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

blackangus · Jan 28, 2024

adroc_thurston said:
Somewhere late'25 or early'26.

So Far Away!

GodisanAtheist said:
Are we already on the RDNA5 hype train even though the RDNA4 hype train never really left the station?

Not hype train just wondering when we have a chance to see another AMD high-end.(ish)

adroc_thurston said:
RDNA4 uarch details themselves are very interesting but there's no point discussing them here.

This is the RDNA4 thread.... details should go here. =)

SolidQ said:
it's reffering to RDNA4? Hope not RDNA5.

Well if no high-end RDNA4 is there, then Nvidia has no pricing pressure for Blackwell will be (extra) arm and leg.

jpiniero · Jan 28, 2024

SolidQ said:
it's reffering to RDNA4? Hope not RDNA5

Blackwell probably. The top end stuff should be rather fast but also ungodly expensive.

blackangus said:
Not hype train just wondering when we have a chance to see another AMD high-end.(ish)

Don't think you will see AMD compete in the high end again until they get scalable chiplets working.

blackangus · Jan 28, 2024

jpiniero said:
Don't think you will see AMD compete in the high end again until they get scalable chiplets working.

Heres to hope!

SolidQ · Jan 28, 2024

blackangus said:
Heres to hope!

it's was seems referring "not everyone can afford 3k$ GPU". Anyway i think there should be top RDNA5

adroc_thurston · Jan 28, 2024

jpiniero said:
Don't think you will see AMD compete in the high end again until they get scalable chiplets working.

They can make a 700mm^2 GCD just fine, but who will buy it?

adroc_thurston · Jan 28, 2024

Duplicate.

branch_suggestion · Jan 28, 2024

Examining AMD’s RDNA 4 Changes in LLVM

As 2024 continues on, because time never stops, AMD has been working on their upcoming RDNA 4 architecture. Part of this involves supporting open source projects like LLVM. If done right, merging t…

chipsandcheese.com

linkgoron · Jan 29, 2024

SolidQ said:
RDNA5 is difference story, and gonna be faster than 4090, maybe even x2

So AMD will have a significantly faster card than a 1/9th cut AD102 more than three years later? It's not that great of a target IMO. The 4090 is "only" about 20% faster than a 7900XTX, and RDNA3 is very mediocre IMO. Nvidia might already be more than half-way to their 6000 series by then as well.

adroc_thurston · Jan 29, 2024

linkgoron said:
Nvidia might already be more than half-way to their 6000 series by then as well.

NV can't do much, the shrinks are gone.

jpiniero · Jan 29, 2024

adroc_thurston said:
NV can't do much, the shrinks are gone.

N3E does have a decent amount of logic shrink. But yes that's it. So one more.

adroc_thurston · Jan 29, 2024

jpiniero said:
N3E does have a decent amount of logic shrink

GPUs are like half SRAM.
Especially NVs (that chungus L1/shmem slab per SM is yes.png).

Saylick · Jan 29, 2024

adroc_thurston said:
NV can't do much, the shrinks are gone.

I wouldn’t count out Nvidia to also be working on a chiplet-based architecture moving forward.

Edit: Missed a key word in the original post.

adroc_thurston · Jan 29, 2024

Saylick said:
I wouldn’t count out Nvidia to also be working on a architecture moving forward.

Oh everyone does.
RDNA5 might be the funniest thing ever, too.

linkgoron · Jan 29, 2024

adroc_thurston said:
NV can't do much, the shrinks are gone.

Same node arch changes have shown that significant gains can also be achieved on the same node.
Kepler to Maxwell, or relatively more recently with 7nm with GCN5 to RDNA and RDNA2.

They've also shown that they're willing to increase die sizes when required (e.g. Turing arch).

adroc_thurston · Jan 29, 2024

linkgoron said:
Same node arch changes have shown that significant gains can also be achieved on the same node.

uArch gains are a non-trivial thing.

linkgoron said:
Kepler to Maxwell

That was rather expensive wrt to compute density and in general was bigger dies for less money.

linkgoron said:
or more relatively more recently with 7nm with GCN5 to RDNA and RDNA2.

RDNA1 was a shrink, 2 yeah.
But then again, bigger RDNA2 parts were worth more money so GPUs are dead without cost-effective shrinks.

linkgoron · Jan 29, 2024

adroc_thurston said:
uArch gains are a non-trivial thing.

That was rather expensive wrt to compute density and in general was bigger dies for less money.

RDNA1 was a shrink, 2 yeah.
But then again, bigger RDNA2 parts were worth more money so GPUs are dead without cost-effective shrinks.

What I meant was from Radeon 7 to RDNA1, which was not a shrink.

Again, this was mostly in reference to:

adroc_thurston said:
NV can't do much, the shrinks are gone.

Which those examples show, is not true.

adroc_thurston · Jan 29, 2024

linkgoron said:
What I meant was from Radeon 7 to RDNA1, which was not a shrink.

Those are way different parts.

linkgoron said:
Which those examples show, is not true.

It is true, Turing, which was lobotomized Volta, bloated the dies iso node to an immense degree.
Free lunch is over.

Tigerick · Jan 30, 2024

adroc_thurston said:
There was never such a rule.

Yeah it's a mobile baby die.

What? RDNA4s will be slower than R7700XT, cannot be..this is new generations of cards which supposed to be faster than RTX4080...

AMD’s new GPU reportedly matches RTX 4080 for half the price

The latest AMD Radeon 8000 rumors suggest AMD is abandoning its multi chiplet approach for RDNA 4 graphics cards, and leaving the high end to Nvidia.

www.pcgamesn.com

Also, you said they are not going to be called RX8000?

jpiniero · Jan 30, 2024

Tigerick said:
What? RDNA4s will be slower than R7700XT, cannot be..this is new generations of cards which supposed to be faster than RTX4080...

Their source is MLID, which yeah.

SteinFG · Jan 30, 2024

Tigerick said:
What? RDNA4s will be slower than R7700XT, cannot be..this is new generations of cards which supposed to be faster than RTX4080...

The small $300 gpu (8600 series) will be slower than 7700 XT, at most on par with 6700XT.

The big one ($500+) is probably as fast as 4070 TiS , just my guess. Based on the current rumor of Navi44 having 32CU, and Navi48 having 64CU.

Tigerick said:
Also, you said they are not going to be called RX8000?

Don't know who keeps saying that, but it most likely is RX 8000, unless AMD goes silly like in the vega days and names it something like RDNA4 64, or RDNA IV, or RX Fury

Glo. · Jan 30, 2024

Tigerick said:
Also, you said they are not going to be called RX8000?

What GPUs are called is meaningless.

AMD can make complete rebranding of the GPU lineup into something like Radeon A600, A700, A800, A900. Then Radeon B600, B700, B800, B900. For iGPUs it would be then: A30, A50, A60, A80, B30, B50, B60, B80.

Why it matters?

Because we are entering era in which entry level GPUs in mobile lineup from AMD will be powerful enough to be entry level GPUs. Rebranding of the whole lineup and simplifying the naming would help AMD make AMD better marketing.

adroc_thurston · Jan 30, 2024

Tigerick said:
this is new generations of cards which supposed to be faster than RTX4080

No?

Tigerick said:
Also, you said they are not going to be called RX8000?

Two cheapo dies is all you get there, yes.

Timorous · Jan 30, 2024

SteinFG said:
The small $300 gpu (8600 series) will be slower than 7700 XT, at most on par with 6700XT.

The big one ($500+) is probably as fast as 4070 TiS , just my guess. Based on the current rumor of Navi44 having 32CU, and Navi48 having 64CU.

Don't know who keeps saying that, but it most likely is RX 8000, unless AMD goes silly like in the vega days and names it something like RDNA4 64, or RDNA IV, or RX Fury

For those guesses to be true it would mean AMD have missed with RDNA 4 as well although the 64CU guess is probably closer to the mark with the typically large pre release error bars.

An OCd 32CU 7600XT can roughly match a 6700XT.

An OC'd 60 CU 7800XT is right on the heels of a 4070Ti.

In both cases the core clock is around 3Ghz up from the measured stock average of 2.76Ghz for the 7600XT and 2.54Ghz for the 7800XT.

I suspect the smaller part will land between the 6700XT and 7700XT which is a pretty large range granted but it really does depend on clocks and if there is an IPC uplift.

I suspect the larger part will land between 7900XT and 7900XTX depending on clocks and IPC.

I doubt the die size for the larger part is much bigger than 250mm otherwise there is no way AMD will sell it for $600 or less. The smaller part would probably need to be 150mm or so to hit a $300-$350 price point.

EDIT to add.

For a 150mm or smaller N44 with similar density to Ada on N4 we are talking between 13.3M xtors and 18M xtors. That gives a bit of room to bump clocks and improve RT.

For a 250mm N48 that would be around 30M Xtors which is more than the 28.1M N32 has with the MCDs. Again with the xtor saving from the interconnects there is probably a bit of room to bump clocks and improve RT in that kind of budget so I guess it looks doable from that kind of perspective and N4 may be a small bit desner than the modified N5 node NV use.

If AMD hit their targets it does look like AMD can release a $600 part with 7900XTX performance while maintaining desired margins. Lets home RDNA 4 does not miss.

MrTeal · Jan 30, 2024

jpiniero said:
Their source is MLID, which yeah.

Wait, I've been a little out of it. Do you mean MLID was wrong and N31 isn't +80% to +120% (or optimistically +130-140%) over N21?
My world is rocked.

misuspita · Jan 30, 2024

I hope they can manage the power consumption, cause now it's ridiculous.

I want to buy AMD because the prices are better than nVidia, at the same perf, but if you want a sff, it's no contest, a 4060ti is smaller and more silent because they consume 50-100w lower than their counterparts. And I want a sff, small and quiet

Discussion RDNA4 + CDNA3 Architectures Thread

Golden Member

Member

Lifer

Member

Senior member

Platinum Member

Platinum Member

Member

Platinum Member

Platinum Member

Lifer

Platinum Member

Diamond Member

Platinum Member

Platinum Member

Platinum Member

Platinum Member

Platinum Member

Senior member

Lifer

Senior member

Diamond Member

Platinum Member

Golden Member

Diamond Member

Senior member