*** AMD "Navi 3X" (RDNA 3 discussion thread) ***

[4K8KW10]

Now we know what RDNA 2 in Navi 21 is capable of, let's speculate and wish what improvements can be made.

Ultra HD 2160p gaming with higher framerates looks like one of them.
Shrink to TSMC N5 process?
Adding more Infinity Cache - up to 256 MB or more?
Increasing the width of the memory bus and/or better memory - GDDR-next-gen?


AMD Zen 4, RDNA 3 to repeat recently seen generational leaps
https://hexus.net/tech/news/cpu/146824-amd-zen-4-rdna-3-repeat-recently-seen-generational-leaps/

As per our headline the AMD EVP was then quizzed about product refreshes going forward. "Everything is scrutinized to squeeze more performance out," replied Bergman and indicated that a similar long list of optimisations strategies are available that made the 19 per cent Zen 3 IPC gain possible – and Zen 4 will be moving to 5nm. On the topic of RDNA 3 GPUs, Bergman later confirmed that AMD was again targeting a 50 per cent plus GPU performance per watt improvement from generation to generation.

 

[humbug]

And so it starts again.... we just had a hard 12 months of RDNA2 speculation, can we have a rest now?

 

[4K8KW10]

And so it starts again.... we just had a hard 12 months of RDNA2 speculation, can we have a rest now?

Hard? Yes, I agree the wait for proper graphics from AMD was difficult for everyone - since now the pressure over AMD is much lower, we can relax and enjoy this discussion.

The news won't wait us to rest.

It would also be nice if AMD improves the ray-tracing performance or releases new features and better ecosystem support for the existing ones - Contrast Adaptive Sharpening - Fidelity FX, better artificial intelligence for some upscaling and increasing the framerate?
Next get DirectX? DirectX 13?

 

[varkanoid]

And so it starts again.... we just had a hard 12 months of RDNA2 speculation, can we have a rest now?

yeah at least give it a week. Let people enjoy their queue position for once. sigh.

 

[4K8KW10]

Navi 3X on TSMC N5 process - with I/O? What is I/O - another chiplet?
The timeframe is also not given - if we are lucky sometime in 2021, if not - in 2022...


NVIDIA ADA LOVELACE GPU Gets First Rumored Specs: Absolute Monster At 18432 CUDA Cores And 64 TFLOPs Of Graphics Horsepower (wccftech.com)

 

[Mean Mr Mustard]

The way things are going this will be officially announced right around the time a decent supply of 6800XTs hits the shelves.

 

[humbug]

Navi 3X on TSMC N5 process - with I/O? What is I/O - another chiplet?
The timeframe is also not given - if we are lucky sometime in 2021, if not - in 2022...


NVIDIA ADA LOVELACE GPU Gets First Rumored Specs: Absolute Monster At 18432 CUDA Cores And 64 TFLOPs Of Graphics Horsepower (wccftech.com)

I terms of Ryzen The IO Die is separate to the cores, could be the same thing, it means the IO Die can be made on a cheap node, in this case 12nm GloFo and more expensive TSMC 7nm 8 core CPU chiplets glued on to it.

 

[4K8KW10]

Yeah, very interesting

Could be only for yields and decreased production costs, could be a new breakthrough like the Infinity Cache that greatly improves the performance?
The so called "uncore" part of the GPU to go to its own silicon, while the compute units to sit on their own.

I thought the ray-tracing cores could have their own silicon, too.

 

[humbug]

Yeah, very interesting

Could be only for yields and decreased production costs, could be a new breakthrough like the Infinity Cache that greatly improves the performance?
The so called "uncore" part of the GPU to go to its own silicon, while the compute units to sit on their own.

I thought the ray-tracing cores could have their own silicon, too.

Yes exactly

 

[Kaapstad]

Close the thread for 12 months, at least after that amount of time people can make informed guesses.

Did anyone get much right about RDNA 2 before its launch?

 

[FoxEye]

Close the thread for 12 months, at least after that amount of time people might be able to buy Navi 2 cards.

FTFY

 

[JediFragger]

Did anyone get much right about RDNA 2 before its launch?

Lisa Su did when she called it Navi 2x*

*Best case scenerio

 

[4K8KW10]

Close the thread for 12 months, at least after that amount of time people can make informed guesses.

Well, by doing so, we will miss any leaks or information from AMD about the new generation - there might be even presentation / teases / previews of Navi 31 at some point.

Did anyone get much right about RDNA 2 before its launch?

Actually, yes
There was an original leak from China stating that the new Navi 21 would be a large, 505 sq. mm GPU - this is from December 2019:

AMD's high-end Navi GPU details: twice as fast as Radeon RX 5700 XT?!
https://www.tweaktown.com/news/6954...tails-twice-fast-radeon-rx-5700-xt/index.html

Charlie wrote about Navi 21's tape out as early as November 2019:

A new high end GPU just taped out
A new high end GPU just taped out - SemiAccurate

 

[Dontrocktheboat]

Will it be able to play CP2077 with RT on?

That's the question? Because RDNA2 cant.

 

[nvidiamd]

Will it be able to play CP2077 with RT on?

That's the question? Because RDNA2 cant.

No it won't. We are all devastated already.
https://twitter.com/MokoMokie/status/1343806792384589829

 

[ANDREW GREEN]

Close the thread for 12 months, at least after that amount of time people can make informed guesses.

Did anyone get much right about RDNA 2 before its launch?

I predicted nearly twice as fast as a 5700 XT at 4k and was right

 

[LtMatt]

No it won't. We are all devastated already.
https://twitter.com/MokoMokie/status/1343806792384589829

That game is so bad it is painful to watch.

 

[Poneros]

I am curious to see what they do to improve RT performance. The rumours of Lovelace are that it will double-down on huge core counts (and consequently RT cores) which means that it will further smack any previous gen cards of any brand around when that's enabled, and make no mistake about it, 2 years from now RT will be a staple feature for most AAA games and even some AA studios, even if it it's only 1 or 2 effects as a toggle. While I bought an RX6800 and am ok with the more meagre performance in RT simply because there's not that many RT games I care about, I definitely don't want to make a similar purchase in 2 years. So it will be very important for them to focus on this aspect because I'm not alone in that assessment.

I have to say it is disappointing that AMD had 2 years to figure out DLSS and couldn't, when all they did for 2.0 was use temporal reconstruction but add an ML model for clamping. Hell even a push for more VRS would've been enough because at the end of the day all these solutions are comparable in performance & visuals. The RDNA 2 driver situation was stable at launch so kudos for that but gotta say feature-wise the AMD software team has not delivered much at all this year.

They've been lucky the demand this year has been insane regardless, that's all I'll say.

 

[4K8KW10]

AD102 is rumoured to have 18,432 cores, only 75% more than the 10,496 in GA102.
And no one guarantees that with such large shaders count, you will see any even close to linear performance scaling.

But that's normal - N5 is a full die shrink of N7.

 

[Poneros]

It's not an issue in the slightest for 4K + RT, utilisation will be there. For lower res, non-RT, high framerate gaming? Yeah, sure, but I don't care about that & for the most part cards are already surpassing monitor limits for that at least where it matters (i.e. esports). Doubt many people buying high-end GPUs care about it either.