Discussion RDNA 5 / UDNA (CDNA Next) speculation

[basix]

See? There is no market for a high-end Radeon.
Selling more for less is just not a sustainable market position anyway. They're not a charity.

You are thinking wrong.

AMD has currently nothing to offer in that space. If you release such a card with let's say 154 CU and 384bit SI (36 GByte VRAM) and fulfill the mentioned conditions, AMD knocks off both 5080 and 5090 (although only temporarily until Rubin release). So AMD would enter an new market space with relatively high margins (1500$ AT0 should yield in higher margins for AMD than a 600$ 9070XT).

So now you can choose:
1. Do not make this, make 0$ revenue and net income
2. Or make XX$ additional revenue and net income

2.) makes much mure sense in case you have the chip anyways developed and manufactured.
And when you beat Nvidia for once, even for a short time frame, it creates positive publicity for AMD GPUs.

The crucial part about all that is the release date. If AMD has 2-3 months time advantage 2.) will turn out well. If AMD releases at the same time or later than Rubin, it will get more difficult. But you might still win because of better P/L agains 6080 / 6090.

 

[olrak29]

That's perfectly assuming people will buy those cards lol

 

[Tarkin77]

https://twitter.com/x/status/2028320123342458885

whats this all about?? AMD/META Training ASIC?

 

[adroc_thurston]

1500$ AT0 should yield in higher margins for AMD than a 600$ 9070XT

You don't sell a winner for less.
How many times do I have to repeat that

 

[1250]

whats this all about?? AMD/META Training ASIC?

FP6? Is it META?

 

[basix]

That's perfectly assuming people will buy those cards lol

Best P/L >500$ and absolut best total performance than any other card on the market. (assuming AT0 releases before Rubin).
If releasing after Rubin it would still have better performance, P/L and more VRAM than a comparably expensive card (probably 6080).
DLSS vs. FSR will be much less of a factor by end of 2026 because FSR4+ narrows the gap. The same applies for RT and PT, where RDNA5 should perform well.

Why do you in return assume that people will not buy such a card?

And there will always be people not buying Nvidia cards because they are a bit unlikeable as a company.

You don't sell a winner for less.
How many times do I have to repeat that

You can repeat it as often as you want. If it is a winner, AMD might sell it for more. Does not really matter. AMD still makes additional money compared to a scenario not releasing such a card. But AMD needs to reduce prices later on, when Rubin is faster. Might be unpleasant for initial AT0 buyers. If it is not a winner (e.g. after Rubin release), AMD can sell it at that (1500$) price point and still earn some additional money from those cards. AMD would sell cards in a new (to them) and additional market region.

1500$ for a winner card would simply result in a very positive impression to gamers all around the world. Most gamers are fed up with Nvidia and their GPU prices. And you do not offend initial or AMD fan buyers, because the price stays fairly stable over time. Either you create short term value (high starting price) or long term value (normal starting price, loyal AMD buyers and positive AMD mindshare in general). Choose one.
A lower starting price might also attract a broader buyer pool. So a lower price might yield in overall higher revenue and profit. And each AMD card buyer will not buy Nvidia and at the same time convince 5 friends that AMD is cool. But that calculation is up to AMD.

 

[adroc_thurston]

AMD still makes additional money compared to a scenario not releasing such a card.

Oh no, g2m costs will make it a loss leader given AMD unit share.

1500$ for a winner card would simply result in a very positive impression to gamers all around the world

How many times do I have to repeat it?
They need to price it *higher* if it's faster.
You can only win by doing the winning. If they price it cheaper, they remain the poverty brand, and poverty is Death.

 

[basix]

1500$ card = poverty brand impression? You live in some sort of dream world...

1500$ is much money for most people. Most people buy at ther budget maximum. Because of that it does not matter, if there are even higher priced cards (5090's are not available anyways). You should just get your marketing right and even with 1500$ you get the longest benchmark bars. Benchmark bars matter most, if you want brag to your friends, not the price tag. The fastest car is often also more impressive compared to the most expensive car. Speed / performance can be felt. Price not.

 

[olrak29]

It will only happen if:
A. 9070 XT sold more than the 5070 Ti (despite being cheaper)
B. 9060 XT 8/16GB sold more than the 5060 Ti (again, both models were cheaper than the Nvidia equivalents)

But neither scenarios happened. AT0 cannot exist in a market that demands Radeon to be cheaper, and offer higher perf at the same time. Besides, at that higher price people will naturally gravitate to Nvidia because it offers more features for the price

 

[adroc_thurston]

1500$ card = poverty brand impression?

Yes.
If it's 30% faster than 5090 it needs to be priced at $2399.

 

[luro]

Ppl are delusional. AMD will NOT sell you top end hardware for cheap. I hope they launch it for $3k

 

[1250]

There's a rumor going around that Hynix won't be able to supply to Nvidia this year. It seems to be due to buffer die issues. I hope AMD gains hegemony ㅋㅋ

 

[rntongo]

Ppl are delusional. AMD will NOT sell you top end hardware for cheap. I hope they launch it for $3k

It will still be cheaper than an RTX 5090 and likely with better compute performance. $3k sounds realistic at current prices because of the memory crisis but who knows in 2027, it could be $2k.

Yes.
If it's 30% faster than 5090 it needs to be priced at $2399.

I think even with a halo RDNA5/UDNA card priced at $2399 it will sell out and offer crazy competition to Nvidia.

 

[eek2121]

Rolex sells watches.
Ferrari sells cars.
Swarovski sells diamonds.
etc


They sell walled garden devices that look like a PC, but this ain't a Personal Computer: Macs have 10% "PC" market share, it's a niche even for Apple.

btw I am not an Apple hater - I use iPhone/iPad, but won't use Macs even if they were free.

Macs aren’t a walled garden, it is just that most developers can’t be bothered to port software to the OS/hardware. Unlike an iDevice, you can install software from any source, Apple does not control this.

I do agree, however, that outside of niches such as software development, they are largely irrelevant. You aren’t going to find any serious gamers gaming on a Mac, for example.

 

[Win2012R2]

Not a chance.

If what MLID says is true about Magnus been ready for production, then so should be AT0 - now Xbox is in deep crisis and memory costs will only make it worse, so it makes sense for Xbox to wait till late next year, or just cancel it, but AMD should also have AT0 near ready, if they put it into production now then should be out end of 2026, which will sell as hotcakes.

They're not a charity.

Selling AT0 for $1500 should give at least 50% margin, that's indeed not a charity at all, and since they have already sunk costs into R&D then production should be no brainer: it will prep ground well for RDNA5/6 also, which is - priceless.

Easiest fix for everyone is abandoning Execute Indirect. Only GPU Work Graphs with DXR 1.3.

And backwards compatibility? Any new card will be expected to run old stuff faster, otherwise it will be a bust. It will take at least few years after PS6 launch before work graphs are used by some seriously popular game.

I think it could end up around $2000.

If it costs that much to double my 4090 perf then I am in, more - no.

 

[adroc_thurston]

Selling AT0 for $1500 should give at least 50% margin

Much less at their units.
It needs to cost twice that to actually have healthy margins.

 

[Win2012R2]

You don't sell a winner for less.

You said there was no market for it and now it's a winner? 🙄

It needs to cost twice that to actually have healthy margins.

Prove it. Here is my proof that you are wrong.

5090 chip size: 24x31=744 sq mm, ideal yield: 73 per wafer, also assume total worst case of 50% get thrown away completely (which seems way too high even for this big chip):

Chips/wafer
Process	Wafer cost	73 chips	36 chips
N3	$ 20,000	$ 274	$ 556
N2 (maybe)	$ 25,000	$ 342	$ 694
N2 (max)	$ 30,000	$ 411	$ 833

​

That's 5090 sized chip, not AT0 which is smaller, even with high memory prices selling GPU for $1500 will make very good money, as long as it sells in volume, which it will IF it is substantially quicker than 5090 (which in UK currently sold for ~ £ 3000 (GBP)), sell it for $2000 at start, discount later to $1500. Very clear business case to make good money, have halo part that will rub off nicely on the rest of the range: easily marketed as long term support because RDNA5 is in PS6.

And this is 5090 sized chip, which AT0 ain't, i fit is reported 600 sq mm then we are looking at total of 88 chips per wafer (in ideal case). What would be real total write off detects, max of 25%? Certainly not 50% I assumed above.

Here is same calculation for AT0 (22x28=616 sq mm): 88 total chips, 66 - 25% total write off, 44 is 50%.

Process	Wafer cost	88	66	44
N3	$ 20,000	$ 227	$ 303	$ 455
N2 (maybe)	$ 25,000	$ 284	$ 379	$ 568
N2 (max)	$ 30,000	$ 341	$ 455	$ 682

​

And AMD won't make a bucket load of money if core GPU wafer costs are $303???

Since memory costs are high then make it on N2 then and sell for $2k, heck I'd pay $2500 for a good N2 GPU like that, will be good enough for a fair few years.

 

[adroc_thurston]

You said there was no market for it and now it's a winner? 🙄

Yes, it's the eternal Radeon problem.

Very clear business case to make good money

High program costs, expensive g2m, no units. Dead part.

 

[Momoka_]

https://twitter.com/x/status/2028320123342458885

whats this all about?? AMD/META Training ASIC?

It may be about the customized SoC used by Meta Phoenix (next-gen MR/VR headset).

 

[Win2012R2]

High program costs

R&D already done for AT0, they can't obviously justify using N2 right now because that will all go to MI455 and Venice, but N3 is a no brainer, with plan to follow up with N2 version when volume allows in 2027/28.

That's the way.

 

[adroc_thurston]

R&D already done for AT0, they can't obviously justify using N2 right now because that will all go to MI455 and Venice, but N3 is a no brainer, with plan to follow up with N2 version when volume allows in 2027/28.

N3 is a far more contested node lmao.
In any case, hard R&D and TO are not the only parts of the program cost.

 

[Win2012R2]

N3 is a far more contested node lmao.

Contest then, offer 10% more per wafer, extra 30 bucks per GPU to make a shedload of money.

And it's the node for PS6, a mass market cheap device.

You are just coming up with new ridiculous excuses that are demonstrably bs.

Personally I'd certainly prefer to pay extra $500 for N2 version of a chip with all the same (as in memory) being equal, but it's not realistic for gaming device right now.

 

[adroc_thurston]

Contest then, offer 10% more per wafer, extra 30 bucks per GPU to make a shedload of money.

NV can pay much higher premiums.
Why are you doing this?

And it's the node for PS6, a mass market cheap device.

That was the intention, yes. But now the rules of the game changed.

You are just coming up with new ridiculous excuses that are demonstrably bs.

You gotta pay man, no one's selling you a pile of N3 for $1500.
And yes it's H2'27 (very very technically so).

 

[Win2012R2]

NV can pay much higher premiums.

My guess is that they are not limited by N3 wafers and already pre-paid for needed capacity a while ago.

 

[adroc_thurston]

My guess is that they are not limited by N3 wafers and already pre-paid for needed capacity a while ago.

Yes, they've made the node more crowded and contirubted to sequential wafer ASP hikes.
tough luck for client GPUs (especially AMD ones).