• Competitor rules

    Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.

*** The AMD UDNA Rumour Mill *** (next generation architecture)

Grim5

Grim5

Grim5

Soldato
Joined
6 Feb 2019
Posts
20,046
Let the rumours begin

What is known so far is that UDNA is a new hybrid architecture from AMD that replaces both RDNA and CDNA with a unified single architecture to service both gamers as well as AI and servers.

It is so far rumoured that UDNA will be built on TSMC N3E process node. It's not yet known if the GPUs will be monolithic or chiplets. Rumoured release is for 2026
 
Last edited:
Yes will be keeping an eye on this one the next few months. I suspect the new consoles will be using some aspects of this so hopefully good leaks by next year. 9070xt on the way so we'll keep me very happy until we get the big boys udna cards in 2027 ?
 
If it truly is a one-size-fits-all, maybe AMD will return to form with FP64 performance that isn't awful (GCN/Vega10 was 1:16, RDNA is 1:32 which is half as good, Vega20 remains king of consumer at 1:8 but that was an HPC card in consumer clothing that won't be repeated again). HPC needs good FP64 AI doesn't, consumer they will at least reduce the ratio so as not to cannibalise the server cards. I hope they don't reduce the ratio by much, going back to 1:16 at worst would be what I'd like to see. More likely (unfortunately) IMO is that they will keep cdna going for bespoke hpc that needs good fp64, and absolutely trash FP64 possibly even to the extent that nvidia does (1:64) with udna to gain whatever fraction of a percent die utilisation elsewhere and set back progress in what I care about for yet another 5 years.
 
I doubt they do chiplets. UDNA 1 will be setting up the next console chip which will 100% be monolithic so I don't think they'll go in a radically different direction with the PC GPUs, especially with all the driver work they'll have to do in order to set a new foundation for the 10 years of next-gen support. Plus I doubt they'll try to compete with a 6090 (which would be the main reason for chiplets), that will be on both a much improved architecture and a node shrink, when they can't even square up to a 5080 that's basically half the chip of gb202 and on the same node and a barely changed arch. Maybe UDNA 2 will see them try chiplets again but I wouldn't get my hopes up for much given their track record with RDNA.
 
Grim5 said:
Let the rumours begin

What is known so far is that UDNA is a new hybrid architecture from AMD that replaces both RDNA and CDNA with a unified single architecture to service both gamers as well as AI and servers.

It is so far rumoured that UDNA will be built on TSMC N3E process node. It's not yet known if the GPUs will be monolithic or chiplets. Rumoured release is for 2026
Click to expand...
 
Poneros said:
I doubt they do chiplets. UDNA 1 will be setting up the next console chip which will 100% be monolithic so I don't think they'll go in a radically different direction with the PC GPUs, especially with all the driver work they'll have to do in order to set a new foundation for the 10 years of next-gen support. Plus I doubt they'll try to compete with a 6090 (which would be the main reason for chiplets), that will be on both a much improved architecture and a node shrink, when they can't even square up to a 5080 that's basically half the chip of gb202 and on the same node and a barely changed arch. Maybe UDNA 2 will see them try chiplets again but I wouldn't get my hopes up for much given their track record with RDNA.
Click to expand...

That's a faulty assumption. The next console chips will be closest to Strix Halo, which is chiplet based. CDNA3 is chiplet based. They likely just rowed back on RDNA4 using chiplets to keep costs under control and work out the kinks; the midrange target for the 9070XT was likely based on that decision. They undoubtedly had chiplet based RDNA4 in testing internally; but decided they couldn't manufacture it to the right cost.
 
Last edited:
harmattan said:
Are we thinink monolithic or chiplet?
Click to expand...
IMHO.

all monolithic with the exception of the 6090 competitor which will be dual chiplet with an interposer.

the problem with a dual-chiplet RDNA4 consumer card was not cost, it was power draw and heat. It would have had similar characteristics to a 5090 but would have been slower.
 
Besty said:
IMHO.

all monolithic with the exception of the 6090 competitor which will be dual chiplet with an interposer.

the problem with a dual-chiplet RDNA4 consumer card was not cost, it was power draw and heat. It would have had similar characteristics to a 5090 but would have been slower.
Click to expand...
Yeah I can imagine them thinking that 2 * 64 CU chiplets at 250W each + 100W for the interposer and such would have been around 600W for a single GPU and coming to the conclusion that no gamer in their right mind would buy a card that needed that much power.
 
If they do do chiplets I'd expect them to do something along the lines of a chiplet that's as powerful as a 9070, so most consumer is catered for by a single chiplet with cheaper packaging. Maybe even doing a zen1 and have the memory controller etc on the chiplet, so for all intents and purposes anything not high end consumer or server is simply monolithic. One way they might be able to do this and still have reasonable chiplet sizes is to do the stacking thing but up to 11, meaning a chiplet contains no L3 and it fully relies on stacking for L3. Stacking higher than 1 is also possible, that could be another way to cater to various segments with the same chiplet.

Zen1-style memory access has its drawbacks, but they're less apparent for server needs who can often get away or even naturally have the workload split into multiple concurrent jobs that can each be pinned to a particular chiplet and set of memory. For AI I have no idea. For gaming scaling is probably tricky and dubious but not impossible. If you can get a 40% performance boost from a second 9070 chiplet bolted on then it's enough for AMD to be in another perf tier for gaming they're not currently in. Even just crossfire with some optimisations that come from being in the same package should be able to get there at least. Whether the lift required to make chiplets work for gaming is worth it is unclear, but it definitely makes sense for server.
 
Xelene said:
That's a faulty assumption.
Click to expand...
It's an assumption, not faulty, unless you're from the future.

Xelene said:
The next console chips will be closest to Strix Halo
Click to expand...
Says who, and based on what? Please don't tell me because it's also an APU and has a unified memory structure.

Xelene said:
CDNA3 is chiplet based.
Click to expand...
And irrelevant to consumer graphics, and with entirely different economics behind it. Even with them "unifying" RDNA + CDNA, it will be of little relevance, just like it is for Nvidia consumer vs DC/pro cards, these things are just too specialised today.

Xelene said:
They likely just rowed back on RDNA4 using chiplets to keep costs under control
Click to expand...
Agreed - so for a new gen on a new (much more expensive) node why would they not do the same?

Xelene said:
and work out the kinks;
Click to expand...
And how do you know they can? If anything their repeated failure to do so should put you in the camp of 'less likely' rather than 'more likely', else it's just wishful thinking.
 
They have already said the problem with chiplets on consumer graphics cards is bandwidth. Both chips have to be treated as one unit so the communication bandwidth between both chips needs to be orders of magnitude higher than anything available at the moment.
 
eeii said:
They have already said the problem with chiplets on consumer graphics cards is bandwidth. Both chips have to be treated as one unit so the communication bandwidth between both chips needs to be orders of magnitude higher than anything available at the moment.
Click to expand...

Some scientists have managed to produce chiplets connections that can provide ample bandwidth; the promising ones I've seen is using fibre optic wires instead of copper. However I don't know how difficult it is to produce on scale or how expensive it is but it is technology that will come in future
 
Grim5 said:
Some scientists have managed to produce chiplets connections that can provide ample bandwidth; the promising ones I've seen is using fibre optic wires instead of copper. However I don't know how difficult it is to produce on scale or how expensive it is but it is technology that will come in future
Click to expand...
The QSFP28 is eye watering expensive. We have that as network back bone at work. Let’s hope they solve the problem soon.
 
You must log in or register to reply here.
Back
Top Bottom