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

 

[adroc_thurston]

Maybe cost wouldn't justify the outcome?

More like they didn't plan on Blackwell to suck.

Not that it's strong representative, but take note of XT -> XT-OC, the scaling still holds while the output is 4x more pixels.

It very clearly falls off in a bunch of games for obvious reasons.

 

[Mahboi]

Can they make that clamshell 32Go VRAM 9070 XT have GDDR7?

 

[adroc_thurston]

Can they make that clamshell 32Go VRAM 9070 XT have GDDR7?

32G is too much for a mainstream product.

 

[Hail The Brain Slug]

Can they make that clamshell 32Go VRAM 9070 XT have GDDR7?

Going from 20gbps gddr6 to 28gbps gddr7 is already going to increase BOM quite a lot. Doubling the amount is going to make it crazy expensive.

 

[Mahboi]

Going from 20gbps gddr6 to 28gbps gddr7 is already going to increase BOM quite a lot. Doubling the amount is going to make it crazy expensive.

That's the point. Try and salvage a semi-halo product.

 

[Hail The Brain Slug]

That's the point. Try and salvage a semi-halo product.

Would the market accept a $1200 9070XTX 32GB if it was on par with a 5080?

 

[gdansk]

Would the market accept a $1200 9070XTX 32GB if it was on par with a 5080?

No. More than 16GB VRAM is of little utility for gamers. And RNDA4 is still a gamer card.

 

[exquisitechar]

That's the point. Try and salvage a semi-halo product.

No gamer would buy that.

 

[Mahboi]

Ok first it's not doubling the price just for 32Go of GDDR7.
And adding $200 to make it a $800 base price with GDDR7 is definitely doable. Probably even $300 to make it $900. No clue how difficult it would be to implement GDDR7 for them.

 

[adroc_thurston]

More than 16GB VRAM is of little utility for gamers

Eh there are cases for 24G somewhere.

 

[PJVol]

You're missing one data point in that chart, one where the 9070 XT is only using mem oc.

Maybe I am, can you point me out where?

 

[Hail The Brain Slug]

Ok first it's not doubling the price just for 32Go of GDDR7.
And adding $200 to make it a $800 base price with GDDR7 is definitely doable. Probably even $300 to make it $900. No clue how difficult it would be to implement GDDR7 for them.

Yeah, that would be to simply swap gddr6 for gddr7. Add +16GB more GDDR7 and you're at $1200

 

[adroc_thurston]

Maybe I am, can you point me out where?

Says 22.5GT/s right there.
That's an 11.25% memory overclock.

 

[blckgrffn]

Eh there are cases for 24G somewhere.

24GB 9070 XTX refresh seems like... it's right there. If they aren't working towards it, they should.

If they had kept the 9700 nomenclature from previous gens, a refresh of all the SKUs with DDR7 would have been an easy 9750, 9750XT and 9750XTX (24GB when available) but now its weird. Oh well.

 

[PJVol]

Says 22.5GT/s right there.

Yes, Gfx is also oc'ed there. I thought that he literally meant only mem OC, didn't he?

one where the 9070 XT is only using mem oc.

 

[MangoX]

Very nice cards and performance. But still price too high. Was hoping for a Ryzen moment, where they could've seriously undercut nvidia and flood and reset the market. Like they did with the 1800x 8c for $499 compared to Intel's 8c for $999.

 

[coercitiv]

We needed Nvidia to deliver to bring us better AMD prices. They let us down.

This, I don't know how else to say it more clearly, but we need lower priced Nvidia cards so we can buy AMD.

I'll see myself out.

 

[linkgoron]

24GB 9070 XTX refresh seems like... it's right there. If they aren't working towards it, they should.

If they had kept the 9700 nomenclature from previous gens, a refresh of all the SKUs with DDR7 would have been an easy 9750, 9750XT and 9750XTX (24GB when available) but now its weird. Oh well.

Clearly, 9070XT Super.

 

[GaiaHunter]

If the problem is the memory shouldn't it be worse at 4K?

 

[Timorous]

24GB 9070 XTX refresh seems like... it's right there. If they aren't working towards it, they should.

If they had kept the 9700 nomenclature from previous gens, a refresh of all the SKUs with DDR7 would have been an easy 9750, 9750XT and 9750XTX (24GB when available) but now its weird. Oh well.

9070 XT Super.

9070 XT 4Ghz Edition.

 

[coercitiv]

I thought that he literally meant only mem OC, didn't he?

I did, and my point is that the graph you were using is not sufficient evidence. Whether you're right or wrong, I can only guess. We need to isolate one metric, either by scaling mem bw or scaling clocks.

 

[adroc_thurston]

Yes, Gfx is also oc'ed there. I thought that he literally meant only mem OC, didn't he?

Well memOC is the relevant thing there.

Clearly, 9070XT Super.

Let's keep it ATi-flavored.
9070XT RAGE.

 

[blckgrffn]

Well memOC is the relevant thing there.

Let's keep it ATi-flavored.
9070XT RAGE.

9070XT FURY

Not to be confused with the 9070XT FURRY

 

[Timorous]

Let's keep it ATi-flavored.
9070XT RAGE.

Might as well go all the way.

9070 XT Rage Fury Maxx AI + Pro

 

[coercitiv]

If the problem is the memory shouldn't it be worse at 4K?

This is what's surprising for me as well, it's an atypical behavior.