Discussion RDNA4 + CDNA3 Architectures Thread

[DisEnchantment]

With the GFX940 patches in full swing since first week of March, it is looking like MI300 is not far in the distant future!
Usually AMD takes around 3Qs to get the support in LLVM and amdgpu. Lately, since RDNA2 the window they push to add support for new devices is much reduced to prevent leaks.
But looking at the flurry of code in LLVM, it is a lot of commits. Maybe because US Govt is starting to prepare the SW environment for El Capitan (Maybe to avoid slow bring up situation like Frontier for example)

See here for the GFX940 specific commits

History for llvm/lib/Target/AMDGPU - llvm/llvm-project

The LLVM Project is a collection of modular and reusable compiler and toolchain technologies. - History for llvm/lib/Target/AMDGPU - llvm/llvm-project

github.com

Or Phoronix

More AMD "GFX940" Enablement Work Landing In LLVM - Phoronix

www.phoronix.com

There is a lot more if you know whom to follow in LLVM review chains (before getting merged to github), but I am not going to link AMD employees.

I am starting to think MI300 will launch around the same time like Hopper probably only a couple of months later!
Although I believe Hopper had problems not having a host CPU capable of doing PCIe 5 in the very near future therefore it might have gotten pushed back a bit until SPR and Genoa arrives later in 2022.
If PVC slips again I believe MI300 could launch before it

This is nuts, MI100/200/300 cadence is impressive.



Previous thread on CDNA2 and RDNA3 here

Question - Speculation: RDNA3 + CDNA2 Architectures Thread

Man I have been dying to make this one for a while now. First rumours for RDNA3 are here so new thread time! Just going to start off with this one for now: kopite7kimi on Twitter: "@VideoCardz Ah, I mean a simple mcm design with 10240 cores is not enough. Because the lift from RDNA2 to RDNA3...

forums.anandtech.com

 

[Win2012R2]

No. It's down to FSR4.

Should have shipped hardware in time for Xmas and promise to deliver wonderful FSR4 in 2025

 

[poke01]

LOL at recommending a 12GB VRAM 5070 in year 2025. It's only there to upsell people to the 5070 Ti.

Blackwell relies on DLSS4 and AI and is a rather sad improvement otherwise. Not to say AMD won't crap the bed though. They practically arlready have "vanished from the market".

Blackwell is a total skip gen if your on 40 series, especially with DLSS4 coming to older cards. I don’t care about Multi-frame

 

[SolidQ]

Should have shipped hardware in time for Xmas and promise to deliver wonderful FSR4 in 2025

gamers are very crazy on forums, they impatience. I'm also gamer, but i can wait easy, because mostly not playing anymore AAA games, just AA and indie

 

[Win2012R2]

Blackwell is a total skip gen if your on 40 series

I'll wait for reviews, find it hard to believe 5090 is just 25%-ish over 4090 (without fake frames obviously)

 

[adroc_thurston]

Should have shipped hardware in time for Xmas and promise to deliver wonderful FSR4 in 2025

No no, it gotta be a launch day thing or people like Alex cumraglia gonna torture radeons again.

 

[SolidQ]

find it hard to believe 5090 is just 25%-ish over 4090

would be worse even from 1080 ti to 2080ti. Even if it's 35% still bad. When RT 6xxx gonna announce huang will say dear ADA users time to upgrade, like it's was with AMPERE announce to Pascal users

Alex cumraglia gonna torture radeons again.

He already got orgasm from NV announcement

 

[Thunder 57]

Blackwell is a total skip gen if your on 40 series, especially with DLSS4 coming to older cards. I don’t care about Multi-frame

Not too surprising considering there is a minimal node advantage.

gamers are very crazy on forums, they impatience. I'm also gamer, but i can wait easy, because mostly not playing anymore AAA games, just AA and indie

Same. I mostly play older games these days. Newest one is probably 2022. I just don't find the latest ones all that interesting.

 

[Thunder 57]

I'll wait for reviews, find it hard to believe 5090 is just 25%-ish over 4090 (without fake frames obviously)

Look at the Far Cry numbers. They were the only ones to exclude DLSS. And they weren't great. I am certainly interested in reviews though especially regarding fake frames. I'm seeing some people say how they are great and the future now. What's the point of have 144hz if latency is still the same? It's just "bigger bar better". Reviewers may need to change how they review cards in the future and focus on frame time.

 

[Keller_TT]

let's wait official info. If it's like 5070ti, why they should sell it cheaper, than 5070?

Highly unlikely. AMD have been very late to the ML game on Radeon and NV can step on the gas gen on gen to the point that the trained models can do some incredible stuff.
If DLSS 4 can do one more step up from 3.5's RR and denoising to minimize artefacts and motion smear, the gap is only getting wider. So much hinges on FSR 4 being polished and at least as good as up to date DLSS 3.x.

 

[SolidQ]

Newest one is probably 2022. I just don't find the latest ones all that interesting.

I'm still play on Pascal 2024 year games, but they mostly Japan games. Even Alone in the Dark 2024 work fine on High setting.
Doom and new Residentevil gonna force me to Upgrade

 

[Gideon]

9070xt 14% faster than 7900xtx lol

View attachment 114456

Whoa, if this is true then this would be the true RV770 successor (granted AMD is also willing to not charge that much more than 7700 - 7800 XT)

Then again, if this is true i second and third this motion:

fire the person who stopped RDNA4 halo…

If the CU counts were even remotely true, this would have steamrolled the 5090 (and no, not in the AMD Bulldozer type of way)

 

[gdansk]

Look at the Far Cry numbers.

Too much drawn from too little. It's gonna be better than that. Nvidia launches do this over and over again. Wow amazing, oh that's fake crap, oh wait it's actually okay.
But since it isn't a big node shrink it's no Ada or Ampere.

 

[Gideon]

N31's GCD was around that size if I recall correctly, assuming you consider that a monolithic GPU like RDNA 4 has IO on it as well. If N48 can't even hit similar levels of performance, that would be just sad.

N32 in comparison is only 346 mm² in total, the GCD is 200 mm² If the rumored CU count is accurate, they really must have done something to beef them up considerably. But since the rumors were wrong about most of the other stuff as well, who knows ....

 

[Thunder 57]

I'm still play on Pascal 2024 year games, but they mostly Japan games. Even Alone in the Dark 2024 work fine on High setting.
Doom and new Residentevil gonna force me to Upgrade

Maybe not Doom. idTech has been incredibly forgiving. I;m not sure if that will interest me though as I wasn't happy with the changes from Doom 2015 to Doom Eternal. I never even finished The Ancient Gods part 2. It seems to try to make it force you to play "their way" ie using certain weapon mods to kill demons. And then the excessive platforming. Just my opinion though.

 

[Win2012R2]

Look at the Far Cry numbers.

I am trying hard to unlook them...

Reviewers may need to change how they review cards in the future and focus on frame time.

And stop getting samples of 6000 series? I think not!

 

[Thunder 57]

Too much drawn from too little. It's gonna be better than that. Nvidia launches do this over and over again. Wow amazing, oh that's fake crap, oh wait it's actually okay.
But since it isn't a big node shrink it's no Ada or Ampere.

Marketing tends to use the best numbers but if they are sandbagging I'm clearly OK with that. Better than overpromising and underdelivering.

 

[gdansk]

Whoa, if this is true then this would be the true RV770 successor

Too much drawn from too little.

If the CU counts were even remotely true, this would have steamrolled the 5090 (and no, not in the AMD Bulldozer type of way)

It would be insanely huge. Nearly 850mm^2 total if disaggregated? Who knows. On review day we'll see why it was a prudent decision.

 

[insertcarehere]

For reference comparison from a Chinese review website, it looks like the 'extreme' preset in COD 6 doesn't actually push every setting to the max, which may be causing some confusion here..

 

[SolidQ]

I never even finished The Ancient Gods part 2

I'm finished every new Doom on Nighmare difficult, and thank god I finish DLC before they nerf to ground it

from a Chinese review website,

Doesn't seem it's in game bench.
Where do you see 125 fps on 3ghz 4090?

 

[reaperrr3]

How many WGP do you think it needs for that die size?

WGP count is actually a poor metric for that.

In N10, the 20 WGP made up merely 80 out of 250mm2.
In N22, probably ~90 out of 330.
And that was on N7.

Basically, a modest WGP count doesn't guarantee a small die, and a big die doesn't guarantee a high WGP count.

RDNA4 WGPs are likely chunkier than RDNA3 WGPs, normalized to the same process.

Who knows, maybe RDNA4 truly doubles ALUs per CU and brings a really beefy IPC increase per WGP.

AMD also have to compete with 4070 Ti/TiS.

Don't know about other countries, but in Germany, the 4070Ti is all but extinct, while the 4070TiS is still over 830€ (roughly equivalent to 750$).

If the 9070XT brings ~79XT raster and 4070TiS RT perf, and FSR4 can at least close the gap to older DLSS revisions, at $550/€600 it would definitely sell.
9070 for $500/€550 would probably sell even better, certainly better than the 7700XT did.

 

[Thunder 57]

I'm finished every new Doom on Nighmare difficult, and thank god I finish DLC before they nerf to ground it


Doesn't seem it's in game bench.
Where do you see 125 fps on 3ghz 4090?

I have done Doom 2016 on Nightmare several times and I found it fun. Doom Eternal with the constant resource farming and platforming not so much. It was fun but just not my preference. Still good fun on UV though.

 

[Thunder 57]

WGP count is actually a poor metric for that.

In N10, the 20 WGP made up merely 80 out of 250mm2.
In N22, probably ~90 out of 330.
And that was on N7.

Basically, a modest WGP count doesn't guarantee a small die, and a big die doesn't guarantee a high WGP count.

RDNA4 WGPs are likely chunkier than RDNA3 WGPs, normalized to the same process.

Who knows, maybe RDNA4 truly doubles ALUs per CU and brings a really beefy IPC increase per WGP.


Don't know about other countries, but in Germany, the 4070Ti is all but extinct, while the 4070TiS is still over 830€ (roughly equivalent to 750$).

If the 9070XT brings ~79XT raster and 4070TiS RT perf, and FSR4 can at least close the gap to older DLSS revisions, at $550/€600 it would definitely sell.
9070 for $500/€550 would probably sell even better, certainly better than the 7700XT did.

Germany seems to be a more pro-AMD country. My theory is because of the fabs they built in Dresden.

 

[gdansk]

WGP count is actually a poor metric for that.

Unless radically different configuration, most structures scale with compute.

RDNA4 truly doubles ALUs per CU and brings a really beefy IPC increase per WGP.

No sign in open RDNA4 software patches thus far.

 

[insertcarehere]

Doesn't seem it's in game bench.
Where do you see 125 fps on 3ghz 4090?

They explicitly said in the review that it is.

Your vid is showing a custom preset (I am guessing max settings) which the extreme preset in COD 6 is not.

 

[poke01]

RDNA4 halo would have beaten 5090 but at the cost of a very big die size.