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 8, 2025

insertcarehere said:
Putting aside that 6600s were barely available outside the US for much of the past few years (think 2022/23). Nvidia is practically the only choice for CS Majors who need to learn programming with GPUs (and game a bit for R&R) . 6600 being 40% more powerful or whatever doesn't matter when RoCm is/was a tire fire.

I meant for gamers as that is the primary seller and I wouldn't know about outside the US. I do know that inside the US though that 3050's were all over the top sellers of Amazon compared to 6600's. I doubt many of them were doing RoCM. But I'll shut up about this now as it's already pretty well known anyway and no need to create more controversy.

Win2012R2 · Jan 8, 2025

poke01 said:
RDNA4 is capable of Neural rendering. Lets see if they release support for it next month

NV 50 series appear to have integrated simple tensors deeper to support that, most likely as with RT in RDNA2 in practice NV solution will be a lot quicker

PJVol · Jan 8, 2025

Saylick said:
N31's GCD was around that size if I recall correctly, assuming you consider that a monolithic GPU like RDNA 4 has IO on it as well. If N48 can't even hit similar levels of performance, that would be just sad.

N32 is 200mm2 n5 + 150mm2 n6. Consider going back to monolithic package on N4 lets you save some area from bulky fabric phy on n48. This alone makes one doubt the estimated number of CUs for the 380mn2 sku is correct, or that they are comparable to the old ones, to which "optimized" hinted at in presentation slide.

Thunder 57 said:
I never even finished The Ancient Gods part 2. It seems to try to make it force you to play "their way" ie using certain weapon mods to kill demons

Actually they did that since Part 1 ("spirit"), but I agree that the 2nd dlc is not as good.

SolidQ said:
before they nerf to ground it

Not the first time I heard this quite an exaggeration. More like they rebalanced both parts.

eek2121 · Jan 8, 2025

looked back a few pages, but didn’t see this linked, sorry if I missed: https://videocardz.com/newz/amd-rad...on-average-at-4k-extreme-settings-without-fsr

RX 9070 non-XT as fast as RTX 4080 (no upscaling) in COD benchmark with non-optimal drivers.

marees · Jan 8, 2025

Thunder 57 said:
I do know that inside the US though that 3050's were all over the top sellers of Amazon compared to 6600's. I doubt many of them were doing RoCM.

So many personal anecdotes like this

https://twitter.com/x/status/1809136881151733989

Jan Olšan · Jan 8, 2025

VideoCardz mentions the CoD test showed artifacts, so there is possibility of FPS being lower once the rendering is fixed with proper drivers.

eek2121 · Jan 8, 2025

Jan Olšan said:
VideoCardz mentions the CoD test showed artifacts, so there is possibility of FPS being lower once the rendering is fixed with proper drivers.

The drivers didn’t even recognize the card. Very early drivers. AMD also said that nearly all performance leaks weren’t accurate (rare statement, usually they decline to comment) and that no OEMs had final drivers.

My original prediction was that the top part would be close to 7900XTX in terms of performance, even possibly beating it. I stand by that statement. I suspect raster may lag a bit (up to 15%), heavy RT will be on par/ahead, and light RT will vary on workload, but mostly be close.

While there is a huge SM difference, the clocks on the 9070 are significantly higher, it has RT and there are architectural improvements.

Predictions:

Price : Between $449 and $499

It will trade blows or possibly beat the 5070 (ignoring NVIDIA’s marketing nonsense)

9070 non XT will be a bit behind and will be $50 less.

EDIT: The non-XT has 15% fewer shaders and slightly different clocks according to the rumor mill (which could be right or wrong)

SolidQ · Jan 8, 2025

eek2121 said:
heavy RT will be on par/ahead, and light RT will vary on workload, but mostly be close

I think it will be faster, just only BVH is doubled, not mention other changes

Thunder 57 · Jan 8, 2025

marees said:
So many personal anecdotes like this

https://twitter.com/x/status/1809136881151733989

I saw that posted once here and it is very accurate/sad.

SolidQ · Jan 8, 2025

ASUS Tuf 9700XT

3x8 pin for Extreme OC?

jpiniero · Jan 8, 2025

SolidQ said:
ASUS Tuf 9700XT

3x8 pin for Extreme OC?

There's a model that uses the 12V too:

https://videocardz.com/newz/asrock-adopts-12v-2x6-power-connector-on-radeon-rx-9070-xt-graphics-card

SolidQ · Jan 8, 2025

Powercolor models

Visiting PowerColor at CES 2025 - New Radeon GPUs and more! - OC3D

PowerColor had a trio of new Radeon RDNA 4 GPUs to show us at CES 2025. Red Devil, Hellhound, and Reaper models revealed!

overclock3d.net

Heartbreaker · Jan 8, 2025

Win2012R2 said:
I'll wait for reviews, find it hard to believe 5090 is just 25%-ish over 4090 (without fake frames obviously)

I don't. At some point people have to realize that the fat gains of the good old days are gone. The last time I saw people happy with a release was GTX 10 series.

Everything is flatlining on the process side. Dennard scaling fell apart years ago, transistor scaling slowed down, and cost per transistor really took the biggest hit. It all has consequences.

There is still progress but that progress has slowed. Instead of 70% gains with a new generation, 20%-30% may be the new normal.

SolidQ · Jan 8, 2025

Heartbreaker said:
The last time I saw people happy with a release was GTX 10 series

That because it's was insane gaing between gens + price, and also Pascal still playable today in many games, unless it's UE5, playable games, but feels slow even at 60 fps

Gideon · Jan 8, 2025

eek2121 said:
The drivers didn’t even recognize the card. Very early drivers. AMD also said that nearly all performance leaks weren’t accurate (rare statement, usually they decline to comment) and that no OEMs had final drivers.

My original prediction was that the top part would be close to 7900XTX in terms of performance, even possibly beating it. I stand by that statement. I suspect raster may lag a bit (up to 15%), heavy RT will be on par/ahead, and light RT will vary on workload, but mostly be close.

While there is a huge SM difference, the clocks on the 9070 are significantly higher, it has RT and there are architectural improvements.

Predictions:

Price : Between $449 and $499

It will trade blows or possibly beat the 5070 (ignoring NVIDIA’s marketing nonsense)

9070 non XT will be a bit behind and will be $50 less.

EDIT: The non-XT has 15% fewer shaders and slightly different clocks according to the rumor mill (which could be right or wrong)

My predictions were initially somewhat similar, but if the die size is correct (and why shoudn't it be?) at ~390 mm² it changes the equation a bit

It's comparable to the GB203 (that is in 5070 Ti and 5080) and about ~100 mm² larger than the AD104 that's used in 4070 SUPER and 4070 Ti. (GB205 used in 5070 probably has similar size, but not confirmed yet)

If AMD only "trades blows" with 5070 (in rasterization!) with a chip that's bigger than the one used in 5080 It'll be a bit dissapointing (miles better than intel, but still)

TBF the RT uplift on RTX 5070 appears to be quite large, but i'm pretty sure the raster uplift is actually <30%, e.g. within 10% from the 4070 SUPER

https://www.reddit.com/r/nvidia/comments/1hvvrqj/50_vs_40_series_nvidia_benchmark_exact_numbers

SolidQ · Jan 8, 2025

Gideon said:
If AMD only "trades blows" with 5070 (in rasterization!) with a chip that's bigger than the one used in 5080 It'll be a bit dissapointing

Unless it's big WGP's, RT, ML's etc

Preorders? Lol

Hitman928 · Jan 8, 2025

del42sa said:
all prediction went wrong ? remember this one ?

View attachment 114471

Yep, it's pretty funny watching the leakers and people who claim to be in the know continuously speaking confidently about things, only to be proven wrong with reality over and over again.

Saylick · Jan 8, 2025

Heartbreaker said:
I don't. At some point people have to realize that the fat gains of the good old days are gone. The last time I saw people happy with a release was GTX 10 series.

Everything is flatlining on the process side. Dennard scaling fell apart years ago, transistor scaling slowed down, and cost per transistor really took the biggest hit. It all has consequences.

There is still progress but that progress has slowed. Instead of 70% gains with a new generation, 20%-30% may be the new normal.

Even with Nvidia’s approach to maximize tensor performance so that they can leverage software improvements instead of pure raster improvements is going to come to a halt soon enough. Right now, Blackwell’s tensor units go down to INT4/FP4… I suppose they can try INT2/FP2, but with precision that low I’m not convinced it will be a meaningful improvement.

Heartbreaker · Jan 8, 2025

Saylick said:
Even with Nvidia’s approach to maximize tensor performance so that they can leverage software improvements instead of pure raster improvements is going to come to a halt soon enough. Right now, Blackwell’s tensor units go down to INT4/FP4… I suppose they can try INT2/FP2, but with precision that low I’m not convinced it will be a meaningful improvement.

LOL!

I'm surprised they can get anything useful out FP4/Int4. I wonder if people are really using those, or is their best use case inflating benchmarks (Kind of an analog to Fake Frame generation in games).

Kepler_L2 · Jan 8, 2025

SolidQ said:
if that real, what is there new? RT?

https://videocardz.com/newz/amd-navi-48-rdna4-gpu-for-radeon-rx-9070-pictured-may-exceed-nvidia-ad103-size

New RT and Matrix cores mostly, but I think that 390mm² measurement is wrong, I got it closer to 300-330 using Krackan (190mm²) or Zen5 CCD (66mm²) as baseline.

gdansk · Jan 8, 2025

Saylick said:
Even with Nvidia’s approach to maximize tensor performance so that they can leverage software improvements instead of pure raster improvements is going to come to a halt soon enough. Right now, Blackwell’s tensor units go down to INT4/FP4… I suppose they can try INT2/FP2, but with precision that low I’m not convinced it will be a meaningful improvement.

The people repeating "raster is dead" never finish the sentence. Raster is dead and raytracing will never catch up.

But since machines will be inferring most pixels and most frames it makes raytracing relatively more appealing and increasingly smaller deficit.

It seems most the models used to do this will be overfit and hard to replicate. So we're heading toward something really swell I'm sure.

Saylick · Jan 8, 2025

Kepler_L2 said:
New RT and Matrix cores mostly, but I think that 390mm² measurement is wrong, I got it closer to 300-330 using Krackan (190mm²) or Zen5 CCD (66mm²) as baseline.

No real matrix cores, though, right? It’s just improved WMMA.

Saylick · Jan 8, 2025

Heartbreaker said:
LOL!

I'm surprised they can get anything useful out FP4/Int4. I wonder if people are really using those, or is their best use case inflating benchmarks (Kind of an analog to Fake Frame generation in games).

I mean, how else would they keep this plot going exponentially upwards if it doesn’t include lower precision? If the precision bottoms out, it will start to plateau at the same rate you mentioned of 20-30% per new node generation.

gdansk · Jan 8, 2025

Saylick said:
I mean, how else would they keep this plot going exponentially upwards if it doesn’t include lower precision? If the precision bottoms out, it will start to plateau at the same rate you mentioned of 20-30% per new node generation.
View attachment 114495

We're going to infer inference. Why actually calculate 4 whole bits when we can calculate 1 and predict the other 3. And since we're predicting the least significant bits it has little impact, really.

adroc_thurston · Jan 8, 2025

Kepler_L2 said:
New RT and Matrix cores mostly, but I think that 390mm² measurement is wrong, I got it closer to 300-330 using Krackan (190mm²) or Zen5 CCD (66mm²) as baseline.

Tlogic measured 12.3 by 28.4 so yeah 350 or below.

Discussion RDNA4 + CDNA3 Architectures Thread

Golden Member

Diamond Member

Senior member

Senior member

Diamond Member

Senior member

Senior member

Diamond Member

Golden Member

Diamond Member

Golden Member

Lifer

Golden Member

Diamond Member

Golden Member

Golden Member

Golden Member

Diamond Member

Diamond Member

Diamond Member

Senior member

Diamond Member

Diamond Member

Diamond Member

Diamond Member

Diamond Member