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

gaav87 · Jan 29, 2025

Well it looks like
9070 xt (304w) vs 5080 (360W)
14% slower in shadow of tombraider
10% slower in horizon zero dawn
8% slower in cyberpunk 2077

SolidQ · Jan 29, 2025

adroc_thurston said:
It's not late even with RDNA4.
They only have to commit.

answer to super series?

9070 xt (304w) vs 5080 (360W)
14% slower in shadow of tombraider
10% slower in horizon zero dawn
8% slower in cyberpunk 2077

with those difference they just need 9070XT GDDR7 version to match 5080

gaav87 · Jan 29, 2025

SolidQ said:
answer to super series?

with those difference they just need 9070XT GDDR7 version to match 5080

24GB gddr7 9080xt 749$ and ur done for next 1.5-2years.

adroc_thurston · Jan 29, 2025

SolidQ said:
with those difference they just need 9070XT GDDR7 version to match 5080

Yea.

gaav87 said:
24GB gddr7 9080xt 749$ and ur done for next 1.5-2years.

Make that $799 and your golden.

Timorous · Jan 29, 2025

What was the Navi 4C spec again? 9 SEDs with 1 SE per SED with 16 to 20 CUs per SE? Something like that? So 144 - 180 CU. So even at the low end the gap between N48 and N4C in raw shader count would be similar to the gap between 5080 and 5090.

So some very basic napkin math would put a N4C part with a 60% performance advantage over N48 at 4K somewhere in the region of 50% faster than the 7900XTX give or take a bit. That puts it between 4090 and 5090 performance.

Obviously there is a lot of assumptions in that very back of the envelope calculation but it seems plausible that N4C could compete with the 5090 in performance. I guess the question would be does the multi GCD config throw up any software issues? Does the BOM cost make sense and would the feature set make people even consider it for ~$1.5K (no way could AMD command a $2K price tag like the 5090).

Even if you said right now that it would be 20% faster than the 5090 in both raster and RT at native res and it had a lower TDP than the 5090 it would still be a tough sell and it would take up a lot of wafer and packaging capacity that could be used for 9800X3D which clearly has far better margins and is selling like hot cakes. We all know even in the scenario where AMD have a clear performance per $ and performance per watt advantage NV would push DLSS and Frame Gen hard and plenty of sites would go along with that. I could imagine some utterly changing their testing method to do something like what [H] did with highest playable settings where they say something like 'DLSS Q looks as good as native so we are using that on the 5090 but FSR Q not so much so we needed to run at native res on the AMD card and in that scenario with very equal IQ the 5090 has a clear performance advantage'.

From a competitive perspective it seems like an L but from a business standpoint I am not so sure about that, there is a lot of market friction to over come and NV have managed to muddy the typical bigger number better metric to such a degree even a decisive performance, efficiency and cost advantage may not be enough to over come it.

adroc_thurston · Jan 29, 2025

Timorous said:
What was the Navi 4C spec again?

9 SED times 3 SA each, giving you a multiple of 27 on the top end.

Timorous said:
So 144 - 180 CU.

W-rong.
It was a real funky part

basix · Jan 29, 2025

ToTTenTranz said:
I wonder if XDNA2 can realistically be used for an upscaler. How close is the NPU to the GPU?
I'm guessing there's no direct bus between each others' caches, so everything must go through the RAM.

I suspect, that this should work. But I expect the offloading only for the DNN part, not FSR4 as a whole. 50 TFLOPS Block-FP16 compared to whatever FP8 TFLOPS of N44/N48 are looking good enough for an iGPU.

I mean, 50 TFLOPS are much more than what e.g. Strix Point's iGPU can deliver (max. 25 TFLOPS FP16 WMMA). And offload compute load from the already stressed iGPU.

AND: It would be a massive uplift and selling point improvement compared mobile RTX40/50, which all get DLSS4.

uzzi38 said:
I don't think putting extra stress on the memory controller as you pass data from the GPU to the NPU and back again (going through main memory in both cases) is a very sensible idea for an already memory-bound APU...

And I also don't think the NPU is fast enough to warrant it.

Why should that go through main memory? The NPU is (in my opinion) only supposed to accelerate the DNN part of FSR4 and not execute FSR4 as a whole. Anything fetched by the NPU would have to be fetched by the iGPU as well.

And shouldn't it be possible to transfer data via Infinity Fabric within IP-Blocks of the APU, if necessary? So if you have concurrent processing on NPU and iGPU (could be, depending on the architecture of FSR4) I suspect that most of it can stay on-chip, so no or only minor main memory bandwidth requirement increase.
And I suspect that the NPU is more bandwidth efficient in general regarding DNN execution (local SRAM) and should execute it faster (latency wise, better TFLOPS utilization).

And as I have written above: 50 TFLOPS Block-FP16 > 25 TFLOPS FP16 WMMA

H T C · Jan 29, 2025

AMD has a "golden opportunity" here to REALLY mess up nVidia, now that "the cat is out of the bag" and the 5090 is not as powerful (relative to the 4090) as they feared.

It OBVIOUSLY depends on the performance of their cards BUT if AMD's top GPU is:

- better than the 5080 (VERY unlikely), then price it @ ... say ... $600
- a bit worse than the 5080 (only a bit), then price it @ ... say ...$500
- considerably worse than the 5080 (likely), then price it @ ... say ... $400

UNLESS they're THIS aggressive with their pricing, i don't see AMD having better luck with this generation than they had in their current one.

OBVIOUSLY, this is gonna "wreck havoc" with the current cards still up for sale and that will cost A TON of money: it's SUPPOSED to be a "calculated risk", designed to MASSIVELY increase their market share.

If it works, they'll end up losing QUITE A BIT of money NOW and earn A LOT of money later ... but if it doesn't ... they'll lose A LOT of money now AND later ...

adroc_thurston · Jan 29, 2025

H T C said:
AMD has a "golden opportunity" here to REALLY mess up nVidia, now that "the cat is out of the bag" and the 5090 is not as powerful (relative to the 4090) as they feared.

It OBVIOUSLY depends on the performance of their cards BUT if AMD's top GPU is:

- better than the 5080 (VERY unlikely), then price it @ ... say ... $600
- a bit worse than the 5080 (only a bit), then price it @ ... say ...$500
- considerably worse than the 5080 (likely), then price it @ ... say ... $400

UNLESS they're THIS aggressive with their pricing, i don't see AMD having better luck with this generation than they had in their current one.

OBVIOUSLY, this is gonna "wreck havoc" with the current cards still up for sale and that will cost A TON of money: it's SUPPOSED to be a "calculated risk", designed to MASSIVELY increase their market share.

If it works, they'll end up losing QUITE A BIT of money NOW and earn A LOT of money later ... but if it doesn't ... they'll lose A LOT of money now AND later ...

These are words and no they're not gonna be aggressive with their pricing. No one buys Radeon.

H T C · Jan 29, 2025

adroc_thurston said:
These are words and no they're not gonna be aggressive with their pricing. No one buys Radeon.

It's PAINFULLY OBVIOUS the current tactics of selling similar performing cards for slightly less or slightly better performing cards for the same price as their nVidia counterparts DON'T WORK, so it's about time they resort to some "more drastic measures".

The idea is to "change that mindset", even by resorting to sell @ little to no profit NOW, in order to MASSIVELY increase the market share, and reap the benefits of it in the NEXT generation of cards: it's NOT SUPPOSED to "pay off now".

AMD simply could not afford to take an aggressive approach like this before, but they can now: question is ... will they ...?

Hans Gruber · Jan 29, 2025

H T C said:
AMD has a "golden opportunity" here to REALLY mess up nVidia, now that "the cat is out of the bag" and the 5090 is not as powerful (relative to the 4090) as they feared.

It OBVIOUSLY depends on the performance of their cards BUT if AMD's top GPU is:

- better than the 5080 (VERY unlikely), then price it @ ... say ... $600
- a bit worse than the 5080 (only a bit), then price it @ ... say ...$500
- considerably worse than the 5080 (likely), then price it @ ... say ... $400

UNLESS they're THIS aggressive with their pricing, i don't see AMD having better luck with this generation than they had in their current one.

OBVIOUSLY, this is gonna "wreck havoc" with the current cards still up for sale and that will cost A TON of money: it's SUPPOSED to be a "calculated risk", designed to MASSIVELY increase their market share.

If it works, they'll end up losing QUITE A BIT of money NOW and earn A LOT of money later ... but if it doesn't ... they'll lose A LOT of money now AND later ...

The early leaked reviews show the 5080 is 8% better performance than the 4080. That means the uplift if any over 4080 is due to GDDR7 vs. GDDR6X. Plus more power consumption with the 5080 vs. 4080.

AMD does not have drivers equal to Nvidia. AMD does not update their drivers weekly or bi-weekly like Nvidia. AMD does not see their GPU's as 2nd tier to Nvidia and they do not price them accordingly. The GPU sector will probably destroy the PC industry with these paltry generational improvements.

del42sa · Jan 29, 2025

H T C said:
AMD has a "golden opportunity" here to REALLY mess up nVidia, now that "the cat is out of the bag" and the 5090 is not as powerful (relative to the 4090) as they feared.

It OBVIOUSLY depends on the performance of their cards BUT if AMD's top GPU is:

- better than the 5080 (VERY unlikely), then price it @ ... say ... $600
- a bit worse than the 5080 (only a bit), then price it @ ... say ...$500
- considerably worse than the 5080 (likely), then price it @ ... say ... $400

UNLESS they're THIS aggressive with their pricing, i don't see AMD having better luck with this generation than they had in their current one.

OBVIOUSLY, this is gonna "wreck havoc" with the current cards still up for sale and that will cost A TON of money: it's SUPPOSED to be a "calculated risk", designed to MASSIVELY increase their market share.

If it works, they'll end up losing QUITE A BIT of money NOW and earn A LOT of money later ... but if it doesn't ... they'll lose A LOT of money now AND later ...

my guess is $670 for RX9070XT and $549 for RX9070 with some discount later as usual ...

adroc_thurston · Jan 29, 2025

H T C said:
It's PAINFULLY OBVIOUS the current tactics of selling similar performing cards for slightly less or slightly better performing cards for the same price as their nVidia counterparts DON'T WORK, so it's about time they resort to some "more drastic measures".

The idea is to "change that mindset", even by resorting to sell @ little to no profit NOW, in order to MASSIVELY increase the market share, and reap the benefits of it in the NEXT generation of cards: it's NOT SUPPOSED to "pay off now".

AMD simply could not afford to take an aggressive approach like this before, but they can now: question is ... will they ...?

I have some bad news for you, ATi tried all this with RV700s and Evergreens and they lost horribly. I get it, you're a zoomer and missed out on all the oldskool GPU wars.
No one buys Radeon, man. Just not a brand people consider worthwhile.

H T C · Jan 29, 2025

adroc_thurston said:
I have some bad news for you, ATi tried all this with RV700s and Evergreens and they lost horribly. I get it, you're a zoomer and missed out on all the oldskool GPU wars.
No one buys Radeon, man. Just not a brand people consider worthwhile.

They didn't buy AMD CPUs either ... until ZEN.

What changed, when Zen arrived?

They priced it VERY AGGRESSIVELY.

So they have a history of doing EXACTLY THIS, with CPUs, and IT WORKED then: why can't it work now, with GPUs?

Win2012R2 · Jan 29, 2025

adroc_thurston said:
No one buys Radeon, man. Just not a brand people consider worthwhile.

Who would, with this defeatist attitude?

They had chiplets and should have made NVidia+50% fast chip - with better RT obviously, that's the loss (breakeven) leader they could have afforded.

H T C said:
why can't it work now, with GPUs?

Two problems -
1) way more silicon needed - 600 sq mm is like 8 Zen5 chiplets (with better yields per chip) - making 64 core Epyc that sells for $4-5k.
2) Nvidia also uses TSMC so no process advantage

adroc_thurston · Jan 29, 2025

H T C said:
They didn't buy AMD CPUs either ... until ZEN.

What changed, when Zen arrived?

They priced it VERY AGGRESSIVELY.

So they have a history of doing EXACTLY THIS, with CPUs, and IT WORKED then: why can't it work now, with GPUs?

Ryzen wasn't a thing until 3950X, or more precisely, 5950X. You know, the parts that won it all and were priced obscenely, far above anything non-HEDT ever.

Win2012R2 said:
Who would, with this defeatist attitude?

They were defeated in detail 15 years ago when Fermi outsold Evergreen. Everything after was just slow agony.

Win2012R2 said:
They had chiplets and should have made NVidia+50% fast chip - with better RT obviously, that's the loss (breakeven) leader they could have afforded.

"Should have would have" is not how you justify running a pricey program in front of dr. Lisa Su.

basix · Jan 29, 2025

9070XT on 5070Ti level for 600$ would already deliver a quite attractive Perf/$ ratio. AMD is not the Salvation Army, but they need to deliver 1.2x or higher Perf/$ ratio in order to convince people to buy AMD. Anything lower and people will still buy Nvidia cards.
600$ MSRP should also be maintainable for AMD. The 7800XT was released at 500$, with same VRAM capacity and similar TDP. N48 will probably be a notch more expensive to produce than N32 (assuming 330...350mm2 Die size vs. 200 + 4*37.5mm + organic RDL packaging) but I think not by too much.

But it will stand or fall with FSR4. DLSS4 has improved by so much regarding image clarity and stability, that Transformer-DLSS-P is quite usable even at 1440p. 4K Transformer-DLSS-P looks very good and in my opinion better than CNN-DLSS-Q and in some regards even CNN-DLAA (mainly image clarity and ghosting/smearing, not so much regarding noise and flickering).
- If FSR4 achieves DLSS3 quality levels (Super Resolution part), it will already be a great improvement compared to FSR2/3. But the competitor is DLSS4
- Ray reconstruction should be there as well. I am optimistic here, because AMD has already teasered that a few times
- MFG is not that important, but could be the icing on the cake. 5070 Ti / 9070 performance should often be sufficient to use FG, especially at 1440p. If you have FG, MFG should not be a too far step from it (frame pacing and image quality are getting more important with MFG, that is clear)
- RDNA4 differentiation possibilites vs. Blackwell: FG Extrapolation (no latency increase), RSR @ AutoSR, AFMF3 @ ML/AI/Extrapolation

If FSR4 comes close to DLSS4 regarding quality, RDNA4 could be a banger. And if the last point (differentiation features) gets delivered as well, it could even surpass Blackwell regarding gaming relevant feature set. Unseen from AMD.

But: It's optimistic, I know.

adroc_thurston · Jan 29, 2025

basix said:
9070XT on 5070Ti level for 600$ would already deliver a quite attractive Perf/$ ratio. AMD is not the Salvation Army, but they need to deliver 1.2x or higher Perf/$ ratio in order to convince people to buy AMD. Anything lower and people will still buy Nvidia cards.
600$ MSRP should also be maintainable for AMD. The 7800XT was released at 500$, with same VRAM capacity and similar TDP. N48 will probably be a notch more expensive to produce than N32 (assuming 330...350mm2 Die size vs. 200 + 4*37.5mm + organic RDL packaging) but I think not by too much.

But it will stand or fall with FSR4. DLSS4 has improved by so much regarding image clarity and stability, that Transformer-DLSS-P is quite usable even at 1440p. 4K Transformer-DLSS-P looks very good and in my opinion better than CNN-DLSS-Q and in some regards even CNN-DLAA (mainly image clarity and ghosting/smearing, not so much regarding noise and flickering).
- If FSR4 achieves DLSS3 quality levels (Super Resolution part), it will already be a great improvement compared to FSR2/3. But the competitor is DLSS4
- Ray reconstruction should be there as well. I am optimistic here, because AMD has already teasered that a few times
- MFG is not that important, but could be the icing on the cake. 5070 Ti / 9070 performance should often be sufficient to use FG, especially at 1440p. If you have FG, MFG should not be a too far step from it (frame pacing and image quality are getting more important with MFG, that is clear)
- RDNA4 differentiation possibilites vs. Blackwell: FG Extrapolation (no latency increase), RSR @ AutoSR, AFMF3 @ ML/AI/Extrapolation

If FSR4 comes close to DLSS4 regarding quality, RDNA4 could be a banger. And if the last point (differentiation features) gets delivered as well, it could even surpass Blackwell regarding gaming relevant feature set. Unseen from AMD.

But: It's optimistic, I know.

That's a lot of words and none of them mean anything.
People won't buy Radeon, simple as.
It's a good product that is worthless without a 3 kilobucked tiled chop chop on top of it.

basix · Jan 29, 2025

You emphasize the Top Dog too much. Most people are interested in good Perf/$ at affordable price points. And that is 600$ range and below.

Yeah, 5090 is cool and such. But most people immediately exlcude it from their buy option evaluation because just too expensive.

In earlier days, the Top Dog argument would have worked. But not today anymore at 2000$ MSRP.

adroc_thurston · Jan 29, 2025

basix said:
You emphasize the Top Dog too much.

The only thing that ever matters.

basix said:
Most people are interested in good Perf/$ at affordable price points. And that is 600$ range and below.

Only if the Brand is worth it. And you can only brand build with a chop chop hatchet for 3 kilobuck apiece.

RV770 turned Radeon into a poverty toy no one truly wants and it hasn't recovered ever since.

basix · Jan 29, 2025

Top Dog / Halo part helps to build branding, for sure.

But having features which the other has not (differentiation features above) would help as well.

And as I said, with such extreme high prices people wake up and look, what is below? And there you compare much more what you get for your money.
Anything else (just seeing the shiny top part) is just stupid and not worth being an adult person. And I believe, that most people are capable of a little bit of thinking

adroc_thurston · Jan 29, 2025

basix said:
But having features which the other has not (differentiation features above) would help as well.

Shill factors are trivial to shuffle around.
Remember when NVENC was Mega Important and everyone was a streamer?

basix said:
Top Dog / Halo part helps to build branding, for sure.

It's the only way to brand build in cultish markets.
Behead the current pastor, take his place, lead the cult. Pretty simple. Only that AMD doesn't have the guts.

basix · Jan 29, 2025

adroc_thurston said:
It's the only way to brand build in cultish markets.
Behead the current pastor, take his place, lead the cult. Pretty simple. Only that AMD doesn't have the guts.

You imply that most people are just stupid. I have a more positive point of view: People are able to think and reason Especially because the user base (aka gamers) gets older and therefore also their reasonig should get more sophisticated and nuanced.

"I just buy Nvidia because the have the shiniest card" might work for kids and teens, but should not for 25y or 30y+ grown-up old adults (who have the money to buy the cards). At least I firmly hope that this is the case, otherwise people are really stupid and we are doomed (not regarding GPU buying decisions, but as a society as a whole).

ToTTenTranz · Jan 29, 2025

adroc_thurston said:
It's not late even with RDNA4.
They only have to commit.

I wonder what would be the chances for other RDNA4 chips releasing before next arch, considering what they know now.

poke01 · Jan 29, 2025

drop the Radeon name then. Call it something else. Worked for lots of companies

Discussion RDNA4 + CDNA3 Architectures Thread

Golden Member

Senior member

Golden Member

Senior member

Diamond Member

Golden Member

Diamond Member

Member

Senior member

Diamond Member

Senior member

Platinum Member

Member

Diamond Member

Senior member

Senior member

Diamond Member

Member

Diamond Member

Member

Diamond Member

Member

Diamond Member

Member

Senior member

Diamond Member