Discussion RDNA4 + CDNA3 Architectures Thread

DisEnchantment · Mar 23, 2022

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 · Nov 23, 2023

jpiniero said:
Depends on your definition of midrange these days, lol.

And not much has leaked on Blackwell. So while I am very pessimistic about the gen-on-gen perf improvement outside of the $$$ stuff it is too early to be definitive about it.

Competitive can mean many things, FWIW, I certainly think RDNA4 will be cheaper to make for a given performance tier.
N4P vs N3E makes that straightforward. It coming out first is the biggest advantage overall.

adroc_thurston · Nov 23, 2023

branch_suggestion said:
N4P vs N3E makes that straightforward

Nothing stops NV from making mainstream GB20x on N4p.

branch_suggestion · Nov 24, 2023

adroc_thurston said:
Nothing stops NV from making mainstream GB20x on N4p.

Very true, but NV hasn't bothered dual sourcing nodes for a single client uarch gen since Pascal, and those were 2 very similar nodes. They are usually happy to mix the previous gen to balance it out. Then again the 700 and 600 series did do this to a large degree.
Considering DC demand for N3E, you are probably right.
RDNA4 should still hold the performance/area edge on similar nodes, so I stand by my prediction.

beginner99 · Nov 24, 2023

Timorous said:
This is exactly what I don't get.

We have the canned super complex multi compute part at the halo end. We have the small monolithic dies at the low and middle end. What happened to the upper middle and high end parts? Why were they canned?

I assume everything canned was chiplet-based. This includes a gpu with just one "compute" chiplet or however the organization would be, just the minimal config. If the can't get this to work at all, all products using the tech from minimal config to monster gpu are canned.
And even if the "single chiplet" version would work, probably not worth it to invest in the "base cost" just for 1 or 2 SKUs.

EDIT: the not working part could also entirely be due to driver/software. If the driver teams says: we are not sure if we can get it working till then, then they have no other option as to delays or can the products.

adroc_thurston · Nov 24, 2023

beginner99 said:
the not working part could also entirely be due to driver/software

It wasn't about the software.

beginner99 said:
If the driver teams says: we are not sure if we can get it working till then

It's not about that.
It's about the possibility of slips.

igor_kavinski · Nov 24, 2023

adroc_thurston said:
It's about the possibility of slips.

So they DID learn something from RDNA3

Ajay · Nov 24, 2023

adroc_thurston said:
Nothing stops NV from making mainstream GB20x on N4p.

Seems mainly like a cost vs power efficiency question for NV. I suspect they'll go with power efficiency, but who really knows at this point.

adroc_thurston · Nov 24, 2023

igor_kavinski said:
So they DID learn something from RDNA3

?
RDNA3 was an oopsie. Hardly related to any schedules.

Ajay said:
Seems mainly like a cost vs power efficiency question for NV.

Cost is king.

igor_kavinski · Nov 24, 2023

adroc_thurston said:
?
RDNA3 was an oopsie. Hardly related to any schedules.

RDNA4 high end not canned coz of RDNA3's stumble?

adroc_thurston · Nov 24, 2023

igor_kavinski said:
RDNA4 high end not canned coz of RDNA3's stumble?

Not in any way, shape or form related.

igor_kavinski · Nov 24, 2023

adroc_thurston said:
Not in any way, shape or form related.

Please say that we are getting an RDNA3 high end refresh. 7950XTX. Please. I beseech you!

adroc_thurston · Nov 24, 2023

igor_kavinski said:
Please say that we are getting an RDNA3 high end refresh

why.

igor_kavinski said:
7950XTX. Please. I beseech you!

why would you need that when the bigger of the N4M's is close enough?

igor_kavinski · Nov 24, 2023

adroc_thurston said:
why would you need that when the bigger of the N4M's is close enough?

It's not gonna be faster than N31, right? Need something that 4080 can't touch.

adroc_thurston · Nov 24, 2023

igor_kavinski said:
It's not gonna be faster than N31, right?

No.

igor_kavinski said:
Need something that 4080 can't touch.

Arbitrary stuff is arbitrary, not how product stacks work.

SteinFG · Nov 26, 2023

igor_kavinski said:
Please say that we are getting an RDNA3 high end refresh. 7950XTX. Please. I beseech you!

Vf cuve is still shit without re-making the die. 3D-stacking MCD for a part that barely beats 4080 isn't worth it. They can put faster vram there, and nothing else really. Doubt that'll help though, as its bandwidth is already reaching 4090 levels.

igor_kavinski · Nov 26, 2023

SteinFG said:
Vf cuve is still shit without re-making the die.

I'm sure there are some engineering tweaks they could do. Or maybe the node characteristics may have improved over time to let them get less leaky dies. Or even they might have been collecting the golden dies for a 7950XTX launch later that not only clock higher but also leak less, consuming less power and getting less hot.

If RDNA4 isn't gonna target the high end, they would want something at the top still, to make the extra margins on the high end. If the top RDNA4 part gets something like 85% of 7900XTX performance, they would want to increase that difference to at least 25% higher performing halo part, to get the gamers with deeper pockets to go for 25% more performance for maybe like 35% more money.

PJVol · Nov 26, 2023

SteinFG said:
Vf cuve is still shit without re-making the die.

There's actually nothing wrong with the graphics core of the rdna3. The issue is with the uncore power consumption.

PJVol · Nov 26, 2023

Btw, gfx1200 and gfx1201 are listed under RDNA3 arch in llvm processors )

User Guide for AMDGPU Backend — LLVM 19.0.0git documentation

igor_kavinski · Nov 26, 2023

PJVol said:
Btw, gfx1200 and gfx1201 are listed under RDNA3 arch in llvm processors )

So it IS very possible that 7950 XT and XTX are inbound!

linkgoron · Nov 26, 2023

PJVol said:
Btw, gfx1200 and gfx1201 are listed under RDNA3 arch in llvm processors )

User Guide for AMDGPU Backend — LLVM 19.0.0git documentation

https://web.archive.org/web/20200919132836/https://llvm.org/docs/AMDGPUUsage.html#processors

RDNA2 was also listed in the same group as RDNA1 before they split it up later.

Tigerick · Nov 26, 2023

PJVol said:
Btw, gfx1200 and gfx1201 are listed under RDNA3 arch in llvm processors )

User Guide for AMDGPU Backend — LLVM 19.0.0git documentation

Wow, nice findings. I think it is pretty much confirmed RDNA4 GPUs are based on RDNA3 architecture even though they use gfx1200 series. As I said, N43 and N44 are just a die shrink of RDNA3, there won't be much change in core design and we should hear more about 7600XT and 7700 in CES2024....

PJVol · Nov 26, 2023

linkgoron said:
RDNA2 was also listed in the same group as RDNA1 before they split it up later.

Very well may be. There's one more though, where rdna 1 & 2 aren't split, and 3 & 4 described under the same MachineModel scheduling definition either:

Code:

//===----------------------------------------------------------------------===//
// GCN GFX10.
//===----------------------------------------------------------------------===//

def : ProcessorModel<"gfx1010", GFX10SpeedModel,
  FeatureISAVersion10_1_0.Features
>;

def : ProcessorModel<"gfx1011", GFX10SpeedModel,
  FeatureISAVersion10_1_1.Features
>;

def : ProcessorModel<"gfx1012", GFX10SpeedModel,
  FeatureISAVersion10_1_2.Features
>;

def : ProcessorModel<"gfx1013", GFX10SpeedModel,
  FeatureISAVersion10_1_3.Features
>;

def : ProcessorModel<"gfx1030", GFX10SpeedModel,
  FeatureISAVersion10_3_0.Features
>;

def : ProcessorModel<"gfx1031", GFX10SpeedModel,
  FeatureISAVersion10_3_0.Features
>;

def : ProcessorModel<"gfx1032", GFX10SpeedModel,
  FeatureISAVersion10_3_0.Features
>;

def : ProcessorModel<"gfx1033", GFX10SpeedModel,
  FeatureISAVersion10_3_0.Features
>;

def : ProcessorModel<"gfx1034", GFX10SpeedModel,
  FeatureISAVersion10_3_0.Features
>;

def : ProcessorModel<"gfx1035", GFX10SpeedModel,
  FeatureISAVersion10_3_0.Features
>;

def : ProcessorModel<"gfx1036", GFX10SpeedModel,
  FeatureISAVersion10_3_0.Features
>;

//===----------------------------------------------------------------------===//
// GCN GFX11.
//===----------------------------------------------------------------------===//

def : ProcessorModel<"gfx1100", GFX11SpeedModel,
  FeatureISAVersion11_0_0.Features
>;

def : ProcessorModel<"gfx1101", GFX11SpeedModel,
  FeatureISAVersion11_0_1.Features
>;

def : ProcessorModel<"gfx1102", GFX11SpeedModel,
  FeatureISAVersion11_0_2.Features
>;

def : ProcessorModel<"gfx1103", GFX11SpeedModel,
  FeatureISAVersion11_0_3.Features
>;

def : ProcessorModel<"gfx1150", GFX11SpeedModel,
  FeatureISAVersion11_5_0.Features
>;

def : ProcessorModel<"gfx1151", GFX11SpeedModel,
  FeatureISAVersion11_5_1.Features
>;

//===----------------------------------------------------------------------===//
// GCN GFX12.
//===----------------------------------------------------------------------===//

def : ProcessorModel<"gfx1200", GFX11SpeedModel,
  FeatureISAVersion12.Features
>;

def : ProcessorModel<"gfx1201", GFX11SpeedModel,
  FeatureISAVersion12.Features
>;

adroc_thurston · Nov 26, 2023

PJVol said:
Btw, gfx1200 and gfx1201 are listed under RDNA3 arch in llvm processors )

That's a stub.

Tigerick said:
I think it is pretty much confirmed RDNA4 GPUs are based on RDNA3 architecture

They're nothing alike.

Tigerick said:
As I said, N43 and N44 are just a die shrink of RDNA3

They're not even called that.

igor_kavinski said:
So it IS very possible that 7950 XT and XTX are inbound!

They don't exist, much like X3D N31 does not exist.

PJVol · Nov 26, 2023

adroc_thurston said:
They're nothing alike.

"Nothing alike" in relation to ISA seems like an exaggeration, doesn't it?

adroc_thurston · Nov 26, 2023

PJVol said:
"Nothing alike" in relation to ISA seems like an exaggeration, doesn't it?

The hardware.
AMD hasn't altered their base ISA much since GCN1 aka 2011.

Discussion RDNA4 + CDNA3 Architectures Thread

Golden Member

Senior member

Diamond Member

Senior member

Diamond Member

Diamond Member

Lifer

Lifer

Diamond Member

Lifer

Diamond Member

Lifer

Diamond Member

Lifer

Diamond Member

Senior member

Lifer

Senior member

Senior member

Lifer

Platinum Member

Senior member

Senior member

Diamond Member

Senior member

Diamond Member