I went to the originally article too but was feeling lazy. lol
 
  • Like
Reactions: usertests
After reading a Ryzen 9 9955X6D post about Moore's Law is Dead leaks on ElChapuzazInformatico, a Spanish IT site that mostly regurgitates news from other sources without giving credit or naming sources, but sometimes comes up with juicy bits anyway...

I've come up with an idea that finally resolves my many questions about the lack of economy for the Strix Halo CCDs:

I believe the 9955X6D and potentially some other Zen 5.5 „refresh“ products might actually be based on the Strix Halo CCDs and using a new stripped down, potentially even iGPU less IOD, which then also leaves the door open to better RAM speeds with a new memory controller (Strix Halo derived).

Reason: Strix Halo CCDs using a „sea of wires“ interconnect for Strix Halo-only are economic nonsense for Strix Halo alone, it simply doesn’t have the potential for scale. But with Strix Halo out the organic die interconnect that may replace the current die carrier Inifinity Fabric also for Zen 6/7 is proven and may offer much better die-to-die latencies, almost monolithic quality, important for desktop gamers, but not suitable for EYPC *).

Thus by only swapping out the IOD for an alternative that replaces memory lanes with PCIe lanes, and the iGPU with either nothing (or NPU or µ-GPU) they have access to a much improved gamer/desktop Strix Halo CCD they are already producing, to make new „Halo Zen 5.5“ parts.

It fits at least partially giving up the increasingly tight EPYC/desktop shared parts strategy in the longer term, or it may signal also a shift in how future EPYCs might be packaged, but with AMD increasing market share and scale (and Intel unable to catch them), they may depend less on that shared strategy anyway, and they wouldn’t want it to become a liability.

How exactly that will play out between EPYC, Thredripper, MegaAPUs, Desktop and mobile remains to be seen, but AMD has consistently shown that modularity can take many forms and that they play it smart.

This move at least for me shows a way around what I thought was a rather atypical stupid move by AMD for what seems a bespoke Strix Halo design and now may be the base for a gamer desktop revolutionary "refresh".

Bonus feature: now if, how and where the Strix Halo IOD might also see some re-use outside just the current Strix Halo, is another interesting bit to consider.
 
Last edited:

TRENDING THREADS