Discussion RDNA 5 / UDNA (CDNA Next) speculation

[ToTTenTranz]

They killed Xbox with their own dumb moves.

They killed xbox with their crass inability to manage 1st party studios.

They demand their studios to apply for competitive public financing which screws up deadlines, and too much outsourcing which prevents talent and knowledge from staying in the company. Everything takes way too long and they need to constantly reinvent the wheel.


The funny part is Phil Spencer has been sinking the platform/brand for over a decade now, but xbox fans still blame Don Mattrick for its state nowadays.


Even the "day one gamepass" concept is completely moronic, as it screws up people's perception of a game's value (not that their games have that much value at release though).

 

[SolidQ]

https://twitter.com/x/status/1896672391457349981

RDNA5? dev kit for next Xbox is already in wild?

 

[HuesToo4]

https://twitter.com/x/status/1896672391457349981

RDNA5? dev kit for next Xbox is already in wild?

Possibly Zen 6 as well

 

[Kepler_L2]

https://twitter.com/x/status/1896672391457349981

RDNA5? dev kit for next Xbox is already in wild?

Fake

 

[Tuna-Fish]

https://twitter.com/x/status/1896672391457349981

RDNA5? dev kit for next Xbox is already in wild?

The early dev kits are not usually some unreleased stuff that's far away, they are just a pc with components that roughly match the performance target. Quite possibly it's just a 9700X and a 9070XT. Or it might even have an NVidia GPU, iirc the first PS4 dev kit did.

There will be "higher fidelity" dev kits that have actual real hardware once the first chips start coming off the line.

 

[SolidQ]

From Mark Cerny quotes

So that seems RDNA5 is targeting more ML/RT/PT perf, also Neural denoising

 

[adroc_thurston]

So that seems RDNA5 is targeting more ML/RT/PT perf, also Neural denoising

It's targeting more juice per CU (again, yes. Again).

 

[SolidQ]

It's targeting more juice per CU

How do you think there will be CU's? 96 max?

Matrix Cores should improve FSR4 hit perf?

 

[adroc_thurston]

How do you think there will be CU's? 96 max?

No idea.

Matrix Cores should improve FSR4 hit perf?

Yeah kinda the point really.

 

[ToTTenTranz]

From Mark Cerny quotes

So that seems RDNA5 is targeting more ML/RT/PT perf, also Neural denoising

Honestly, this says very little about post-RDNA4.

Amethyst so far is basically a joint venture between AMD and Sony to develop a ML-based upscaler.
AMD makes no software but makes GPUs and Sony vice versa, so they're using Sony's dataset to train a ML model optimized for AMD's latest compute units.

Which means PSSR is probably a percursor to FSR4, and it seems likely to me that PSSR will eventually be upgraded to whatever the latest version of FSR4 Sony's QA engineers decide to validate. Even if both companies do decide to keep the naming separate for marketing reasons.

 

[adroc_thurston]

AMD makes no software

What the

so they're using Sony's dataset to train a ML model optimized for AMD's latest compute units.

Yes AMD trains stuff on their own hardware for their own hardware.

Even if both companies do decide to keep the naming separate for marketing reasons.

Yes they're completely separate, unrelated products.

 

[soresu]

AMD makes no software

Take a quick look at the Mesa/Gallium3D codebase for Linux and tell me with a straight face that drivers do not count as hella complex software development.

That's not even counting all the other stuff AMD have been up to over the years on the software side like Tress FX and various other things GPUOpen deals with.

 

[ToTTenTranz]

What the

Take a quick look at the Mesa/Gallium3D codebase for Linux and tell me with a straight face that drivers do not count as hella complex software development.

I meant videogame development, obviously.
Driver development isn't going to help much if what you need is a huge dataset of rendered frames using sky-high quality settings for training a CNN.

Yes AMD trains stuff on their own hardware for their own hardware.

Of course they do.
It's the dataset that is coming from people who make games. Sony's dev teams own the original, uncompressed textures and can dial IQ settings above the commercial game's settings in order to produce very high quality frames.

Yes they're completely separate, unrelated products.

Dude, Mark Cerny literally said they're not.

"The neural network (and training recipe) in FSR 4's upscaler are the first results of the Amethyst collaboration," Cerny told us.
(...)
Our target is to have something very similar to FSR 4's upscaler available on PS5 Pro for 2026 titles as the next evolution of PSSR; it should take the same inputs and produce essentially the same outputs.

They're using the same inputs, i.e. the same dataset.

 

[adroc_thurston]

They're using the same inputs, i.e. the same dataset.

Dataset is whatever, training recipe is where it at.

 

[SolidQ]

is That need a lot AI power?

 

[ToTTenTranz]

Dataset is whatever, training recipe is where it at.

Anyone who's meddled in ML training knows that producing the dataset is >90% of the work.

 

[adroc_thurston]

Anyone who's meddled in ML training knows that producing the dataset is >90% of the work.

Not really?
All the current ML developments are squarely related to model architecture and training recipes.
Which is yes, what's FSR4 is all about.

 

[marees]

Wrong thread

 

[blackangus]

Anyone who's meddled in ML training knows that producing the dataset is >90% of the work.

All the current ML developments are squarely related to model architecture and training recipes.

So you are both kinda right here.
In some industries building the dataset is difficult, there is alot to making sure your data is of high quality and the information is being presented appropriately.
Then there is still alot of scientific development going on with the mechanics, traceability, accountability, data formatting, and data ingesting areas. Lots of feature and optimization work to do in many areas based on where we are today.

I get to see both sides of the house in my current position.

 

[ToTTenTranz]

https://twitter.com/x/status/1899822577213645148

Project Amethyst got Sony collaborating with AMD in the FSR4 models, as now confirmed by Jack Huynh.


PSSR and FSR4 are the same project.

 

[poke01]

Is this still accurate meaning no real high end GPU like 7900XTX for RDNA5?

 

[adroc_thurston]

View attachment 120175

Is this still accurate meaning no real high end GPU like 7900XTX for RDNA5?

Yeah.

 

[marees]

View attachment 120175

Is this still accurate meaning no real high end GPU like 7900XTX for RDNA5?

Can be intrepreted as 3 chips instead of 2 (as in RDNA 4).

 

[Tuna-Fish]

I'd highlight that it covers a range including both what 7600 is now, and a range exceeding the top of N48. (Remember, this is not performance, this is market segments, performance in segment goes up every generation.)

So, on the low end probably a 96-bit GDDR7 chip with 9GB or something, and on the top end probably something above 256-bit.

 

[marees]

I'd highlight that it covers a range including both what 7600 is now, and a range exceeding the top of N48. (Remember, this is not performance, this is market segments, performance in segment goes up every generation.)

So, on the low end probably a 96-bit GDDR7 chip with 9GB or something, and on the top end probably something above 256-bit.

So segmented by VRAM, it is likely to be

1. 24gb - N51
2. 18gb - cut down N51
3. 12gb - N52
4. 9gb - N53

??