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

SolidQ · Mar 3, 2025

carrotmania said:
magine a full fat 128CU 9090XT

it's was 144CU?

Hail The Brain Slug · Mar 3, 2025

Josh128 said:
The only thing we can be fairly sure of regarding 7900XTX is that they missed their published frequency target (~3GHZ) for the RDNA3 architecture by a good margin. It released with a boost clock of 2.5GHz, 20% below 3GHZ. RDNA 4 has achieved 3GHz+. Whether that was related to MCM vs monolithic design or something else, is unclear.

If we assume the provided performance figures during the announcement weren't intended to be egregiously misleading, I dont know if the clock frequency difference is quite enough to explain it. Seems like there's something else holding it back.

But since the rumors of a "fixed N31 refresh" never came true I dont think we will ever know for sure.

HuesToo4 · Mar 3, 2025

SolidQ said:
it's was 144CU?

Regardless, Navi4c would have crushed 5090 if they didn't cancel it.

SolidQ · Mar 3, 2025

Unlimited Geometric Detail = RTX Mega Geometry?

carrotmania · Mar 3, 2025

SolidQ said:
it's was 144CU?

Awww.
If perfect scaling, that would have been 144 FPS in CP2077, vs 5090s 108
ie, as much over the 5090 as the 5090 is over the 4090

... and 399FPS in DOOM...

MrTeal · Mar 3, 2025

carrotmania said:
VERY rough numbers, but a hypothetical 96CU 9080XT could looked like this:

View attachment 118743 View attachment 118744 View attachment 118745

Imagine a full fat 128CU 9090XT

That seems wildly optimistic. It seems like you're just going off 9070 XT estimates +50%, but it's almost certainly going to be way less. The 5090 is basically double a 5080, and in those games:
CP2077 5090 108.8 5080 70.9 (+53.5%)
BG3 5090 177.4 5080 119.0 (+49%)
DOOM 5090 249.2 5080 178.3 (+39.8%)

The 5080 is faster in all those games than a 7900 XTX (and likely 9070 XT), so expecting a 96CU RDNA4 part to outperform a 5090 would require some crazy perfect scaling magic.

Hail The Brain Slug · Mar 3, 2025

MrTeal said:
That seems wildly optimistic. It seems like you're just going off 9070 XT estimates +50%, but it's almost certainly going to be way less. The 5090 is basically double a 5080, and in those games:
CP2077 5090 108.8 5080 70.9 (+53.5%)
BG3 5090 177.4 5080 119.0 (+49%)
DOOM 5090 249.2 5080 178.3 (+39.8%)

The 5080 is faster in all those games than a 7900 XTX (and likely 9070 XT), so expecting a 96CU RDNA4 part to outperform a 5090 would require some crazy perfect scaling magic.

Just looking at rough figures, AMD's designs have tended to scale better than nvidia. I would expect better scaling than 5090 vs. 5080, albeit nowhere near 100%.

carrotmania · Mar 3, 2025

MrTeal said:
That seems wildly optimistic. It seems like you're just going off 9070 XT estimates +50%, but it's almost certainly going to be way less. The 5090 is basically double a 5080, and in those games:
CP2077 5090 108.8 5080 70.9 (+53.5%)
BG3 5090 177.4 5080 119.0 (+49%)
DOOM 5090 249.2 5080 178.3 (+39.8%)

The 5080 is faster in all those games than a 7900 XTX (and likely 9070 XT), so expecting a 96CU RDNA4 part to outperform a 5090 would require some crazy perfect scaling magic.

I did not.
I looked up 7900GRE numbers for them, and added the % uplift AMD gave in those games from their presentation, or a best guess (around 40%, but not bang on). Then, yes, I added 50% on that for a scaling of CUs. I said it was rough, but some thought went in to it.

Also, I went for 50% CU scaling, in the hope (for a non existent part) that it would employ GDDR7, so the greater bandwidth would also help offset other losses.

PJVol · Mar 3, 2025

carrotmania said:
Also, I went for 50% CU scaling,

You'll need +50% power for that. In practice, taking CU scaling, EDC and other factors into account you'll end up with +35-40% at best and @ 460W.

SolidQ · Mar 3, 2025

PJVol said:
You'll need +50% power for that. In practice, taking CU scaling, EDC and other factors into account you'll end up with +35-40% at best and @ 460W.

There based on TPU reviews, difference between XTX and 4090 in 4k is 29%. Should be around same for 9070XT. So for AMD would be best scenario make ~4090 perf for 1k with 24GB, and close RDNA4 line up

MrTeal · Mar 3, 2025

carrotmania said:
I did not.
I looked up 7900GRE numbers for them, and added the % uplift AMD gave in those games from their presentation, or a best guess (around 40%, but not bang on). Then, yes, I added 50% on that for a scaling of CUs. I said it was rough, but some thought went in to it.

Also, I went for 50% CU scaling, in the hope (for a non existent part) that it would employ GDDR7, so the greater bandwidth would also help offset other losses.

It kinda seems like you did? Isn't 7900GRE * 1.4 the 9700XT estimate and then 50% on top of that?

gdansk · Mar 3, 2025

There's no point extrapolating over discarded paper.
We simply don't know how RDNA4 scales.

SolidQ · Mar 3, 2025

gdansk said:
We simply don't know how RDNA4 scales.

for that we need N44

carrotmania · Mar 3, 2025

PJVol said:
You'll need +50% power for that. In practice, taking CU scaling, EDC and other factors into account you'll end up with +35-40% at best and @ 460W.

No one batted an eyelid when NV blew the "agreed" power budget with the 5090, so I didn't bother with bothering about that.

PJVol · Mar 3, 2025

SolidQ said:
There based on TPU reviews, difference between XTX and 4090 in 4k is 29%.

Yeah, that's what I (ok, we) ended up with while speculating just two pages ago

Page 368 - Discussion - RDNA4 + CDNA3 Architectures Thread

Page 368 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

carrotmania · Mar 3, 2025

MrTeal said:
It kinda seems like you did? Isn't 7900GRE * 1.4 the 9700XT estimate and then 50% on top of that?

OK, so TPU numbers for the 7900GRE are old, I looked at them, and the most recent benchies, which both included XTXs but not GREs, and I added the small increase in % in that time. Then I added the right % uplift, ie AMD claims 48% not 40%. So, again, I said VERY ROUGH, but I didnt "just add 50%" as you claimed.

PJVol · Mar 3, 2025

carrotmania said:
so I didn't bother with bothering about that.

Congrats! But I do, as does a lot of people I know.

carrotmania · Mar 3, 2025

PJVol said:
Congrats! But I do, as does a lot of people I know.

Then a 5090 isn't for you either, and it doesn't matter what a top end AMD card could put out. Reductio ad absurdum.
But that would still be a 450W card, knocking a 4090 around, also at 450W, and sometimes beating a 575W 5090. Don't sand bag.

Josh128 · Mar 3, 2025

carrotmania said:
OK, so TPU numbers for the 7900GRE are old, I looked at them, and the most recent benchies, which both included XTXs but not GREs, and I added the small increase in % in that time. Then I added the right % uplift, ie AMD claims 48% not 40%. So, again, I said VERY ROUGH, but I didnt "just add 50%" as you claimed.

AMD claims +42%, not +48%, at 4K mixed raster/RT, over 7900GRE. The real world numbers are almost certainly going to be a bit less due to the "first party benchmarks" tax.

PJVol · Mar 3, 2025

carrotmania said:
Then a 5090 isn't for you either, and it doesn't matter what a top end AMD card could put out.

A remarkable foresight of yours; and no, it does matter.

carrotmania · Mar 3, 2025

Josh128 said:
AMD claims +42%, not +48%, at 4K mixed raster/RT, over 7900GRE. The real world numbers are almost certainly going to be a bit less due to the "first party benchmarks" tax.

View attachment 118750

Bah, I missed the end of that sentence, my 5yr old was jumping on me.
I'll re-write it:

"Then I added the right % uplift for the right game, ie AMD claims 48% not 40% (PJ said I just added 1.4x) for CP2077."

Taken from:

carrotmania · Mar 3, 2025

PJVol said:
A remarkable foresight of yours; and no, it does matter.

Bad day at work? You're being remarkably p%(*&y for a quick and dirty "point of interest".
But nice to know you think its OK for NV to make a 450W card, and then a 575W card, but AMD absolutely HAS to match or beat them at 300W. Faaaantasy laaaand.

SolidQ · Mar 3, 2025

Some real fps you can get on AMD sites
Here: https://www.amd.com/en/products/graphics/desktops/radeon/9000-series/amd-radeon-rx-9070xt.html
Also there RT numbers on another page

Josh128 · Mar 3, 2025

carrotmania said:
Bah, I missed the end of that sentence, my 5yr old was jumping on me.
I'll re-write it:

"Then I added the right % uplift for the right game, ie AMD claims 48% not 40% (PJ said I just added 1.4x) for CP2077."

Taken from:

View attachment 118751

Interesting. The average of these figures, which dont appear to include RT results, is +37% raster at 4K.

carrotmania · Mar 3, 2025

Josh128 said:
Interesting. The average of these figures, which dont appear to include RT results, is +37% raster at 4K.

Its a subset. The full image says that they worked out the 42% average from 30+ games, far more than shown on the marketing.

Discussion RDNA4 + CDNA3 Architectures Thread

Golden Member

Golden Member

Diamond Member

Junior Member

Golden Member

Member

Diamond Member

Diamond Member

Member

Senior member

Golden Member

Diamond Member

Diamond Member

Golden Member

Member

Senior member

Member

Senior member

Member

Senior member

Senior member

Member

Member

Golden Member

Senior member

Member