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

 

[branch_suggestion]

Okay, so initial thoughts of RDNA4/N48.
Firstly, N48 vexes me greatly, My first thought was on adding another SA to each SE, adding GDDR7 and that is a ~4090 right there at around 420mm^2.
The frontend is absolutely cracked, it could handle it no problem.
Lots of changes are iterations of RDNA3, but so much meaner. I would say the area investments are justified, especially considering what it can do at a large membw disadvantage, this is RDNA2 vs Ampere all over again.
AMD said their goal with each generation is to make the most of the games launched in the lifespan of the GPU, which is far more pertinent to nearly every consumer.
Soon however, the time will come when you must build something that will stand the test of time, and RDNA4 is a step towards that.
R300 lasted a fairly long time for what it was, GCN aged well over time thanks to consoles. In recent times this has been the struggle of RDNA, and they need to deliver something that can stand up to anything anybody throws at them.
Finally, I really want to see microbenchmarks and proper die analysis before saying more, I feel this is the kinda part that will age well but already perform well in established titles.
Lots of raw power for devs to tap into without relying on shader spam.
64CUs has come a very long way since Fiji and Vega, sure the xtor count has also gone a long way but diminishing returns etc.

revenge of the Larrabee (good ending).

As usual, Intel has good ideas but the board gets scared about leaving the cushy x86 money printer, so they always get canned before they are iterated enough to really show their potential.
IDC is a symptom of this management.

 

[adroc_thurston]

Lots of raw power for devs to tap into without relying on shader spam.

Yeah, it's the most bad code tolerant GPU design since Bifrost (and Bifrost sucked).

My first thought was on adding another SA to each SE, adding GDDR7 and that is a ~4090 right there at around 420mm^2.

Well one gotta keep the long-standing ATi tradition of undersizing configs in a really-really obnoxius way.

 

[coercitiv]

Oh wow I'M SAVED

Spoiler

Not a single model of the new GPUs will fit in my tiny 315mm length mATX case
yayyyy.............................................................................

RIP @Mahboi 's case, it won't make it 'till summer.

 

[Win2012R2]

I think the upcoming game AMD showed off looked good.

For sure, but reviewers will use CP2077 for next 10 years

 

[poke01]

For sure, but reviewers will use CP2077 for next 10 years

Oh they definitely will.

It’s funny, CD Projekt RED will port the game to Mac ( highly likely that Apple paid for it) and support its APIs but won’t add FSR3.1. If FSR3.1 doesn’t come when the Mac port launches, it’s definitely in their contract with Nvidia to never include the latest AMD upscaling.

 

[adroc_thurston]

Oh they definitely will.

It’s funny, CD Projekt RED will port the game to Mac ( highly likely that Apple paid for it) and support its APIs but won’t add FSR3.1. If FSR3.1 doesn’t come when the Mac port launches, it’s definitely in their contract with Nvidia to never include the latest AMD upscaling.

That's ok AMD has CoD for a pet shill game.

 

[itsmydamnation]

For sure, but reviewers will use CP2077 for next 10 years

Hopefully they just use mods to do fsr4.

I play cp2077 with fsr3.1 fg + xess on my 7900xtx. Optiscaler is so good you can custom tweak the src rez , target rez( super sampling) , sharpness , all on the fly as well . Really minimax quality to target frame rate.

It sux the 7900xtx can't do full rate int8. Hopefully it can do enough bf16/fp16 int 8 to get some version of fsr4.

 

[PJVol]

Also looking at RDNA2: 6900XT had 11% more CUs than 6800XT without any other changes: Same clocks, same TDP.

How do you think they managed to make 6900xt 8-9% faster if there were no other changes?
Binning is not just about disabling faulty and symmetrical CUs.

What you wrote basically sounds like we are at a point where it is acceptable for AMDs drivers to be bad at the beginning and it's great when they mature sometime in the future.

No, I didn't mean they're good or bad. To put it simple think of driver development as part of an ongoing GPU validation step.

 

[Win2012R2]

It sux the 7900xtx can't do full rate int8. Hopefully it can do enough bf16/fp16 int 8 to get some version of fsr4.

FSR4 appears to be using FP8

 

[itsmydamnation]

FSR4 appears to be using FP8

Yes I know but anything fp8 can be handled by fp/bf16 if the 7900 has enough raw throughput.

Then the question is , is there any advantage / not much loss using int8 when appropriate.(Int8 is 1/2 rate so same throughout of tops as fp16)

Then the final question is can we get acceptable performance with it.

 

[adroc_thurston]

can be handled by fp/bf16 if the 7900 has enough raw throughput.

yes but your math, bandwidth and mem footprint is 2x. stings.

 

[Dezii90]

Oh they definitely will.

It’s funny, CD Projekt RED will port the game to Mac ( highly likely that Apple paid for it) and support its APIs but won’t add FSR3.1. If FSR3.1 doesn’t come when the Mac port launches, it’s definitely in their contract with Nvidia to never include the latest AMD upscaling.

Why is amd sitting in their useless ass and not take Nvidia to court for anticompetitive practices.

 

[Darkmont]

What would you even sue for lol

 

[Dezii90]

anticompetitive practices...paying devs to block competitors products

 

[Win2012R2]

What would you even sue for lol

Abusing dominant market position

 

[Timorous]

So what are the chances N44 matches the 7700XT?

9070 has the same CU count as the 7700XT and looks to be 45% faster or there abouts in 1440p.

9060XT with 32CUs and a 45% uplift over the 7600XT would be in the same ballpark as 7700XT if not slightly ahead. Bandwidth will probably limit it a bit though but certainly looks as though it can be in the region of 7700XT.

To me this means AMD really do just need to make a 9060XT 16GB and they can probably sell it for $330 as a direct replacement for the 7600XT.

I suspect N44 is sub 200mm so with that in mind I would suspect yields are amazing and I wouldn't bother with a cut part. I wouldn't even bother with an 8GB variant if performance is 7700XT tier because at 1440p it will absolutely need more than 8GB and the performance would be there for that.

Compared to what I suspect will be a fake $450 MSRP for the 5060 Ti 16GB the 9060XT would hold up very well.

 

[itsmydamnation]

yes but your math, bandwidth and mem footprint is 2x. stings.

yeah probably only the 7900XT / XTX would have a chance. we just dont know how heavy the upscaler load is.

 

[MS_AT]

Hmm seems Charlie got upset https://www.semiaccurate.com/2025/02/28/amds-radeon-9070-isnt-very-good/

 

[Win2012R2]

Hmm seems Charlie got upset

He is never happy.

 

[reaperrr3]

So what are the chances N44 matches the 7700XT?

Good.

computerbase.de has the 7700XT at only ~35% faster than the 7600XT in FullHD, and in most cases that's gonna be the resolution the 9060XT will play games at (at least the internal resolution before upscaling).
35-45% faster than the 7600XT should be absolutely doable, judging by what we've seen so far.
The 5060Ti might actually struggle to beat the 9060XT in native raster.

 

[Joe NYC]

They had a late prod kickoff and no will to delay the launch any further. It happens.

There was a theory floated by Tech Potato Head, that NVidia may have trashed some datacenter Blackwell wafers, and to recover from that, NVidia used its gaming wafer allocation.

 

[Win2012R2]

Gaming Blackwell is rubbish regardless of its availability, even if prices were "MSRP" too - it's far worse than Turing because at the time even +30% were pushing 4K gaming to be viable, now with better upscaling that is even less of an issue, frankly I am more excited about RDNA5 now than 60 series and I had been Nvidia buyer since Riva128

 

[ryanjagtap]

Good.

computerbase.de has the 7700XT at only ~35% faster than the 7600XT in FullHD, and in most cases that's gonna be the resolution the 9060XT will play games at (at least the internal resolution before upscaling).
35-45% faster than the 7600XT should be absolutely doable, judging by what we've seen so far.
The 5060Ti might actually struggle to beat the 9060XT in native raster.

If that's the case then AMD should launch a 16GB clamshell Navi 44 alongside the 8GB one.
The 7700XT is a decent 1440p card and and 16GB Navi 44 could be a good product. (maybe name it 9060XTX or just name it 9060XT and put the 8GB model as 9060.... then the cut down version if any, can be 9050 series)

 

[Timorous]

If that's the case then AMD should launch a 16GB clamshell Navi 44 alongside the 8GB one.
The 7700XT is a decent 1440p card and and 16GB Navi 44 could be a good product. (maybe name it 9060XTX or just name it 9060XT and put the 8GB model as 9060.... then the cut down version if any, can be 9050 series)

N44 is tiny so I don't see the point in an 8GB or a cut version at all if it is 7700XT or higher performance, it will be received badly unless it is sub $250 and then the reduction in price Vs a 16GB variant will be greater than the ram cost.

They can keep pumping out N33 and drop the price of the 7600 if they want a low end product to replace the 6600 for the $200 crowd and that won't use N4 wafers either.

 

[poke01]

will games like Kingdom come 2 improve with RDNA4?

I'm looking forward to seeing some improvement