Discussion RDNA 5 / UDNA (CDNA Next) speculation

[adroc_thurston]

the only real case where AMD get really close to Nvidia in performance wise was RDNA2 with 520 mm2 big chip

great news man, AT0 is a chungus die with 96WGPs and a wee little bit of SRAM.

RDNA3/4 "chiplet GPU king"

That was a self-own yes, they've never actually built the big part.

 

[Win2012R2]

RDNA 4 was not a chiplet.

It seems to work very well, so that's hardly a disadvantage.

If RDNA5 is 50% quicker than 5090 (which will be then close to double of 4090) then I'll be very interested, as long as ray tracing does not totally suck. I find it hard to believe it will be though, maybe if it was on N2 but it ain't.

 

[CakeMonster]

There are reports of Sony contemplating delaying PS6 to 2028 or even 2029.

2028 seems like a no brainer given memory prices, and in any case they will benefit from having as large PS5 installment base as possible.

Now N3 was released in 2023, so how silly it would be to use N3 design in 2029 for a console meant to last like 7+ years, that's like Switch 1 territory rather than next gen.

Would be really pathetic if they don't bite the bullet and make new design for N2.

If delay is what is needed to not cut down the amount of RAM because of cost, I'm perfectly fine with it. The length of generations might even be increasing, so its pick your poison in this case.

 

[Win2012R2]

If delay is what is needed to not cut down the amount of RAM because of cost, I'm perfectly fine with it.

Me too, it's sensible to delay even without explosion in memory prices. I just hope they will port current design to better node, in N2 in 2029 should be in theory reasonably plentiful.

 

[marees]

Me too, it's sensible to delay even without explosion in memory prices. I just hope they will port current design to better node, in N2 in 2029 should be in theory reasonably plentiful.

unlikely to change node. example is switch 2

 

[marees]

from AMD:

Speed is the Moat: Inference Performance on AMD GPUs​

Feb 17, 2026

Instinct MI450 Series GPUs and Helios are doing great in the labs and we are actively going through our hardware and software validation steps. We remain on track for second-half production ramp and initial customer deployments.

https://www.amd.com/en/developer/resources/technical-articles/2026/inference-performance-on-amd-gpus.html

https://twitter.com/x/status/2023808516117463427

 

[1250]

Speed is the Moat: Inference Performance on AMD GPUs​

Feb 17, 2026

Instinct MI450 Series GPUs and Helios are doing great in the labs and we are actively going through our hardware and software validation steps. We remain on track for second-half production ramp and initial customer deployments.

https://www.amd.com/en/developer/resources/technical-articles/2026/inference-performance-on-amd-gpus.html

I think lpddr5 should do well.

 

[Tachyonism]

While the whole PS5 RDNA1.5 situation is going on again, I wonder how much does Sony/Mark Cerny actually influent RDNA5's design/features set.

 

[marees]

While the whole PS5 RDNA1.5 situation is going on again, I wonder how much does Sony/Mark Cerny actually influent RDNA5's design/features set.

my understanding is that AMD gives a buffet of options to Sony. Sony picks & chooses what it needs
this happens pretty early in lifecycle

Microsoft doesn’t seem to contribute anything — other than those that influence direct3d architecture. Basically MS tries to ensure that console & GPU have same tech to guarantee PC gaming capabilities

 

[Win2012R2]

unlikely to change node. example is switch 2

Yes, agree, but that's not Nintendo we talking about and it's not a handheld for much more price sensitive market segment.

Since PS6 does not use a lot of SRAM one would have thought porting this to N2 would shrink nicely, so smaller chip and gain some extra frequency also, problem is Xbox by then will be a footnote in history, so no real pressure on Sony to do it.

 

[tsamolotoff]

they are a professional outfit that obviously makes Nvidia want them to cover them

And surely nV doesn't give them specific 'recommendations' on how to tests their HW and what new features are now 'very important'. But then we had their tests where they've not noticed horrible artifacts with frame generation (and never emphasized that this is not real performance, just motion smoothening algo like SVP for videos), tested Doom Eternal in the resolution / settings that allowed it to fit into 8gb vram buffer of the card they were 'previewing' etc. Of course, it's all legal, proper and technically correct, but that doesn't preclude the fact that nV foundry is closer to first party tech PR rather than objective journalism (and they've been championing TAA derivatives for years and scoffing at people who don't like their details and motion clarity sacrificed for 'progress' that mostly serves to cut game dev costs)

 

[Win2012R2]

objective journalism

Sadly this died in tech long time ago (and barely exists now in other segments), it was only financially possible to do at the time when there were 40-50 original GPU makers.

Pretty sure we'll soon see AI reviewing this stuff, time for NV to cut out he middlemen!

 

[techjunkie123]

Pretty sure we'll soon see AI reviewing this stuff, time for NV to cut out he middlemen!

Introducing the 7060! Now with AI-enabled 16 GB of virtual RAM that quadruples the 4 GB of VRAM we used for this product.

50% faster than the 5060** at the same price.

**8x MFG on, vs no MFG.

 

[marees]

this is not what happened with the 9070xt tho

https://twitter.com/x/status/2024287221361045955

 

[marees]

is this GFX1310 ??

Wccftech thinks samsung exynos 2700 launching this year could have RDNA 5 GPU

the graphics processor might be called the Xclipse 970, and given that the Exynos 2600’s Xclipse 960 is based on a modified version of ARM’s RDNA 4 architecture, the newer GPU could be based on RDNA 5.​

Exynos 2700 Specifications, Features, Architecture, Launch Details - Here’s All You Need To Know About Samsung’s Next 2nm Chipset

The next 2nm chipset from Samsung is expected to be the Exynos 2700, and here’s everything you need to know about the silicon

wccftech.com

According to The Korea Economic Daily citing Kiwoom Securities, Samsung plans to start mass production of the Exynos 2700 in the second half of 2026. This chip will be manufactured using Samsung Foundry’s advanced second-generation 2nm Gate-All-Around (GAA) fabrication process, known as SF2P.​

​

Early production yields for the Exynos 2700 have reached approximately 50%, which is considered a strong performance for an initial output from a cutting-edge 2nm process. ​

​

the chip will integrate the new Xclipse 970 GPU, developed through Samsung’s continued collaboration with AMD. Early engineering samples show this GPU’s clock speed and performance are still modest, but there is room for improvement before official launch.​

Samsung’s Exynos 2700 Mass Production Set to Shake Up Flagship Chip Market, Challenging Snapdragon’s Dominance in 2026

Samsung is gearing up for a major shift in its mobile processor strategy with the upcoming Exynos 2700. According to The Korea Economic Daily citing Kiwoom Securities, Samsung plans to start mass production of the Exynos 2700 in the second half of 2026. This chip will be manufactured using...

semiwiki.com

 

[Win2012R2]

This chip will be manufactured using Samsung Foundry’s advanced second-generation 2nm Gate-All-Around (GAA) fabrication process, known as SF2P.

Well, if it is RDNA5 and AMD does not make yuge dGPU on the same process...

 

[reaperrr3]

is this GFX1310 ??

Highly doubt it.
WTFTech are just speculating ("could be based on RDNA5").
IIRC, kepler or adroc already mentioned that GFX131x are the client/Microsoft GPUs.

I'd assume Samsung GFX IP will lag behind AMD consumer GPUs by at least half a year.
Pretty sure Xclipse 970 is mostly a shrink of Xclipse 960 to SF2P, maybe with 2 additional CUs and/or some minor optimizations.

 

[adroc_thurston]

Well, if it is RDNA5 and AMD does not make yuge dGPU on the same process...

Client teams are banned from building big dies by Lisa herself.

 

[Win2012R2]

Client teams are banned from building big dies by Lisa herself.

And AT0 is what?

 

[adroc_thurston]

And AT0 is what?

Cloud VDI part.
Same as Navi12.

 

[Win2012R2]

Cloud VDI part.

So no dGPUs using it will be released?

 

[1250]

Mi455x qual(quality test)
I know Samsung was the only company that passed, but I wonder if other companies also received it.
Suddenly 미안 실례가 안된다면 통과 조건 아시는분

 

[adroc_thurston]

So no dGPUs using it will be released?

Nope.

 

[ToTTenTranz]

Cloud VDI part.
Same as Navi12.

Navi12 went to Apple laptops as well.

Nope.

MLID said AT0 is coming to desktop dGPUs.

 

[jpiniero]

MLID said AT0 is coming to desktop dGPUs.

MLID says a lot of things.