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

 

[ToTTenTranz]

Ok, just don't look up.

View attachment 117326

I see Nvidia's stock falling sharply after Deepseek's release, and then gradually recovered because they managed to pivot their wafers. It's still not on the same value as it was in mid January, though.



I'm not sure that's a subject for the RDNA4 thread, though. Nvidia producing more GB202 chips because there's less demand for B200 does relate to RDNA4's market performance, but calling out the stock market is a bit of a stretch.

 

[coercitiv]

but calling out the stock market is a bit of a stretch.

You're using a quote from someone called MEGAsizeGPU as evidence for Nvidia wafer allocation after someone else tried to gently guide your to more educated sources such as capex spending from hyperscalers. Of course I'm giving you the Nvidia stock screenshot, even gamers can guess that stock price cannot go back into the stratosphere based on consumer GPU sales.

But you're right with one thing, it would be great if we stopped invoking Deepseek as the basis for pure speculation about Nvidia supply in the AMD RDNA4 thread. Deal?

 

[gaav87]

Leaked 9070xt TSE 4K with beta drivers does not look that bad vs 5070ti/4080

 

[inquiss]

That is not what I'm seeing.

https://twitter.com/x/status/1890643714009121073

I mean yeah, you'd have to quote hyperscalers earning transcripts or calls or finance numbers rather than this random stuff.

 

[marees]

Assuming 9070xt matches RT of 4070 ti super then $600 would be a fair price for it (& $500 for 9070)

DLSS4 transformer based upscaling would be a point of differentiation for nvidia

 

[jpiniero]

Assuming 9070xt matches RT of 4070 ti super then $600 would be a fair price for it (& $500 for 9070)

Well the Real MSRP for the 5070 Ti is going to be $899.

 

[blckgrffn]

Well the Real MSRP for the 5070 Ti is going to be $899.

Yeah, I think we will see what the manage to put forth at $749. Of 9070xt MSRP is $699 and real prices are $100 or so more, seems like they’ll be in the right spot given the GPUs that are available now.

Isn’t it great that with a new gen we will probably be paying more for less? This is super duper!

Lmao, 5070ti is going to be $1k 😂. Microcenter listings are live for “coming soon”:

 

[MrTeal]

The recent Mindfactory sales data is useless for market share. They currently have 3 models of 4070 in stock for €600-€630, and not a single higher end Nvidia GPU. Meanwhile they have 23 models of AMD GPUs in stock at the 7800XT or above level, and you can buy a 7900GRE for €604.

 

[GodisanAtheist]

Ok, the following might be incredibly ignorant but:

If FSR4 is a drop in .DLL, is there a technical reason AMD cannot enable it on any game that supports DLSS by replacing/hijacking the DLSS .DLL with their own FSR4 version?

I'm guessing there is some logic in the game itself that says don't allow DLSS to be an option if AMD/Intel is detected, but could AMD find a way around that to immediately boost FSR4 adoption rates?

 

[jpiniero]

Ok, the following might be incredibly ignorant but:

If FSR4 is a drop in .DLL, is there a technical reason AMD cannot enable it on any game that supports DLSS by replacing/hijacking the DLSS .DLL with their own FSR4 version?

I'm guessing there is some logic in the game itself that says don't allow DLSS to be an option if AMD/Intel is detected, but could AMD find a way around that to immediately boost FSR4 adoption rates?

I assume what they mean is that it would be a drop in for a game that supports FSR3.

 

[gdansk]

I assume what they mean is that it would be a drop in for a game that supports FSR3.

He's aware. Sounds like the question is why AMD can't also make a shim DLL which has the same entry points that DLSS games expect.

 

[GodisanAtheist]

I assume what they mean is that it would be a drop in for a game that supports FSR3.

-I got that part, but FSR3 as far as I can tell isn't very broadly supported, certainly nowhere near the adoption rate of DLSS 2.0+.

Making FSR4 available on any DLSS 2.0+ game through a dll swap and some driver trickerywould be a major coup and get FSR4 adoption way up there pretty quickly.

 

[GTracing]

I'm guessing there is some logic in the game itself that says don't allow DLSS to be an option if AMD/Intel is detected

I think this would be the main issue. I doubt there's some "is_rtx" flag they could flip. They would need to basically create a mod for every single game.

Working with game developers to add native support is a better option all around imo.

 

[Heartbreaker]

Ok, the following might be incredibly ignorant but:

If FSR4 is a drop in .DLL, is there a technical reason AMD cannot enable it on any game that supports DLSS by replacing/hijacking the DLSS .DLL with their own FSR4 version?

I'm guessing there is some logic in the game itself that says don't allow DLSS to be an option if AMD/Intel is detected, but could AMD find a way around that to immediately boost FSR4 adoption rates?

It might be because AMD don't want to mod for something they have no control over. NVidia can change their interfaces, and break AMDs mod if they want to.

There was a community mod that did, DLSS to FSR, for FSR3, and I expect there will be a community mod to do it for FSR4.

 

[basix]

They could support the people who create the mods. OptiScaler for example is quite capable: https://github.com/cdozdil/OptiScaler

A driver based swap / mode would be ideal (user convenience) but I do not know if that is allowed (licenses, usage agreement terms or whatever).

 

[Jan Olšan]

I assume what they mean is that it would be a drop in for a game that supports FSR3.

No, FSR 3.1. That's important, 3.1 put in the API support for that. FSR 3.0 is not sufficient.

 

[poke01]

They could support the people who create the mods. OptiScaler for example is quite capable: https://github.com/cdozdil/OptiScaler

A driver based swap / mode would be ideal (user convenience) but I do not know if that is allowed (licenses, usage agreement terms or whatever).

Sadly AMD won’t support this.

 

[poke01]

this, update every dlss game needs to support FSR3.1. But some games would need community mods

 

[Tup3x]

Every game should add DirectSR support and this wouldn't be an issue.

 

[gaav87]

Jfc... Its like im an oracle when i said nitro+/red devil/taichi will be 900euro
I hope they will reduce this after 5070ti reviews i also heard from chinese 5070ti = almost 4080 in gaming... With good OC and power eff.

 

[Jan Olšan]

Leaked 9070xt TSE 4K with beta drivers does not look that bad vs 5070ti/4080

View attachment 117337

This is month old.
There were no performance leaks since,I take it? Besides the Call of Fury which was hard to compare and unclear settings.

Who knows if we can count on that, I hope it's not something way lower in the end.

 

[blckgrffn]

Anecdotal sales data - this morning my local MC had 10 Asus 7800XT in stock at $499. All of them are gone now when I just checked.

Seems like the $500-$1k segment is complete starved right now and anything that hits the channel is gone ASAP.

If AMD can land with stock it seems like they’ll get some wins right now because they are the ones showing up to play. It will be interesting to see what volume nvidia brings to bear of the vanilla 5070 because that seems to be the one they would be selling in massive numbers right now if it existed in retail.

 

[gaav87]

Anecdotal sales data - this morning my local MC had 10 Asus 7800XT in stock at $499. All of them are gone now when I just checked.

Seems like the $500-$1k segment is complete starved right now and anything that hits the channel is gone ASAP.

If AMD can land with stock it seems like they’ll get some wins right now because they are the ones showing up to play. It will be interesting to see what volume nvidia brings to bear of the vanilla 5070 because that seems to be the one they would be selling in massive numbers right now if it existed in retail.

I heard whole 5090 5080 5070ti supply = 2x 9070 so not that great either...

 

[ToTTenTranz]

You're using a quote from someone called MEGAsizeGPU as evidence for Nvidia wafer allocation after someone else tried to gently guide your to more educated sources such as capex spending from hyperscalers.

I pointed to a leaker with a pretty great track record on Nvidia AICs which is why the news has been spreading pretty wide.
And at least I pointed to something, whereas these educated sources existed only in hearsay in this thread so far.


Regardless, hyperscalers are half the market for data center GPUs.

 

[gaav87]

If amd think they can sell 9070xt for 480pln more than 7900xtx with -8gb vram they are mad.