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

Hans Gruber · Apr 5, 2024

Is there any truth to RDNA3 being broken from the outset? Will RDNA4 work correctly upon release? Remember when they were talking about 3ghz GPU's.

SolidQ · Apr 5, 2024

Hans Gruber said:
s there any truth to RDNA3 being broken from the outset? Will RDNA4 work correctly upon release?

That what we know for now

Mahboi · Apr 5, 2024

Hans Gruber said:
Is there any truth to RDNA3 being broken from the outset? Will RDNA4 work correctly upon release? Remember when they were talking about 3ghz GPU's.

You're STILL asking????
Projected clocks were missed by 20%, perf is way below announced, power draw at idle and load are awful. What doubts are there to have still?
RDNA 3.5 will show what RDNA 3 was meant to be in a few months anyway.

Ghostsonplanets · Apr 5, 2024

SolidQ said:
Now we need to know, last part, it's RT/PT Ampere or Ada level.

Rumors say RDNA 4 RT performance is a great uplift. So I guess we should expect something Ada level at least. Let's wait and see.

Mahboi · Apr 5, 2024

I'm not sure what people mean by Ampere/Ada "levels".

IIRC RDNA2/3 suffers with the number of light bounces, I.E, AMD RT is not very good, but acceptable for most games. Some (Nvidia supported ones) like Portal RTX, Cyberpunk and I think Control (possibly Alan Wake?), that use more bounces than necessary, dumpster the AMD RT performance. I think it takes an XTX from being roughly a 3080 Ti competitor in RT down to a 3070.
If they fix that, already we're getting "Ampere levels" even just with RDNA 3 performance. An XTX is somewhere in the relatively higher end of Ampere. Sure, it should compete with Lovelace in the first place, but I don't see how "is it Ampere, or is it Lovelace" is the question when Blackwell will be the competitor.

If it's about "RT with Ampere makes you lose 38% framerate, with Ada it's 31%, RT with RDNA 3 makes you lose 45%", then I'd worry more about how much extra silicon has been pushed into a matmul engine. Which is going to be a cost tradeoff. And we know how much AMD is obsessed with saving every penny.

SolidQ · Apr 5, 2024

Ghostsonplanets said:
So I guess we should expect something Ada level at least

at least they gonna competition with Blackwell, so would be fine if higher, than ADA

adroc_thurston · Apr 5, 2024

Mahboi said:
I'm not sure what people mean by Ampere/Ada "levels".

adding BVH walker.
that's it

Mahboi · Apr 5, 2024

adroc_thurston said:
adding BVH walker.
that's it

So fixing the light bounce inefficiency. That's make everything faster, but I don't see how that'll suddenly give a "performance tier".
I don't think AMD cares to beat NV in RT performance with RDNA 4 anyway, they'll use it as a weaning gen, if they wanted to go for the throat, we'd see Navi 41.
Enjoy the smol & good GPUs and wait for RDNA 5 all over again to see a real attempt. If it is one, that is.

SolidQ · Apr 5, 2024

Mahboi said:
I don't think AMD cares to beat NV in RT performance with RDNA 4

At least they should care, about 4070ti/super level with N48

Mahboi · Apr 5, 2024

SolidQ said:
At least they should care, about 4070ti/super level with N48

Why that level specifically?
4070 Ti/Super is the "touch 4K by the fingertips" tier.
If you're already touching 4K but barely, and you get a heavy workload like RT, I don't see you being good enough for it.

adroc_thurston · Apr 5, 2024

SolidQ said:
At least they should care

Doesn't matter in phones/APUs even zilch so not really.
They'll do it for fun though.

SolidQ · Apr 5, 2024

adroc_thurston said:
Doesn't matter in phones/APUs even zilch so not really.

i'm talking about N44/N48 parts, not APU

Mahboi said:
Why that level specifically?
4070 Ti/Super is the "touch 4K by the fingertips" tier.

1080p/1440p, don't need 4k.
Because it's going to mainstream level. Look how AW2 RT kill RDNA2/3 cards, and alot people buying NV because DLSS/RT.
Persoanlly i don't care about RT and won't care until 2035

Looks how 4060 kill 7600. That why 4060 selling much more, than 7600

adroc_thurston · Apr 5, 2024

SolidQ said:
i'm talking about N44/N48 parts, not APU

same WGP, kiddo.
Very Relevant Things drive the overarching design goals.
Like APUs and exynoses.

SolidQ said:
That why 4060 selling much more, than 7600

because it haez the green sticker.
Product merits matter very little.

branch_suggestion · Apr 5, 2024

Hans Gruber said:
Is there any truth to RDNA3 being broken from the outset? Will RDNA4 work correctly upon release? Remember when they were talking about 3ghz GPU's.

7900XTX at launch.

7800XT after many months in the torture chambers.

You decide.

branch_suggestion · Apr 5, 2024

SolidQ said:
i'm talking about N44/N48 parts, not APU

1080p/1440p, don't need 4k.
Because it's going to mainstream level. Look how AW2 RT kill RDNA2/3 cards, and alot people buying NV because DLSS/RT.
Persoanlly i don't care about RT and won't care until 2035

Looks how 4060 kill 7600. That why 4060 selling much more, than 7600

Every PT implementation so far is signed and delivered under the watch of NV's engineers.
And they have all sorts of very specific hardware blocks and other gimmicky nonsense to make performance somewhat acceptable.
Just ignore the horrible denoising artifacts and light ghosting.

CouncilorIrissa · Apr 5, 2024

branch_suggestion said:
Every PT implementation so far is signed and delivered under the watch of NV's engineers.
Just ignore the horrible denoising artifacts and light ghosting.

It's one of the things that kills RTRT for me at the moment. Yes the lighting is more realistic, but image clarity is just ass. Not exactly a good tradeoff given the performance hit.

SolidQ · Apr 5, 2024

branch_suggestion said:
Just ignore the horrible denoising artifacts and light ghosting.

Still alot people involved in marketing bu****t
like this "we're buying NV because DLSS/RT/Encoder/DLAA etc"
at least AMD would be fine if they gonna make DLSS/DLAA competitor, let alone RT/Encoder
FSR 3.1 from examples gifs doesn't look promised. Reduced ghosting/shimmering, but add much more blur.
For me even at 1080p+ FSR2 is fine, just need sharpness slider. For me more means is Image Quality like Sharped/Blurred, minor shimmering doesn't bother me absolutely.

Hans Gruber · Apr 5, 2024

SolidQ said:
Still alot people involved in marketing bu****t
like this "we're buying NV because DLSS/RT/Encoder/DLAA etc"
at least AMD would be fine if they gonna make DLSS/DLAA competitor, let alone RT/Encoder
FSR 3.1 from examples gifs doesn't look promised. Reduced ghosting/shimmering, but add much more blur.
For me even at 1080p+ FSR2 is fine, just need sharpness slider. For me more means is Image Quality like Sharped/Blurred, minor shimmering doesn't bother me absolutely.

The Radeon Image Sharpening feature in Adrenaline is one of the few useful features that looks really good. There is no tax on the GPU like some of the other features available.

branch_suggestion · Apr 5, 2024

adroc_thurston said:
It's not a driver issue, AMD has a problem with resets

https://twitter.com/x/status/1775999856420536532

Just document it and it fixes itself, or something.

SolidQ · Apr 5, 2024

Hans Gruber said:
The Radeon Image Sharpening feature in Adrenaline

I'm still have GTX, next year planning change to N44 maybe, than buy RX 9700 like in 2002
At least i need x5 perfomance(currently not aviable cards on market) compared to my currently card, and then i can sitting another 7-8 years without problems.

Ajay · Apr 5, 2024

Mahboi said:
Why that level specifically?
4070 Ti/Super is the "touch 4K by the fingertips" tier.
If you're already touching 4K but barely, and you get a heavy workload like RT, I don't see you being good enough for it.

Right - with the only change being hardware assisted BVH traversal. I wouldn't say 4070 Ti is really 'touching' 4k - unless people are playing older games. Maybe the Super gets a bit closer, especially with more GDDR. But RT struggles even with my system.

Tuna-Fish · Apr 5, 2024

igor_kavinski said:
Would be cool if RDNA4 can beat at least 5070.

If the perf rumors (<7900XTX, >7900GRE) are true, the only way this happens is if nVidia really disappoints with 5070.

adroc_thurston · Apr 5, 2024

Tuna-Fish said:
the only way this happens is if nVidia really disappoints with 5070.

you're limited by core IP PPA metrics, which B100 wasn't anything interesting so far.

Ajay · Apr 5, 2024

Tuna-Fish said:
If the perf rumors (<7900XTX, >7900GRE) are true, the only way this happens is if nVidia really disappoints with 5070.

Too bad AMD won't be making a 7900 XTX MAX (or whatever) with RDNA 3.5. Even a 20% boost in raster would be better than nothing for their 'high-end'.

adroc_thurston · Apr 5, 2024

Ajay said:
Too bad AMD won't be making a 7900 XTX MAX (or whatever) with RDNA 3.5. Even a 20% boost in raster would be better than nothing for their 'high-end'.

no point, N4M's are just so much better ppa-wise

Discussion RDNA4 + CDNA3 Architectures Thread

Golden Member

Platinum Member

Senior member

Senior member

Senior member

Senior member

Senior member

Diamond Member

Senior member

Senior member

Senior member

Diamond Member

Senior member

Diamond Member

Senior member

Senior member

Senior member

Senior member

Platinum Member

Senior member

Senior member

Lifer

Golden Member

Diamond Member

Lifer

Diamond Member