Discussion RDNA 5 / UDNA (CDNA Next) speculation

[jpiniero]

Yes, the upcoming exploding demand for VRAM/RAM to load NN models is one of the main reasons AMD is going with smartphone memory for their lower end GPUs. And it's driven by developers (e.g. their close relationship with Sony).

You need to stop smoking MLID. It'll rot your mind

 

[CakeMonster]

Yes, the upcoming exploding demand for VRAM/RAM to load NN models is one of the main reasons AMD is going with smartphone memory for their lower end GPUs. And it's driven by developers (e.g. their close relationship with Sony).

Some models, whether its for dialogue or upscaling or something new that will be invented in 2-5 years will most probably be part of future games. That's why I'm a tad nervous about next gen consoles if they cheap out on VRAM (say 16 or 24GB), after all the expected lifetime of those will be to 2034+ which is a long time...

 

[marees]

Some models, whether its for dialogue or upscaling or something new that will be invented in 2-5 years will most probably be part of future games. That's why I'm a tad nervous about next gen consoles if they cheap out on VRAM (say 16 or 24GB), after all the expected lifetime of those will be to 2034+ which is a long time...

Xbox could cheap out on vram because they can easily replace with RDNA 7 in 4 years time

PlayStation needs large vram — both for console & handheld

 

[marees]

4gb for AI models

https://twitter.com/x/status/1961410371832500294

 

[ToTTenTranz]

You need to stop smoking MLID. It'll rot your mind

Did the people at Ubisoft, Nvidia and InWorld AI also smoked MLID a couple years ago when they built a team to prototype game experiences with LLM-enhanced NPCs, which they demoed at GTC 2024?

How Ubisoft’s New Generative AI Prototype Changes the Narrative for NPCs

Have you ever dreamed of having a real conversation with an NPC in a video game? For the last few years, a small R&D team at Ubisoft’s Paris studio have been experimenting with generative AI in a first step toward making this dream a reality.

news.ubisoft.com

Was Nvidia high on MLID when they created a full set of tools exactly for that?

NVIDIA ACE for Games

Covers every aspect of digital humans—from speech to animation.

developer.nvidia.com

Was this developer high on this week's MLID rumors when they started making a game all about LLM-enhanced NPCs, due out this year?

Was Microsoft high on MLID in 2023 when they announced DirectML for DirectX 12 GPUs to run Llama 2 7B?

Announcing preview support for Llama 2 in DirectML

At Inspire this year we talked about how developers will be able to run Llama 2 on Windows with DirectML and t

blogs.windows.com

If so, then yes I guess I am high on MLID. 🤷‍♂️

 

[marees]

Good summary. Not just a copy paste

The latest AMD RDNA 5 rumours are complicated but it looks like there really is going to be a high-end next-gen GPU to take on Nvidia's best graphics card

And could low-end RDNA 5 use laptop memory?

www.pcgamer.com

 

[dangerman1337]

Some models, whether its for dialogue or upscaling or something new that will be invented in 2-5 years will most probably be part of future games. That's why I'm a tad nervous about next gen consoles if they cheap out on VRAM (say 16 or 24GB), after all the expected lifetime of those will be to 2034+ which is a long time...

Pretty much and hell Canis and Orion will be lasting beyond the PS7's launch (IMV 2035 on a A5 Node) because next-next-gen won't be cheap and if Canis becomes the "entry level baseline" for AAA gaming 2035-2040 for example with LLMs being part of games then it needs a lot of RAM not to be a big concern.

I wouldn't be suprised if it's 2042, 15 years after PS6 comes out and we still see AAA games on the PS6. I mean MWIV next year will be the first CoD dropping last-gen support which is 6 years after this gen came out (Vs IW being the first PS4/XBO only CoD after 3 years) so It wouldn't surprise me.

 

[adroc_thurston]

Agentic AI using smallish 1.5-2B language and text-to-speech models are 100% going into videogames next gen for NPC dialogues, world stage integration, environment adaptability, etc.

lmao no, these are so catastrophically unreliable any sort of gamedesign vision would be gone in an instant.

 

[dr1337]

lmao no, these are so catastrophically unreliable any sort of gamedesign vision would be gone in an instant.

A custom made small model thats trained on dialog the dev wants and not just the entire knowledge of the internet would do a lot better than any distill you can download today.

 

[adroc_thurston]

A custom made small model thats trained on dialog the dev wants and not just the entire knowledge of the internet would do a lot better than any distill you can download today.

It's still gonna be horribly unreliable via the very nature of language models.

 

[ToTTenTranz]

lmao no, these are so catastrophically unreliable any sort of gamedesign vision would be gone in an instant.

Companies are developing agentic AIs to promote their products/services and go as further as close deals, multimodal solutions are being deployed for medical triages in private hospitals, but god forbid a videogame developer could implement a similar system for game character dialogues, lest a NPC do something as dangerous as say stupid lines.

 

[adroc_thurston]

Companies are developing agentic AIs to promote their products/services and go as further as close deals, multimodal solutions are being deployed for medical triages in private hospitals

none of that works.
next.

 

[ToTTenTranz]

none of that works.
next.

Ok I'll send an email to those private hospitals and let them know the triage systems they're deploying don't work because you said so.

 

[Claudiovict]

The Chiphell guy pretty much concurs with everything said so far, 4 dies, 2027 launch, very specifically no AT1.
Oh and also the doubling of CU size.

AT0 154CUs?

 

[branch_suggestion]

AT0 154CUs?

Nothing that specific.

 

[branch_suggestion]

Believing AMD slideware verbatim this far from launch has backfired before.
Their opsec have played games with leakers before, some quite hilariously.
Some leaked things were even indeed real, but were internal targets that were missed.

 

[soresu]

Oh and also the doubling of CU size.

More like they yeeted the RDNA1-4 WGP nomenclature and just started calling them CUs instead.

Whether that is because they simply thought WGP was redundant and/or confusing on top of CU language for consumers, or because WGP -> RDNA5 CU nomenclature indicates some greater interconnection between the dual 'lobes' of ALUs beyond just a clump of cache who knows.

Maybe both.

 

[adroc_thurston]

More like they yeeted the RDNA1-4 WGP nomenclature and just started calling them CUs instead.

No, memory architecture just changed.

beyond just a clump of cache who knows.

Wasn't even cache, WGP mode shared the LDS slab.

 

[soresu]

AT0 154CUs?

That was just a binning of a larger 192/96 CU die according to MLID's rumor.

Personally it sounds weird to me.

80/40 -> 192/96 is an insane shift of 2.4x for the top 2 SKU dies.

Massively more than the 60 -> 96 (1.6x) of the RDNA3 gen and even significantly beyond the 40 -> 80 (2x) of RDNA2 gen.

 

[adroc_thurston]

That was just a binning of a larger 192/96 CU die according to MLID's rumor.

Personally it sounds weird to me.

80/40 -> 192/96 is an insane shift of 2.4x for the top 2 SKU dies.

Massively more than the 60 -> 96 (1.6x) of the RDNA3 gen and even significantly beyond the 40 -> 80 (2x) of RDNA2 gen.

Shader delta only looks big because AT1 got scrapped.

 

[soresu]

Unless perhaps they are just keeping a midway AT1 design for a die shrink.

 

[adroc_thurston]

Unless perhaps they are just keeping a midway AT1 design for a die shrink.

Kilobuck GPUs aren't a real market per se.
Too expensive to be mainstream-ish yet not big enough to be noteworthy on any review charts.

 

[soresu]

Kilobuck GPUs aren't a real market per se.
Too expensive to be mainstream-ish yet not big enough to be noteworthy on any review charts.

That's the problem tho.

If Kepler's take is correct then the uptick in raw ALUs from N48 to AT2 is only 25% for raster, and I don't expect the clock increase to be that good.

That means it's either going to be a lackluster generational increase for the 9070 XT upgrade, or a very expensive 1090 XT unless they have actually finally gone for chiplets.

 

[branch_suggestion]

More like they yeeted the RDNA1-4 WGP nomenclature and just started calling them CUs instead.

Whether that is because they simply thought WGP was redundant and/or confusing on top of CU language for consumers, or because WGP -> RDNA5 CU nomenclature indicates some greater interconnection between the dual 'lobes' of ALUs beyond just a clump of cache who knows.

Maybe both.

Well the conjecture is that the WGP is dead, for starters unifying L0 and LDS already makes the WGP pretty redundant.
Then removing WGP mode/Wave64 makes it completely dead. This makes the DCU behave like a 2x sized single CU without any other changes beyond that (though there should be).
Therefore you can call what was the WGP now a single CU.
To add my personal conjecture, some RDNA4 stuff like the dynamic register alloc only works in CU mode/Wave32, so I think they will lean that way especially if CDNA is doing the same.

Unless perhaps they are just keeping a midway AT1 design for a die shrink.

He's dead, Jim.
AMD used to make parts like AT1, they all failed to move the needle.
Simple fact is there is no market between the halo and the enlightened midrange and their market analysts and product steering group finally figured this out.
Every ATx part has reuse outside of being a client dGPU (okay I don't expect AT0 to do that much in the pro market but it is their best part there in over a decade), so no need to try to fill in every hole in the market when you can sell millions of them in mobile/semicustom APU form.

If Kepler's take is correct then the uptick in raw ALUs from N48 to AT2 is only 25% for raster, and I don't expect the clock increase to be that good.

Look at the clock uplift ARM guys have gotten from N4P to N3P parts, it is a nice node bump.

That means it's either going to be a lackluster generational increase for the 9070 XT upgrade, or a very expensive 1090 XT unless they have actually finally gone for chiplets.

Not sure what uplift you're expecting but it should be good enough to make Ada/RDNA3 owners interested.
AT0 is too small to bother with chiplets when a big mono die works fine.

 

[adroc_thurston]

If Kepler's take is correct then the uptick in raw ALUs from N48 to AT2 is only 25% for raster

16 additional shader cores and each one is far meaner + N3p clks.
It'll be good nuff.

or a very expensive 1090 XT

duhhhhhhhhhhh.
Who doesn't like $1999 GPUs?