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

Timorous · Jan 18, 2025

SolidQ said:
That different story AMD RDNA3 was like "UP x 1.7" no numbers
Here is FPS numbers and remember it's leak, not official announce from AMD

Numbers are kind of irrelevant. Up 1.7x is good enough because the numbers vary scene to scene and game to game. TPU, HUB, GN, CB and PCGH all get different numbers in the same games because their test scenarios are all different.

adroc_thurston · Jan 18, 2025

Timorous said:
Numbers are kind of irrelevant

What the hell.
They're numbers. For different GPUs ISO rest of the system.

Timorous said:
TPU, HUB, GN, CB and PCGH all get different numbers in the same games because their test scenarios are all different.

The test bench config is literally in the slide title and they use built-in benchmarks for games.

SolidQ · Jan 18, 2025

adroc_thurston said:
The test bench config is literally in the slide title and they use built-in benchmarks for games.

but interesting they dodn't test path tracing...

It loses to N31 in 2 bandwidth-bound titles anyway.

i'd assuming hypothetical GDDR7 version can fix it

jpiniero · Jan 18, 2025

adroc_thurston said:
It's not 390mm^2.

But it's a lot closer to GB203 than GB205, no?

Theoretically going low enough that nVidia responds with price cuts doesn't exactly help AMD.

adroc_thurston · Jan 18, 2025

SolidQ said:
but interesting they dodn't test path tracing...

Meme.

SolidQ said:
i'd assuming hypothetical GDDR7 version can fix it

More like help it.

jpiniero said:
But it's a lot closer to GB203 than GB205, no?

In both area and perf, indeed.

jpiniero said:
Theoretically going low enough that nVidia responds with price cuts doesn't exactly help AMD.

AMD can slap GDDR7 to equate their BOM while also leveling the perf delta versus GB203.

You can pricecut Intel to death since they have atrocious PPA. Ain't gonna fly against N4m.

SolidQ · Jan 18, 2025

adroc_thurston said:
Meme.

People will compare in C2077/AW2
Interesting where latest games, why old horizon, if new horizon love AMD, also UE5 games...

adroc_thurston · Jan 18, 2025

SolidQ said:
Interesting where latest games, why old horizon, if new horizon love AMD, also UE5 games...

They picked games from N31/AD103 launch window.
(Or just whatever they had on the testbench. Again, you weren't supposed to see them charts).

DownTheSky · Jan 18, 2025

Is BM:W RT HIGH path tracing? There is no RT HIGH in benchmark settings. Only very high.

Timorous · Jan 18, 2025

adroc_thurston said:
This cope never works when you have concrete FPS numbers 2 pages down.

Nothing to do with cope. Everything to do with 64CUs @ 3Ghz with minimal bandwidth gains vs the 7800XT means without a massive perf per CU uplift the gains will be in the 15-20% region give or take and even with a perf/CU uplift how effective will it be when bandwidth starved? That was the one thing that really held the 7900GRE back.

adroc_thurston said:
What the hell.
They're numbers. For different GPUs ISO rest of the system.

The test bench config is literally in the slide title and they use built-in benchmarks for games.

Great, relative performance is what matters though. always has and the absolute FPS numbers are just a means to that end. This also assumes those numbers are not a fabrication.

If the 9070XT ends up that close to the 7900XTX (geomean has the 7900XTX around 7.5% ahead) then fantastic. AMD did with RDNA 4 what they should have done with RDNA 3 and it make scrapping the halo RDNA 4 parts look entirely stupid and RV770 all over again, which is just Radeon division all over tbh.

adroc_thurston · Jan 18, 2025

Timorous said:
Everything to do with 64CUs @ 3Ghz with minimal bandwidth gains vs the 7800XT means without a massive perf per CU uplift the gains will be in the 15-20% region give or take and even with a perf/CU uplift how effective will it be when bandwidth starved?

Caches are truly magical things and AMD is the most devout worshipper of them.
But they can only go so far, which is why N31 krumps it in two titles that use rather antique inhouse engines.

Timorous said:
relative performance is what matters though

But you get relative perf relative to both N31 and AD103 chops.

Timorous said:
AMD did with RDNA 4 what they should have done with RDNA 3 and it make scrapping the halo RDNA 4 parts look entirely stupid and RV770 all over again, which is just Radeon division all over tbh.

DING-DING-DING-DING-DING.
Right on money, we're back to the ATi hood classic of "amazing IP, zero noteworthy configs shipped".

SolidQ · Jan 18, 2025

DownTheSky said:
Is BM:W RT HIGH path tracing? There is no RT HIGH in benchmark settings. Only very high.

isn't there only like on/off RT? So that mean they test it in PAth Tracing
Looks like ~4080 in wukong

AMD can slap GDDR7 to equate their BOM while also leveling the perf delta versus GB203.

Maybe than can slap 3GB modules, make 24GB and sell it to gamer and AI guys. Vram is matter in AI, sohuld be win over 5080

geomean has the 7900XTX around 7.5% ahead

If they add 5% with final derivr should be 2.5%

gaav87 · Jan 18, 2025

DownTheSky · Jan 18, 2025

@SolidQ That's great news if true.

@gaav87 That's not showing the whole story. Navi 48 looks severely memory bw starved. Hence some games = 7900XTX some lags by 20%

SolidQ · Jan 18, 2025

DownTheSky said:
That's great news if true.

Also found result on 4080 in C2077 4k RT ultra

DownTheSky · Jan 18, 2025

So 1 fps higher than 9070XT.

marees · Jan 18, 2025

del42sa said:
https://videocardz.com/newz/asus-co...n-greek-retailer-list-january-23-availability

23th January ? withouth official launch and specs ? weird ...

I don't see jan-23rd in the source.

What did I miss ?

eek2121 · Jan 18, 2025

adroc_thurston said:
ding-ding-ding, with Blackwell being what it is (poo), AMD can price whatever. $549 is the optimistic pick.

$479 was for the MBA cards, AIB cards will be around $549. Note that that price was prior to the NV 50 series launch. They absolutely could (have) raise(d) it, but if they are smart, they won’t.

Josh128 · Jan 18, 2025

AMD shadow marketing doing its thing to whip up hype and launch pricing. Who doesnt realize this? They have the ball, clear shot to the end zone, only question is at what point do they fumble (prices) it??

Shadow marketing refers to methods that appeal to consumers' subconscious minds and convey a brand's message in ways that are not overtly noticeable. Unlike eye-catching advertisements or direct sales messages, these techniques aim to subtly influence consumers' perceptions and decisions.

Abwx · Jan 18, 2025

eek2121 said:
$479 was for the MBA cards, AIB cards will be around $549. Note that that price was prior to the NV 50 series launch. They absolutely could (have) raise(d) it, but if they are smart, they won’t.

Selling at too low prices is counter productive in all counts, if a card is sold 10% lower than its 10% lower performing competition then the general public will assume that it s for a reason, and that the card has hidden flaws or whatever else that justify the lower price tag.

gaav87 · Jan 18, 2025

Would 12V-2x6 have better overclocking vs 3x8pin on taichi vs nitro+ ?
Would atx 3.1 provide better overclocking potential / more stable voltage ?

SolidQ · Jan 18, 2025

marees said:
I don't see jan-23rd in the source.

What did I miss ?

Pre order data

reaperrr3 · Jan 18, 2025

Timorous said:
Nothing to do with cope. Everything to do with 64CUs @ 3Ghz with minimal bandwidth gains vs the 7800XT means without a massive perf per CU uplift the gains will be in the 15-20% region give or take and even with a perf/CU uplift how effective will it be when bandwidth starved? That was the one thing that really held the 7900GRE back.

The 4070S and Ti had less bw and only 48 MB L2, vs. 4MB L2 and 64MB IC on the GRE.

That means RDNA3 still has quite a bit more room for improvement in terms of culling early, compression algorithms and so on.
Of course, there's no guarantee that they completely caught up to Ada in this regard with RDNA4 already, but my point is, just because the GRE was bandwidth-starved doesn't mean that RDNA4 necessarily has a ceiling that low.

DownTheSky · Jan 18, 2025

No. It has a ceiling 40% higher.

adroc_thurston · Jan 18, 2025

reaperrr3 said:
That means RDNA3 still has quite a bit more room for improvement in terms of culling early, compression algorithms and so on.

Caches.

reaperrr3 said:
just because the GRE was bandwidth-starved doesn't mean that RDNA4 necessarily has a ceiling that low.

Indeed. Caches are magical things.

gdansk · Jan 18, 2025

If it has Radeon's patented Big Cache technology, how much will GDDR7 even help?
I suspect GDDR7 version won't be coming soon (until perhaps next year)

Discussion RDNA4 + CDNA3 Architectures Thread

Golden Member

Golden Member

Diamond Member

Golden Member

Lifer

Diamond Member

Golden Member

Diamond Member

Senior member

Golden Member

Diamond Member

Golden Member

Senior member

Senior member

Golden Member

Senior member

Senior member

Diamond Member

Senior member

Lifer

Senior member

Golden Member

Member

Senior member

Diamond Member

Diamond Member