Discussion RDNA 5 / UDNA (CDNA Next) speculation

[fastandfurious6]

RDNA5 must be a really good architecture

This sentence awakens a sudden clarity in me... third eye wide open 👁️ 👁️ 👁️

 

[dangerman1337]

How the hell do I ship shit much with 30% shrink

Really 288 SMs? Thats 1,53x the SM count on RTX 6000 Pro

I suspect Nvidia will be cutting L2 Cache for their dies in a similar way AMD is; with 36 Gbps GDDR7 there is little need for a lot of L2 cache as with Lovelace & Blackwell. I mean I feel RTX 60 is going to be on Samsung's SF2X (AFAIK on par with TSMC N3 nodes, which is 1.6x or so the logic density over N5 itself) with rumours of Nvidia been talking to them:

[News] Samsung Reportedly in Talks with NVIDIA, Broadcom to Ship Custom HBM4 in 1H26 | TrendForce

Amid rumors of Google moving away from its HBM3E and NVIDIA’s looming test in June, Samsung looks set to turn the tide with its next-gen HBM. Accordin...

www.trendforce.com

AT0 XL is a gaming card. If we're to believe the list that first named the AT0 codename, then we should believe what the list says.

I think the list ain't final, like AMD could do different SKUs with AT0, AT2 etc. I mean RTX 5090 specs was going to be 160 SMs on a 448-bit bus until Nvidia boosted that up to 170 SMs & 512-bit bus.

 

[jpiniero]

I mean I feel RTX 60 is going to be on Samsung's SF2X (AFAIK on par with TSMC N3 nodes, which is 1.6x or so the logic density over N5 itself) with rumours of Nvidia been talking to them:

[News] Samsung Reportedly in Talks with NVIDIA, Broadcom to Ship Custom HBM4 in 1H26 | TrendForce

Amid rumors of Google moving away from its HBM3E and NVIDIA’s looming test in June, Samsung looks set to turn the tide with its next-gen HBM. Accordin...

www.trendforce.com

I think if nVidia does go back to Samsung, it would be for a dual source strategy where the low end parts are fabbed at Samsung.

 

[adroc_thurston]

I think if nVidia does go back to Samsung, it would be for a dual source strategy where the low end parts are fabbed at Samsung.

No one's going back to Samsung since Samsung can't yield.

 

[jpiniero]

No one's going back to Samsung since Samsung can't yield.

Maybe if it's small enough?

 

[gdansk]

Maybe if it's small enough?

When's the last time a small enough GPU was relevant?

I bet even a N44 class product is too big for Samsung to make profitably at a price low enough to entice AMD.

 

[adroc_thurston]

Maybe if it's small enough?

Why bother engaging a different PDK to do a tiny ass irrelevant part?

 

[jpiniero]

Why bother engaging a different PDK to do a tiny ass irrelevant part?

They sell a boat load of those on mobile. But yeah it probally wouldn't be worth it unless it also yielded acceptably.

 

[adroc_thurston]

They sell a boat load of those on mobile.

Just as fine on a TSM node. better, even.

 

[branch_suggestion]

You know, AMD has used Vega and Deneb before, just saying.

 

[ToTTenTranz]

Why bother engaging a different PDK to do a tiny ass irrelevant part?

Did you tell nvidia they did a mistake in engaging with Samsung to make the tiny ass Pascal GPUs while the bigger ones were on TSMC?

 

[adroc_thurston]

Did you tell nvidia they did a mistake in engaging with Samsung to make the tiny ass Pascal GPUs while the bigger ones were on TSMC?

That was way back when design costs were cheaper, and cycles shorter.
Next.

 

[poke01]

If they are truly targeting the RTX6090 with a 154CU 384b 380w part

how likely is this SKU to release?
@Kepler_L2 @adroc_thurston

 

[Kepler_L2]

how likely is this SKU to release?
@Kepler_L2 @adroc_thurston

Ask Lisa

 

[dangerman1337]

On the subject of monolithic Vs chiplet, I think it's a two chiplet deal with a Media & I/O Die that can have functions (i.e. what PCI-E Gen is used because even a 192CU gaming card doesn't even need Gen 6.0)

Ask Lisa

So AT0 does really exist? That said why is there no AT1 documented? Was AT0 & AT2 chosen as examples of what they're developing? Really also what Nvidia may do for their top die.

I can understand reluctance of a 96CU, 256-bit doe if AT2 is 72CUs but I can see the case of a 320-bit, 120CU AT1 where it gets cut down quite a lot. I mean I can see a lineup like this:

1. AT0 - Radeon AI 190 XT (512-bit), 190 (448-bit) & 180 XT (384-bit).
2. AT1 - Radeon AI 180 (320-bit), 170 XT (256-bit) & 170 (224-bit?).
3. AT2 - Radeon AI 160 XT (192-bit) & 160 (160-bit lol).
4. AT3 - Radeon AI 150 XT (128-bit) & 150 (96-bit, yes I know but with 3GB GDDR7 that'd make a compelling, lower power, popular MP gaming card for developing countries).

Just random ass speculation with what I think the specs look like (I think AT2 has 72 CUs, just makes too much sense). I don't think that leaked document is 100% finalized because we don't know the context. But I do think it's very certain that AMD is planning a 512-bit, 192CU die and is pushing ahead of it. They look at the NVIDIA RTX PRO 6000 Blackwell cards being sold and in 2027 or so with 4GB modules AMD & Nvidia can do 128GB Professional cards (4GB modules are definitely a RTX 70/RDNA 6 thing, maybe 5GB Modules with the pro cards in 2029 or so?).

But as you said before yeah Configs aren't necessarily final both for the hardware silicon and what SKUs they take but the former is more fixed.

 

[Win2012R2]

MLID stuff does not seem to have been picked by neither the usual WCCF (who take anything) nor VideoCardz (who are very discerning), he must be upset given that Kepler_L2 confirmed it...

 

[fastandfurious6]

tbh MLID doesnt really care, its a youtube show with semi-accurate predictions, he makes a good living... that's all pretty much

 

[dangerman1337]

I may write out a diagram of what I think RDNA 5's ATx dies & various client dies may look like to explain to much people as possible, scan it and upload it here and other places. Just so it gives a speculative but educated view of how these ATx dies & various client dies may be combined (think a many-to-many relationship as seen in Databases, yes that's my old uni-computer science database course being of some use lol).

 

[Win2012R2]

I may write out a diagram of what I think RDNA 5's ATx dies & various client dies

Put AMD logo on it and email to MLID to see if he makes another video confirming his leak...

 

[marees]

I may write out a diagram of what I think RDNA 5's ATx dies & various client dies may look like to explain to much people as possible, scan it and upload it here and other places. Just so it gives a speculative but educated view of how these ATx dies & various client dies may be combined (think a many-to-many relationship as seen in Databases, yes that's my old uni-computer science database course being of some use lol).

So

48 CU AT3 xt 12gb = 16gb n48 xtx == $450?
40? CU AT3 xl 12gb = 16gb n48 xt == $400??
32?? CU AT3 le 9gb (or 16gb??) = 12gb n48 xl == $350???

Does that mean the n44 xt 16gb (9060xt) will have a very long run at a price around $300 ????

(& n33 below at $250 ? )

 

[PJVol]

I know this may be against forum niceness and etc., but sometimes it feels like an extra reaction icon (for example "stupidity") underneath a post wouldn't be inappropriate. Excuse me my offtopic (and this doesn't refer to the post right above)

 

[dangerman1337]

So

48 CU AT3 xt 12gb = 16gb n48 xtx == $450?
40? CU AT3 xl 12gb = 16gb n48 xt == $400??
32?? CU AT3 le 9gb (or 16gb??) = 12gb n48 xl == $350???

Does that mean the n44 xt 16gb (9060xt) will have a very long run at a price around $300 ????

(& n33 below at $250 ? )

IDK, I don't think AMD will do more than 2 SKUs for AT3 on dGPU desktop. Like 48 and maybe 36-40 CU wise. And maybe instead of N3P they'll fab it on N3C it keep the costs down so 48CU, 12GB = $300 WW (US Tariffs make everything harder to estimate) and 40-36 CU, 9GB = $225 WW.

$200 WW or below for a new dGPU release is off the table with TSMC N3 and equivalent nodes. It's way too expensive these days sadly. We're in a era where cards like the 9070 XT would've been $299 max are now $649 effectively. Considering Inflation + Node prices it's understandable though.

 

[soresu]

Put AMD logo on it and email to MLID to see if he makes another video confirming his leak...

Lol, he could have at least put the effort in by recreating a clean chart himself rather than just pushing out that low res image with his stupid banners all over it.

 

[Mopetar]

I know this may be against forum niceness and etc., but sometimes it feels like an extra reaction icon (for example "stupidity") underneath a post wouldn't be inappropriate. Excuse me my offtopic (and this doesn't refer to the post right above)

There used to be disagreement reactions but too many people were complete dipsticks with them and it was decided to remove them.

I do think it's better this way as if someone is being an idiot or has posted something utterly foolish, it's probably better for you to reply and explain why that's the case instead of clicking the loony react.

I do wish there were a reaction to indicate that something is interesting or thought provoking. Just use the thinking 🤔 emoji.

 

[soresu]

Seems I completely forgot about Altair, which has not yet been used as a GPU codename by AMD, so that could be AT too.

So far I have Altair, Antares and Amethyst as possible codenames.

Anyone else want to chime in?