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

Bryo4321 · Mar 3, 2025

SolidQ said:
i'd assuming it's with FG results, and we know AMD FG is way faster, than NV FG

Mmm yeah that’s true too. I’m looking forward to the 5th.

branch_suggestion · Mar 3, 2025

adroc_thurston said:
boring and niche, if you're spending this R&D you do a 500 on 500, N2 on N4C and you win.

adroc_thurston said:
Oh I'm not advocating for N4c sequel.
Something far far simpler instead.

WoW slab really is the best of both worlds.
No need for the API breaking hacks to scale across multiple frontends but you can scale perf to a similar degree. (you can still implement some of the ideas if they scale well enough)
Also think of how easy it is to market a beyond reticle stack, X3D branding is a legit marketing weapon now and you can tout lots of big, shiny numbers that are > everyone else.
It is such a simple product manager goal, give 1kmm^2, N2 on N3/4 SoIC-WoW and let them cook.
Use some of that Sony/MS console R&D money and build something that will lift all boats.
I really hope AMD sees the situation is very winnable and Mr Hyunh has the mandate to do what is best for client compute. Even if they start it a bit late compare to the mono RDNA5 parts it would still be ready by Computex 2027. Oh, and don't undersize the mono parts this time, 1.5x the N44/N48 CUs and GDDR7 with the new uArch should be good enough for the normies.

Kepler_L2 · Mar 3, 2025

branch_suggestion said:
WoW slab really is the best of both worlds.
No need for the API breaking hacks to scale across multiple frontends but you can scale perf to a similar degree. (you can still implement some of the ideas if they scale well enough)
Also think of how easy it is to market a beyond reticle stack, X3D branding is a legit marketing weapon now and you can tout lots of big, shiny numbers that are > everyone else.
It is such a simple product manager goal, give 1kmm^2, N2 on N3/4 SoIC-WoW and let them cook.
Use some of that Sony/MS console R&D money and build something that will lift all boats.
I really hope AMD sees the situation is very winnable and Mr Hyunh has the mandate to do what is best for client compute. Even if they start it a bit late compare to the mono RDNA5 parts it would still be ready by Computex 2027. Oh, and don't undersize the mono parts this time, 1.5x the N44/N48 CUs and GDDR7 with the new uArch should be good enough for the normies.

Even the monolithic gfx13 stuff has the new CP/GE distributed logic so that's not a problem. If they go with chiplets for dGPUs I expect them to just re-use one ~100mm² MID across the stack with ~150/250/350mm² GCDs, which should be enough for a 2x perf increase over N48 even on N3P.

branch_suggestion · Mar 3, 2025

Kepler_L2 said:
Even the monolithic gfx13 stuff has the new CP/GE distributed logic so that's not a problem.

Very cool, wonder how it will scale between implementations.

Kepler_L2 said:
If they go with chiplets for dGPUs I expect them to just re-use one ~100mm² MID across the stack with ~150/250/350mm² GCDs, which should be enough for a 2x perf increase over N48 even on N3P.

Yeah how to implement the IO is the sticky point, you can have a WoW stack with a PHY that connects to the MID via fanout. Or just leave enough room between all the MALL and G7 PHYs, but this might not leave enough space, especially if they do smaller stacks.
MID approach does make sense only if it scales across the stack, could work for both 2D and 3D paradigms.
350mm^2 of just compute, cache and memory PHYs could get a bit over 1.5x on N3P, 2x would require more area or 3D stacking.

marees · Mar 3, 2025

Timorous said:
Quick and dirty is that the 9070XT has about 50% more performance than the 7800XT with a very similar config.

Doing the same at 96CUs means 7900XTX + 50% with a very similar config.

So sure the numbers are super rough but even if error bars are +/-10% it gives you a ballpark.

AMD's claim is 40% raster increase per CU & doubling of RT

gaav87 · Mar 4, 2025

Ok so that result was with Frame gen vs 4080s and 4090 also with frame gen to point out 9070xt framegen is better
It was stock msrp model no pl/oc vs oced 4080s and 4090 xD
No frame gen its -10fps vs 4080s (high rt) 9070xt(very high rt)

uzzi38 · Mar 4, 2025

adroc_thurston said:
RDNA4 clocks vary a lot more depending on how hard you hammer the memory.
Remember, this thing is not like the others.

Have you been shown the funny?

(IYKYK)

But generally speaking on the topic of overclocking, trust me when I say shader clocks absolutely won't be limiting factor. Power budget or the other silicon in there absolutely will be.

gaav87 · Mar 4, 2025

So yeah anyway hows ur guys morning all good ?
Cause my 6800xt just go crushed by 3.63x vs 9070xt same settings in RT

DownTheSky · Mar 4, 2025

So we're not there yet. Performance looks similar to a 4080. I saw some videos on youtube and it's enough for RT medium aka normal RT. For path tracing, even 1080p you need a 5090.

gaav87 · Mar 4, 2025

DownTheSky said:
So we're not there yet. Performance looks similar to a 4080. I saw some videos on youtube and it's enough for RT medium aka normal RT. For path tracing, even 1080p you need a 5090.

Good enough for a 599$ card that competes vs 1000/1200$ 4080 xD even if -10% xD

soresu · Mar 4, 2025

adroc_thurston said:
Pretty much everything RTRT is that.
Trickier compute shaders are that too.

Any ubershader in an emulator like Dolphin would almost certainly benefit from it.

soresu · Mar 4, 2025

DownTheSky said:
For path tracing, even 1080p you need a 5090

Unlike the fairly stale and nearly tapped out state of raster gfx, the software side of GPU RT/PT is still in its wild west period of discovery, with a lot of potential low hanging fruit remaining to improve performance with the same hardware without using super resolution haxx.

Point of fact, the software side has improved performance many times more than the hardware side since Turing was announced in 2018.

inquiss · Mar 4, 2025

uzzi38 said:
Have you been shown the funny?

(IYKYK)

But generally speaking on the topic of overclocking, trust me when I say shader clocks absolutely won't be limiting factor. Power budget or the other silicon in there absolutely will be.

Will whatever you're hinting at be disclosed tomorrow, do you think?

uzzi38 · Mar 4, 2025

inquiss said:
Will whatever you're hinting at be disclosed tomorrow, do you think?

If the person I talked to doesn't include it in their review, it should be perfectly fine for me to post the screenshot here anyway methinks. It's something anyone could try on N48.

SolidQ · Mar 4, 2025

gaav87 said:
No frame gen its -10fps vs 4080s (high rt) 9070xt(very high rt)

isn't that difference high vs very high?

coercitiv · Mar 4, 2025

more than all all 50 series combined does have a ring to it:

I would have to eat some crow for the mood of my previous posts, but they... would be happy to oblige! Let's see if the MSRP sticks.

Panino Manino · Mar 4, 2025

I'm feeling hard the end to those Tech Deep Dives by The Anandtech that were masterful written in a way that even dummies like me could understand.

Any alternatives?

gaav87 · Mar 4, 2025

SolidQ said:
isn't that difference high vs very high?

Its 250% faster than 7900xtx same settings xD

Josh128 · Mar 4, 2025

uzzi38 said:
If the person I talked to doesn't include it in their review, it should be perfectly fine for me to post the screenshot here anyway methinks. It's something anyone could try on N48.

Because there'll likely be like 50+ reviews tomorrow, and I'll only be skimming through a few, and if video reviews, skimming through them, please come back and post what you are talking about specifically because theres a 99% chance it will be missed.

gaav87 · Mar 4, 2025

30min please don't ban me im, a good boy

sandorski · Mar 4, 2025

"5xxx" would also work. Apparently, we'll soon know.

gaav87 · Mar 4, 2025

sandorski said:
"5xxx" would also work. Apparently, we'll soon know.

Just for you.
Last one i promis.

sandorski · Mar 4, 2025

This could be the magic moment, AMD winning the GPU Generation by Volume alone. RDNA5 may be the next part of the combo that puts them on par with NVidia unlike we have ever seen.

linkgoron · Mar 4, 2025

sandorski said:
This could be the magic moment, AMD winning the GPU Generation by Volume alone. RDNA5 may be the next part of the combo that puts them on par with NVidia unlike we have ever seen.

AMD missing this huge opportunity by canning the big RDNA4 is nuts. RDNA4 looks like an actually successful, competitive, good arch with great engineering and they just squandered this huge opportunity. Who knows if RDNA5 will have the same chance vs Rubin or whatever the next NV arch is.

Tup3x · Mar 4, 2025

coercitiv said:
more than all all 50 series combined does have a ring to it:

View attachment 118838

I would have to eat some crow for the mood of my previous posts, but they... would be happy to oblige! Let's see if the MSRP sticks.

How much is nothing * n?

Discussion RDNA4 + CDNA3 Architectures Thread

Golden Member

Member

Senior member

Senior member

Senior member

Senior member

Senior member

Platinum Member

Senior member

Senior member

Senior member

Diamond Member

Diamond Member

Senior member

Platinum Member

Golden Member

Diamond Member

Golden Member

Senior member

Senior member

Senior member

No Lifer

Senior member

No Lifer

Platinum Member

Golden Member