Discussion RDNA 5 / UDNA (CDNA Next) speculation

[adroc_thurston]

Even Intel is trying to release some sort of high end GPU card pretending to be AI card. It will be using LPDDR5x.

Idgaf.
There is no market for a $2k Radeon. Got it?

It would be foolish for AMD to not have a product offering, or even better, multiple products from Medusa Halo to AT0 for potential use as local AI

You're paying $2k. Why would you waste money on a Radeon?

It has to be 2k? 9070 XT is 357mm² with MSRP of $600, if AMD double the chip to 714mm², they couldnt charge $1500? (or 1600, 1650?)

Not how yields scale lmao.

Is the flexIO the million PCIe6 lanes that Mi455 will have, to be used defiantly for Meta?

They're not PCIe6 at all, but UAL.

 

[1250]

AMD may have done a good job, but they don't know anything about it. They say c HBMs are in the concept development and product development stages.

 

[gdansk]

It has to be 2k? 9070 XT is 357mm² with MSRP of $600, if AMD double the chip to 714mm², they couldnt charge $1500? (or 1600, 1650?)

1. Yield models suggest that would be like 3x the cost. Even with TSMC low defect hacks.
2. N3P costs more
3. GDDR7 costs more
4. Bigger chips need bigger margins to make it through planning
5. MSRP for N48 was only real (in the US) for about November 2025
6. AMD follows the Nvidia Minus policy regardless
7. AT0 is kill anyway

 

[marees]

It has to be 2k? 9070 XT is 357mm² with MSRP of $600, if AMD double the chip to 714mm², they couldnt charge $1500? (or 1600, 1650?)

384 bit bus
24gb or 36gb or 48gb ram (~6x PS5 pro)

maybe a further cut down 320 bit bus (~5x PS5 pro)
that is 20gb or 30gb

all variants would perform better than a 5090
so all variants priced higher than $1500 at the minimum

 

[1250]

I guess b–die development is being led by amd Or is it being developed overseaЅ

 

[Win2012R2]

There is no market for a $2k Radeon. Got it?

I'll buy it*. Got it? Write it down.

* if it's 30%+ faster than 5090 with max 600W.

Not how yields scale lmao.

Woot, no redundancy on big GPUs? Nvidia is using same fab and it works for them, and 5090's production cost isn't $1000 per chip.

1. Yield models suggest that would be like 3x the cost. Even with TSMC low defect hacks.

Yes, that's fine - fab cost per chip will certainly be way way under $1k, maybe they should just sell those direct under their own brand to cut out middlemen, keep the margin.

 

[adroc_thurston]

I'll buy it*. Got it? Write it down.

Not a real market.
Anyone sane will always pay a bit more to get an Nvidia card.

Woot, no redundancy on big GPUs?

Redundancy doesn't help when die sizes go up and up.
NV's just nowhere near bad enough PPA-wise for AMD to manifest a real cost advantage over them.

maybe they should just sell those direct under their own brand to cut out middlemen, keep the margin.

AIBs will riot.

 

[1250]

I wanted to claim that when Su and Jaeyong met, the discussion was primarily about HBM, but I gave up.
Hmm, a dedicated line? Seems like a tight schedule.
Do you think AMD will announce that they will use 3Ѕtar(special) hbm or will someone else do it?

 

[marees]

Anyone sane will always pay a bit more to get an Nvidia card.

I usually think like below

never buy gillette razor if there is an alternative
or buy the cheapest available gillette razor if there is no alternative

(doesn't matter if my face is bleeding with cuts)

 

[adroc_thurston]

I usually think like below

never buy gillette razor if there is an alternative
or buy the cheapest available gillette razor if there is no alternative

(doesn't matter if my face is bleeding with cuts)

It's just the general market sentiment.
People gladly pay more for NV, moreso in segments where money is less of a question.

 

[MrMPFR]

It's just the general market sentiment.
People gladly pay more for NV, moreso in segments where money is less of a question.

Considering NVIDIA's dominance in CUDA + proviz then who is paying for AMD's Radeon Pro GPUs?

 

[adroc_thurston]

Considering NVIDIA's dominance in CUDA + proviz then who is paying for AMD's Radeon Pro GPUs?

I dunno.
Probably no one, AMD doesn't even bracket proviz anywhere.

 

[Win2012R2]

Not a real market.

I had Titan, 3090, 4090 - yes, I am actually the real market, only reasons I did not get 5090 is because FE was not in stock and perf increase is too pathetic for crazy 600W.

Redundancy doesn't help when die sizes go up and up.

Yeah of course it doesn't - that's why Nvidia keeps making huge chips.

 

[adroc_thurston]

I had Titan, 3090, 4090 - yes, I am actually the real market, only reasons I did not get 5090 is because FE was not in stock and perf increase is too pathetic for crazy 600W

Not my point.
I'd rather not deal with NV display cores either, but as far as the market goes, they're the only game in town.

Yeah of course it doesn't - that's why Nvidia keeps making huge chips

For non-cost constrained markets.
AMD is always cost constrained.

 

[Win2012R2]

For non-cost constrained markets.

Like top end gaming cards, if they can't hack it at full size chip then maybe bring back fixed chiplet arch - put bigger, even retina max processing piece as GCD - and collect money.

 

[adroc_thurston]

Like top end gaming cards

Those are still cost constrained if you're a poverty brand.

put bigger, even retina max processing piece as GCD - and collect money.

No one pays top dollar for a Radeon. Get real.

 

[Momoka_]

Even Intel is trying to release some sort of high end GPU card pretending to be AI card. It will be using LPDDR5x.

Crescent Island?

Why not use some versions of AT0 and release it in a way that it could be used for local AI.

Based on this info, it looks like the full-spec AT0 core is being used as an inference card.

 

[adroc_thurston]

Based on this info, it looks like the full-spec AT0 core is being used as an inference card.

It's a cloud gaming GPU.

 

[Elfear]

Not a real market.
Anyone sane will always pay a bit more to get an Nvidia card.

I don't pretend to know all the intricacies of GPU sales & demand, but the 9070XT seems to be selling very well compared to the 5070Ti. This is an AI overview so take it with a few tablespoons of salt:
"AMD Radeon RX 9070 XT has outperformed the GeForce RTX 5070 Ti in sales since their early 2025 launch, often selling 10 times more units at major retailers like Mindfactory. The 9070 XT ($599) offers superior value, often outselling the entire NVIDIA 50-series lineup combined, while the 5070 Ti ($749+) has suffered from lower availability and higher prices"

If the potential street price matchup were AT0 for $2k up against a 6090 for $3-4k or a 6080 for $2k but 20% less performance, I think a lot of folks would grab the Radeon. I know I would. My hypothetical pricing may be off, but I think there is defintely potential for a halo AMD card. *my tarnished and gouged 0.02.

 

[adroc_thurston]

but the 9070XT seems to be selling very well compared to the 5070Ti

1:8/1:7 ratio.
It's doing badly in units and alright in margins.

If the potential street price matchup were AT0 for $2k up against a 6090 for $3-4k or a 6080 for $2k but 20% less performance, I think a lot of folks would grab the Radeon.

$2k is the segment where price is not a consideration and people pay top dollar for the best.
Radeon is for poor people; not the best. ever. at all.

 

[Thunder 57]

Those are still cost constrained if you're a poverty brand.

No one pays top dollar for a Radeon. Get real.

Meanwhile the rich brand continues to have poverty drivers:

https://videocardz.com/newz/nvidia-unlaunches-its-geforce-595-59-driver-after-reports-of-fan-and-clock-issues

 

[adroc_thurston]

Meanwhile the rich brand continues to have poverty drivers:

https://videocardz.com/newz/nvidia-unlaunches-its-geforce-595-59-driver-after-reports-of-fan-and-clock-issues

yes Blackwell screwups would've been perfect, but:
a) AMD ain't doing too hot s/w-wise either
b) no halo part. RIP bozo.

 

[gdansk]

Meanwhile the rich brand continues to have poverty drivers:

https://videocardz.com/newz/nvidia-unlaunches-its-geforce-595-59-driver-after-reports-of-fan-and-clock-issues

I think the general rule, now for both AMD/Nvidia, is not to update your drivers unless you're having problems with something.

They both seem to have some funny SNAFUs from time to time. Though my Blackwell card was a real PITA for the first few months.

 

[Saylick]

Meanwhile the rich brand continues to have poverty drivers:

https://videocardz.com/newz/nvidia-unlaunches-its-geforce-595-59-driver-after-reports-of-fan-and-clock-issues

Hey, when gaming is <10% of Nvidia's revenue, it's natural it only gets <10% of the their effort.

 

[Thunder 57]

I think the general rule, now for both AMD/Nvidia, is not to update your drivers unless you're having problems with something.

That's how I've done it for years. Unless there is a new feature or driver necessary for a game I leave it alone.