Discussion RDNA 5 / UDNA (CDNA Next) speculation

[Tuna-Fish]

I think it is better to have the memory unified.

So are we now assuming that ALL of the consoles will use LPDDR?

MLID is claiming that PS6 normal will use GDDR7, the handheld and the "S" version that's based on the handheld chip use LPDDR5X.

 

[Joe NYC]

MLID is claiming that PS6 normal will use GDDR7, the handheld and the "S" version that's based on the handheld chip use LPDDR5X.

The memory sizes will make it challenging. Full console will surely need to have same or more memory, so the memory costs will start to add up...

 

[Timorous]

None.
None of them do anything

Bethesda might for NPC interactions. It will be hanky as anything but it is about the only use I see for it.

 

[adroc_thurston]

Bethesda might for NPC interactions

Oh no, they're not doing anything unorthodox ever again.

 

[MrMPFR]

Bethesda might for NPC interactions. It will be hanky as anything but it is about the only use I see for it.

The potential for open-ended sandbox and mass scale game worlds to use it is def there such as the virtual Game Master in the upcoming indie game The Waywards Realms.
There are many any uses for AI in video games besides boring NPC interactions.

Impact on AAA low, medium for AA, for indie impact could be huge.

 

[Timorous]

Oh no, they're not doing anything unorthodox ever again.

Does that mean they are dropping their game engine entirely?

 

[GodisanAtheist]

The potential for open-ended sandbox and mass scale game worlds to use it is def there such as the virtual Game Master in the upcoming indie game The Waywards Realms.
There are many any uses for AI in video games besides boring NPC interactions.

Impact on AAA low, medium for AA, for indie impact could be huge.

- I'd love to see an "AI Director" that's almost like a DM that tailors the game experience dynamically to keep a player engaged with a game.

Would make huge "static" open worlds so much more fresh than they've been for a very long time.

 

[adroc_thurston]

Does that mean they are dropping their game engine entirely?

I mean their gamedesign framework for TES6 will be a 1:1 clone of Skyrim.

 

[MrMPFR]

1. 2x Intersection Testing,
2. unified LDS/L0 Cache,
3. Dedicated Stack Management and Traversal HW,
4. Coherency Sorting HW, and
5. 3-coordinate decompression Geometry HW.

How much of this is confirmed by Kepler?

#3 RDNA3 already added stack management. Does this mean dedicated RTU cache like Intel or something else?

Is #5 DGF? Yes likely.

 

[adroc_thurston]

Does this mean dedicated RTU cache like Intel?

FF is Bad and Evil, so nope.

 

[adroc_thurston]

I'd love to see an "AI Director" that's almost like a DM that tailors the game experience dynamically to keep a player engaged with a game.

1) launch Rimworld
2) select Randy Random
3) enjoy!

 

[proelitedota]

The 40 "CU" one

View attachment 129371

Interesting, so a CU === WGP and Magnus disables 4 CUs / WGP.

 

[HurleyBird]

Oh no, they're not doing anything unorthodox ever again.

I'll be extremely surprised if someone doesn't hack something into Daggerfall Unity within the next couple of years.

 

[ToTTenTranz]

There's Medusa Halo with SoC (CPU + I/O) and AT3 GMD (GPU + Memory), Medusa Premium with smaller SoC (CPU + I/O) and AT4 GMD (GPU + Memory) and Medusa Point with SoC (CPU + GPU + Memory + I/O) plus optional CCD.

If each one needs to have their own SoC, it doesn't look like there's much modularity here. AT4 will hardly ever be successful as a dGPU with 135GB/s bandwidth on LP5X.
Even the APU version seems a bit worthless like this, as it's bound to be choked on bandwidth.

Either these things get paired with LPDDR6 or they're a repetition of their predecessors' flaws.

 

[marees]

If each one needs to have their own SoC, it doesn't look like there's much modularity here. AT4 will hardly ever be successful as a dGPU with 135GB/s bandwidth on LP5X.
Even the APU version seems a bit worthless like this, as it's bound to be choked on bandwidth.

Either these things get paired with LPDDR6 or they're a repetition of their predecessors' flaws.

The 9060xt 16gb goes for a street price of $400

Only buyable card below that is the 7400xt 55w

The AT4 can very well fill the niche of $250 to $350 with adequate vram (replacing the 3060 12gb finally)

AT3 is an issue. But here the backup plan would be to continue selling RDNA 4 cards at discounted price

 

[dangerman1337]

The 9060xt 16gb goes for a street price of $400

Only buyable card below that is the 7400xt 55w

The AT4 can very well fill the niche of $250 to $350 with adequate vram (replacing the 3060 12gb finally)

AT3 is an issue. But here the backup plan would be to continue selling RDNA 4 cards at discounted price

Really depends how AT3 & 4 dGPUs are priced because AT4 sounds way below that 9060 XT tier segment, wouldn't even be suprised if AMD prices it aggressively at $200 or below with 8GB of RAM. AT3 I think could hit $300 or so WW.

 

[adroc_thurston]

AMD prices it aggressively at $200 or below with 8GB of RAM

How do I pop 4 channels of LPDDR5X with only 8GB of DRAM.

 

[ToTTenTranz]

The what.
No one's stuffing language models into games.

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. We're probably talking about 8GB VRAM dedicated to these at the very least.

And they're going into the 2027 consoles for a handful of years before going to PC. Thank Nvidia for being super successful at shoving 8GB GPUs down PC gamers' throats on a $300-400 budget for DIY GPUs and <$2000 laptops, for an entire friggin' decade.
We'll be seeing lots of PC Master Race bros with desktop and laptop 60-series on a meltdown because they're either not being to play the games or are getting toned down experiences (or worse: they must pay a fee to get those features from a cloud service).

The 9060xt 16gb goes for a street price of $400

Only buyable card below that is the 7400xt 55w

The AT4 can very well fill the niche of $250 to $350 with adequate vram (replacing the 3060 12gb finally)

AT3 is an issue. But here the backup plan would be to continue selling RDNA 4 cards at discounted price

I just can't see how an AT4 with 12 WGPs / 24 CUs and 135GB/s bandwidth can have the necessary performance to run newer games adequately, even at just 1080p.
PS6 Canis is going with a 192bit bus for 200GB/s on a handheld with 16CUs at 1.3GHz (i.e. probably less than half the compute throughput of AT4).


Medusa dGPUs and iGPUs with LPDDR5X don't sound like good choices at all, save maybe for running low to mid-sized LLMs on a low budget.

 

[marees]

1

I just can't see how an AT4 with 12 WGPs / 24 CUs and 135GB/s bandwidth can have the necessary performance to run newer games adequately, even at just 1080p.

1080p max with FSR 4 (or 5,6 etc.) should be doable for AT4 right ?

 

[jpiniero]

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. We're probably talking about 8GB VRAM dedicated to these at the very least.

That would be cool... but yeah not happening.

The Sony handheld is going to flop if the specs are even remotely close to accurate since it wouldn't be fast enough to run PS5 games without developer intervention. Devs wouldn't bother. But since we are talking about MLID, they likely aren't.

 

[Kepler_L2]

That would be cool... but yeah not happening.

A cool idea if you want your FPS to plummet everytime you speak to an NPC

 

[dangerman1337]

I just can't see how an AT4 with 12 WGPs / 24 CUs and 135GB/s bandwidth can have the necessary performance to run newer games adequately, even at just 1080p.
PS6 Canis is going with a 192bit bus for 200GB/s on a handheld with 16CUs at 1.3GHz (i.e. probably less than half the compute throughput of AT4).

Well there's 12.7 Gbps LPRR5X in the works:

Samsung extends LPDDR5 to 12.7 GT/s: Next-gen devices enjoy a nice speed boost

Samsung's LPDDR5-Ultra-Pro further clings into LPDDR6 territory

www.tomshardware.com

Very concievable that a very cheap (no more than $200) AT4 dGPU with 8GB card targetting 1080p online F2P games (think CS2, DOTA, Fortnite etc). And then 256-bit 16GB AT3 dGPU would be enough for 1080p-"entry level" 1440p if could be sold at $300 would bring a new level of performance to the masses Vs now.

 

[ToTTenTranz]

A cool idea if you want your FPS to plummet everytime you speak to an NPC

Not if you only need like 8 tokens/s (plentiful for conversations, overkill for dynamic environmental changes and world updates) and properly allocate the resources for it.

The latest developments on hardware-aware optimizations like Jet-Nemotron get the old Orin (50 TOPS INT8) to output 55 token/s on a 2B LLM.
Developers can eventually allocate like the equivalent of 10-20 TOPs INT4 of a GPU or NPU for this and would probably suffice. It just needs enough RAM for having the model loaded and a proper context window.


Videogames, mainly all the high profile sandbox RPGs and action games, will be changing drastically with agentic AI.
Just not for the folks who bought a 3060Ti/3070, or a 4060 8GB, or a 5060 8GB, or a 5070 8GB laptop. I.e. all those top sellers in Steam's Hardware Survey.

 

[dangerman1337]

Not if you only need like 8 tokens/s (plentiful for conversations, overkill for dynamic environmental changes and world updates) and properly allocate the resources for it.

The latest developments on hardware-aware optimizations like Jet-Nemotron get the old Orin (50 TOPS INT8) to output 55 token/s on a 2B LLM.
Developers can eventually allocate like the equivalent of 10-20 TOPs INT4 of a GPU or NPU for this and would probably suffice. It just needs enough RAM for having the model loaded and a proper context window.


Videogames, mainly all the high profile sandbox RPGs and action games, will be changing drastically with agentic AI.
Just not for the folks who bought a 3060Ti/3070, or a 4060 8GB, or a 5060 8GB, or a 5070 8GB laptop. I.e. all those top sellers in Steam's Hardware Survey.

Probably why AT3 has a 256-bit bus so they can do 16GB Vs 128-bit with 12GB is my guess. Those AT4 peeps who get a dGPU variant of it are the "I only play CS/LoL/Sims/DOTA/Fortnite at 1080p" etc crowd who exist all around the world like in SEA, Latin America, South Asia, Eastern Europe etc.

 

[ToTTenTranz]

Probably why AT3 has a 256-bit bus so they can do 16GB Vs 128-bit with 12GB is my guess. Those AT4 peeps who get a dGPU variant of it are the "I only play CS/LoL/Sims/DOTA/Fortnite at 1080p" etc crowd who exist all around the world like in SEA, Latin America, South Asia, Eastern Europe etc.

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).