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

Thunder 57 · Jan 10, 2025

soresu said:
I'd be the first to moan about RDNA3's misleading dual issue FLOPS figures, but nVidia have been pushing increasingly silly numbers over the last several generations.

We often take a pot shot at tensor cores taking up room on nVidia's GPUs, but I'm beginning to wonder if they are actually cannibalising the general compute/gfx/CUDA cores in order to make room for more/beefier tensor cores to push the AI craze until the bubble explodes.

Indeed. Just like when Nvidia decided to double their CUDA cores a few generations ago.

soresu · Jan 10, 2025

SolidQ said:
6. RDNA 4 has "much much better, but muuuuuch better" Raytracing performance, he insists saying it twice in the interview

Looking at the Blender Cycles perf between AMD and nVidia I can see why.

Since RDNA1 they've made great strides with raster gfx, but they got caught with their pants down with nVidia's RT pivot, so I'm guessing that they cobbled together something as best they could for RDNA2/3, with RDNA4 being planned to be their proper RT moment.

I suspect that sometime between that original plan and now they decided that splitting the accelerator µArchs for pro/AI/HPC and gaming wasn't working for them in terms of industry participation due to having to split dev attention, so RDNA4 hi end was canned to speed up RDNA5/UDNA.

soresu · Jan 10, 2025

Thunder 57 said:
Indeed. Just like when Nvidia decided to double their CUDA cores a few generations ago.

It makes a lot more sense in light of them pushing DLSS and neural gfx techniques so hard.

They are going to do a 2018 again and push some sort of future iteration of ReSTIR and neural augmentation hoping to sucker punch AMD while fixing the RT perf issue.

gaav87 · Jan 10, 2025

Adroc ur take on my NO decoupled clock theory ?

Win2012R2 · Jan 10, 2025

soresu said:
I'm beginning to wonder if they are actually cannibalising the general compute/gfx/CUDA cores

Looks like it

gaav87 · Jan 10, 2025

Win2012R2 said:
Looks like it

Well grace blackwell is, a AI gpu.

adroc_thurston · Jan 10, 2025

soresu said:
but nVidia have been pushing increasingly silly numbers over the last several generations.

Ehhh, it's just GB202 that feels really off.

soresu said:
if they are actually cannibalising the general compute/gfx/CUDA cores in order to make room for more/beefier tensor cores

Oh hell no. Client Blackwell sees zero throughput improvments on its matrix cores outside of newer (and miserable to use) FP6/4 formats.

adroc_thurston · Jan 10, 2025

soresu said:
so I'm guessing that they cobbled together something as best they could for RDNA2/3, with RDNA4 being planned to be their proper RT moment.

No.
They waited for an API evolution that never came.
Like RTRT is a dead end now. DXR won't ever be flexible and no programmable RTRT hardware will ever ship.

soresu said:
RDNA4 hi end was canned to speed up RDNA5/UDNA.

It was canned because GB202.
It does not speed up any gfx13 part in any way, they're about 8q out.

SolidQ · Jan 10, 2025

adroc_thurston said:
it's just GB202 that feels really off.

Need also CPU bound test without FG mode

branch_suggestion · Jan 10, 2025

They Let Bring a Camera Into a Top Classified US Supercomputer El Capitan

We had the opportunity to take photos and film inside El Capitan, the number 1 Top500 supercomputer as it enters its classified mission

www.servethehome.com

So that is where Dr. Su is.

Keller_TT · Jan 10, 2025

coercitiv said:
If N48 is that much better than their existing lineup and FSR4 is pumping on all cylinders, then price it like 6800XT. However, if the performance jump is incremental and FSR4 still needs time to reach maturity (and more games), then it's 7800XT pricing.

For it to be considered a gen on gen upgrade from N32, assuming that RDNA4 at the least matches the best case uplift achieved with RDOA3, i.e., the upsold XTX, it simply can not be priced like a 6800 XT at the apogee of crypto boom. The goal of RDNA4 was to maximize efficiency and drive down costs.

Since N48 has some meat to it, AMD will have it to peg it at optimum PPW over 7900 XT and price it at $500 to really beat RTX+DLSS+5070, or show to the world that they never learnt anything from 7900 XT greedy launch. It'll be a clown show!

SolidQ · Jan 10, 2025

But we can have fun, AMD!

Nobody has the final driver, not even the board manufacturers, so don’t believe performance claims on the Internet.

– AMD Representative – CES 2025

reaperrr3 · Jan 10, 2025

Keller_TT said:
For it to be considered a gen on gen upgrade from N32, assuming that RDNA4 at the least matches the best case uplift achieved with RDOA3, i.e., the upsold XTX, it simply can not be priced like a 6800 XT at the apogee of crypto boom.

I assume anyone mentioning 6800XT pricing here is talking about its launch MSRP of $649, which was actually a little before the crypto/scalper boom went off (because otherwise AMD and NV would've set the MSRPs of Ampere and RDNA2 much higher).

Keller_TT said:
The goal of RDNA4 was to maximize efficiency and drive down costs.

Considering it seems to be at minimum 330mm2, all in N4P which just had its wafer prices raised, I doubt it's cheaper to make than N32 or N21.

Keller_TT said:
Since N48 has some meat to it, AMD will have it to peg it at optimum PPW over 7900 XT and price it at $500 to really beat RTX+DLSS+5070, or show to the world that they never learnt anything from 7900 XT greedy launch. It'll be a clown show!

The press interview they gave after the cancelled CES presentation to me sounded like they won't go for the PPW sweet-spot, but rather PP$. And since I doubt they want to give the cards away for pennies in terms of margins, that makes me believe at least the XT's clockspeed will be cranked up as much as their yield rate + the quality of AIB partner PCBs/coolers reasonably allows.

But yeah, I wouldn't expect 649$ now even if the XT matched the 5070Ti in every aspect, at that price too many people would find an excuse to go for either the 5070 or 5070Ti.

jpiniero · Jan 10, 2025

reaperrr3 said:
I assume anyone mentioning 6800XT pricing here is talking about its launch MSRP of $649

If the two models are 16 GB, then the rumors of the 9700 Non XT being only a very slight cut are true. Which is dumb, but that's AMD for ya.

I could see them originally intending $649 and $599 (figuring people will buy the $599 model of the two, and the faster one is just so it looks a little better in benchmarks...) but that was with the expectation that it would only be a 5070 competitor.

PJVol · Jan 10, 2025

reaperrr3 said:
I doubt it's cheaper to make than N32 or N21

Probably around n32, so $499-$549

coercitiv · Jan 10, 2025

reaperrr3 said:
I assume anyone mentioning 6800XT pricing here is talking about its launch MSRP of $649

Yup, that's exactly what I was referencing, and it was not meant as an absolute reference to the $650 value, but rather to the $150 delta in price between the two products.

Heartbreaker · Jan 10, 2025

jpiniero said:
If the two models are 16 GB, then the rumors of the 9700 Non XT being only a very slight cut are true. Which is dumb, but that's AMD for ya.

There is nothing wrong with a slight cut if that is enough to recover most of the chips with a bad section.

It's all in the pricing.

There also seem to be some speculation that there are two cuts:

9070 - Slight cut with 16 GB
9060 XT(X?) deeper cut with 12 GB.

That seems rather sensible since it appears Navi 44 is half the size of Navi 48.

SolidQ · Jan 10, 2025

2. He says that most gamers these days don't want to spend $1000 in a GPU, they released $1000 GPUs last gen but only got a very favorable response from the market when they lowered the prices and released the 7900 GRE.

That interesting thing, i think they hint price. around 7800XT/7900GRE

Gideon · Jan 10, 2025

SolidQ said:
That interesting thing, i think they hint price. around 7800XT/7900GRE

That makes the most sense to me as well. With the cut-down 12GB model (RX 9060 XT, RX 9070 GRE or whatever they'll call it) probably taking the 7700XT slot price-wise.

At least that's what I wish they'd do.

Antything less doesn't seem that likely, unless there are significant (performance or feature) caveats we don't know about

beginner99 · Jan 10, 2025

Keller_TT said:
price it at $500 to really beat RTX+DLSS+5070, or show to the world that they never learnt anything from 7900 XT greedy launch. It'll be a clown show!

I mean personally I'm all for that being $500 but no, maybe the non-xt but certainly not the XT IF the performance rumors are true and performance is higher than expected at 5070 Ti levels.

jpiniero said:
If the two models are 16 GB, then the rumors of the 9700 Non XT being only a very slight cut are true. Which is dumb, but that's AMD for ya.

I could see them originally intending $649 and $599 (figuring people will buy the $599 model of the two, and the faster one is just so it looks a little better in benchmarks...) but that was with the expectation that it would only be a 5070 competitor.

I think best case scenario is $499 non-xt and $599 xt.

The non-xt as rumored has slower memory speed and likely slower clocks. so the binning will be mostly on clocks not just defects which makes sense if they aim for high clocks. this will also make the non-xt somewhat relevant as the $50 difference at that price is so little it doesn't really make sense to not spend it for the additional benefits. And lower clocks means less power so smaller cooler, dual-fan designs possible. in essence there will also be a perfromance difference worth that $100 between the 2.

jpiniero · Jan 10, 2025

Heartbreaker said:
There is nothing wrong with a slight cut if that is enough to recover most of the chips with a bad section.

It's all in the pricing.

Performance gap won't really be much.

Heartbreaker · Jan 10, 2025

jpiniero said:
Performance gap won't really be much.

If they are only making one cut and it's rather shallow, and they mess up pricing again, then there will be cause to complain.

If they are making two cuts, one shallow and 16GB, plus one deeper and 12GB, that seems like the right thing to me, but again it's all in the pricing, and not doing another 7900XTX/XT flub.

This is definitely seems like a minor wait and see item.

marees · Jan 10, 2025

The fact that this is named 70 series & 80 series makes me think that MBA boards won't go above $550
The non-xt is likely to come slightly under 7800xt at $480

Since board partner cards have 3x 8-pin they could be priced at $600

As for the further cut down 12gb card, it is likely to be launched very late for $380

jpiniero · Jan 10, 2025

Heartbreaker said:
If they are only making one cut and it's rather shallow, and they mess up pricing again, then there will be cause to complain.

Aparently the two products are the full and the very shallow. Both with 16 GB.

Keller_TT · Jan 10, 2025

PJVol said:
Probably around n32, so $499-$549

My run of the mill guess:
Assuming DRAM has largely stagnated in scaling sub-10nm, and TSMC's stated N4P vs N5 is 6% higher density and 22% better power efficiency, it kind of makes for a 350 mm² die size for N48. If AMD have managed density efficiency like they achieved with Zen 5 over Zen 4, N48 can pack overall roughly 30% more transistors than N32 for a monolithic 350 mm² die.
If the ML and BVH hardware added more bulk, we're probably looking at a modest increase from my base figure.

Now to performance, I would think this package would be at least 20% more than the 7900 GRE at PPW level. The GRE was a direct glimpse of RDNA 3 vs 2: match 335W 80 CUs 6950 XT at 265W. This and XTX were the best case efficiency figures for RDNA3.

For pricing, I would forget Nvidia's games and treat it like 5700 XT to 6700 XT and keep $500 for MBA card (550 max stretch) because 6800 XT to 7800 XT was funny naming, poor efficiency, stagnant performance. The 9700 XT in ideal AMD universe should be at minimum 7900GRE + 10% for all their funny PPT claims.

Discussion RDNA4 + CDNA3 Architectures Thread

Golden Member

Diamond Member

Diamond Member

Diamond Member

Senior member

Senior member

Senior member

Diamond Member

Diamond Member

Golden Member

Senior member

Member

Golden Member

Member

Lifer

Senior member

Diamond Member

Diamond Member

Golden Member

Golden Member

Diamond Member

Lifer

Diamond Member

Senior member

Lifer

Member