Discussion RDNA4 + CDNA3 Architectures Thread

[DisEnchantment]

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

 

[adroc_thurston]

Hmmm GFX12 are showing up.

Yeah, parts should be H1'24 after all.

 

[PJVol]

Not happening (at least not anymore).

So, the top rdna4 performer is somewhere between n31 and n32?

 

[adroc_thurston]

So, the top rdna4 performer is somewhere between n31 and n32?

Probably.
They're tiny dies.

 

[DeathReborn]

wccftech warning
Well this is (could be) troubling for current/future RDNA & CDNA architectures, at least for China.

Dell Prohibits Sales of AMD Radeon RX 7900 XTX, 7900 XT, PRO W7900 & Upcoming MI300 GPUs In China

AMD Radeon RX 7900 XTX, 7900 XT, PRO W7900 & MI300 GPUs sales are now prohibited by Dell in China following more aggressive US bans.

wccftech.com

 

[Joe NYC]

Eh the difficult uArch decisions are roughly the same for both; just less requirements for the specific way of doing stuff correctly for CDNA.

They've actually undercommited.

If Strix Halo and high performing APU are the way for client (especially notebooks), then HBM will look increasingly attractive - except the cost.

There just needs to be some new and elegant ways to get the high cost out of HBM...

 

[igor_kavinski]

They're tiny dies.

Harder to cool they will be? How high will be the clocks? 3.5 GHz? Power consumption lower than RDNA3?

 

[Timorous]

Probably.
They're tiny dies.

This is exactly what I don't get.

We have the canned super complex multi compute part at the halo end. We have the small monolithic dies at the low and middle end. What happened to the upper middle and high end parts? Why were they canned?

 

[igor_kavinski]

Why were they canned?

Could make sense if three months after RDNA4 launch, we get the RDNA5 Halo part, several months before Geforce 5090. That would be a masterful stroke by AMD. Should translate to some good high margin sales.

 

[jpiniero]

What happened to the upper middle and high end parts? Why were they canned?

Guessing anything higher than the low end monolithic parts was using the chiplets.

So, the top rdna4 performer is somewhere between n31 and n32?

I still think that's too much to ask. Just the 7700 XT is 51% faster in raster than the 7600. Maybe in RT but there would have to be big improvements there.

 

[Timorous]

Guessing anything higher than the low end monolithic parts was using the chiplets.

rdna3 uses chiplets.

big difference in packaging cost and complexity between that design and the mi300 like halo part that seems to be canned.

Could make sense if three months after RDNA4 launch, we get the RDNA5 Halo part, several months before Geforce 5090. That would be a masterful stroke by AMD. Should translate to some good high margin sales.

Maybe, I doubt 4-5 will be that fast but who knows.

 

[jpiniero]

rdna3 uses chiplets.

big difference in packaging cost and complexity between that design and the mi300 like halo part that seems to be canned.

Correct but seemingly AMD only planned on doing the scalable chiplets design without any real fallback if they didn't make it work in time.

 

[Timorous]

Correct but seemingly AMD only planned on doing the scalable chiplets design without any real fallback if they didn't make it work in time.

Still does not make sense. Cost structures and packaging volume would be limiting to where that design can be used. Super halo tier is low volume and very expensive so it makes sense. Standard high end tiers in the $600 region don't at all because the RDNA3 layout is far cheaper and is there enough capability to package such a design and actually have reasonable volume of parts at those price points? Seems unlikely to me given how new this tech is.

 

[jpiniero]

Still does not make sense. Cost structures and packaging volume would be limiting to where that design can be used. Super halo tier is low volume and very expensive so it makes sense. Standard high end tiers in the $600 region don't at all because the RDNA3 layout is far cheaper and is there enough capability to package such a design and actually have reasonable volume of parts at those price points? Seems unlikely to me given how new this tech is.

AMD doesn't sell enough to justify 3 different types of products, I'm guessing.

You probally won't see any sub-$1000 Blackwell parts until 2025... so as long as RDNA5 is realistic for like Mid 25 it's NBD.

 

[PJVol]

Could make sense if three months after RDNA4 launch, we get the RDNA5 Halo part, several months before Geforce 5090. That would be a masterful stroke by AMD

It would, but some speculates it's still usual six quarter cadence from gen to gen

 

[adroc_thurston]

Harder to cool they will be? How high will be the clocks? 3.5 GHz? Power consumption lower than RDNA3?

I've no idea, just yet.

What happened to the upper middle and high end parts?

They were still Navi4c.

I still think that's too much to ask

Nah, it's a shrink and a new uarch.

Correct but seemingly AMD only planned on doing the scalable chiplets design without any real fallback if they didn't make it work in time.

The fallback is killing the top-end parts and moving on.
It's that simple.

so as long as RDNA5 is realistic for like Mid 25 it's NBD.

It's nowhere close to mid'25.

 

[linkgoron]

Hmmm GFX12 are showing up.

[AMDGPU] Define new targets gfx1200 and gfx1201 by jayfoad · Pull Request #73133 · llvm/llvm-project

Define target names and ELF numbers for new GFX12 targets gfx1200 and gfx1201. For now they behave identically to GFX11.

github.com

Nothing special yet.

Just for a general time frame, at least for RDNA2 and RDNA3. For RDNA3, gfx11 was added on April 29, 2022 and the cards were launched Dec. 13 2022, so around 8 months. RDNA2 (GFX1030) was added on Jun 16, 2020 and released November 18, 2020, so around five months.

 

[SolidQ]

Just for a general time frame, at least for RDNA2 and RDNA3. For RDNA3, gfx11 was added on April 29, 2022 and the cards were launched Dec. 13 2022, so around 8 months. RDNA2 (GFX1030) was added on Jun 16, 2020 and released November 18, 2020, so around five months.

mlid was saying about it long time ago, like 5700XT release "Summer"

what "leaks" we're have now
RDNA4

RGT Leak
・Radeon RX 8800 Series N31 Refresh
・Radeon RX 8700 Series based on Navi48 32WGP(64CU)
・Radeon RX 8600 Series based on Navi44 20WGP(40CU)

・RDNA 4 highest end SKU about on par with RX 7900GRE
・RDNA 4 Ray Tracing averages 10-20 percent faster than RDNA 3
・Designed to be energy efficient. cheap to produce
・Monolithic dies
-------------
MLID leak
1) it is currentry projected to beat 7900XT
2) We aren't sure if it can beat the 4080 SUPER
3) targeting "midrange" so likely $400-$600

...and i am stil not sure on the exact release
date nor if it'll get close to Ada in RT/Raster

People can believe in both leak or none

 

[Tigerick]

Just for a general time frame, at least for RDNA2 and RDNA3. For RDNA3, gfx11 was added on April 29, 2022 and the cards were launched Dec. 13 2022, so around 8 months. RDNA2 (GFX1030) was added on Jun 16, 2020 and released November 18, 2020, so around five months.

Thanks for the timeline info, I suspect AMD would most likely launch RDNA4 cards before even five months timeframe, that would put the date earlier than April 2024.

I have discussed Shader ISA before as shown table below:- Every gen of RDNA cards has different Shader ISA but there isn't much difference in performance between RDNA2 and RDNA3, including ray tracing performance. Shader ISA appears to be aligned to the new process node AMD is using; thus I believe change of Shader ISA in RDNA4 is more aligned with N4P process, AMD most likely re-arrange some Shader design to fit in more EUV layers thus make the die size smaller. Anybody wants to guess how small the die size of N44?

AMD has to put every resource on the design of RDNA5, do you think they will spare extra team for RDNA4 design? That's why I think RDNA4 GPUs are just die shrink of N32s with integrated Infinity Cache, so no I don't think RDNA4 cards will outperform N32 cards especially without GDDR7 insight...

@TESKATLIPOKA , @DisEnchantment

When I was referring RDNA3/4 difference, I meant the performance increase in rasterization which is pretty meh from past RDNA2 with the same amount of CU. That's why I don't expect any changes on raster performance which is mostly correct. What I am missing is new features upgrade, namely ray tracing performance. As shown on Cyberpunk 2077, AMD managed to increase ray-tracing performance by almost 50% with RX7700XT which is pretty good, but still not up to nVidia level.

With new Shader ISA gfx12 and possible 3rd generation RT cores, if AMD manages to pump up additional 50% fps on ray tracing, then AMD might be reaching NV's level if the amount of CU/SM is in parity...I would say this is the biggest uncertainty on N43/N44, we shall see...

I am more convinced that N43 going to have 40CU with 12GB GDDR6 support, along with N44 which should feature 32CU and 8GB GDDR6 support. Here comes the table with AMD GPU lineup for 2024:- Pending confirmation. PS: I also throw in old RDNA2 and upcoming Strix Halo for comparison:-

Model	SRP	Codename	Process	TDP	Die Size (mm2)	Shader ISA	RT Cores	CU	IC	Memory	Memory BW	TPU Perf
RX 6650 XT	$399	RDNA2 N23	N7	176 W	237	GFX1032	1st Gen	32	32MB	128-bit GDDR6	280.3 GB/s	- 1%
RX 7600	$269	RDNA3 N33	N6	165 W	204	GFX1102	2nd Gen	32	32MB	128-bit GDDR6	288 GB/s	Base
RX 7600 XT ?	$299?	RDNA4 N44	N4P	?	?	GFX12?	3rd Gen	32	32MB	128-bit GDDR6	288 GB/s	?
RX 6750 XT	$549	RDNA2 N22	N7	250 W	335	GFX1031	1st Gen	40	96MB	192-bit GDDR6	432 GB/s	+ 19%
RX 7700 ?	$399?	RDNA4 N43	N4P	?	?	GFX12?	3rd Gen	40	48MB	192-bit GDDR6	432 GB/s	?
Strix Halo	NA	RDNA3.5	N3E	?	?	GFX115?	?	40	32MB	256-bit LPDDR5X-8533	272 GB/s	?
RX 7700 XT	$449	RDNA3 N32	N5 + N6	245 W	346	GFX1101	2nd Gen	54	48MB	192-bit GDDR6	432 GB/s	+ 38%
RX 7800 XT	$499	RDNA3 N32	N5 + N6	263 W	346	GFX1101	2nd Gen	60	64MB	256-bit GDDR6	620.8 GB/s	+ 67%

 

[adroc_thurston]

Shader ISA appears to be aligned to the new process node AMD is using

no.

AMD has to put every resource on the design of RDNA5

Did you mean MI400?

That's why I think RDNA4 GPUs are just die shrink of N32s

Not in any way, shape or form related to any N3x part.

 

[Tigerick]

no.

Did you mean MI400?

Not in any way, shape or form related to any N3x part.

Guess we will know more when AMD make the changes in github library. Do you think AMD will launch RDNA4 cards during CES2024?

 

[adroc_thurston]

Do you think AMD will launch RDNA4 cards during CES2024?

no lol

 

[Bigos]

28nm Tonga and Fiji and 14nm Polaris were both gfx8.

14nm Vega 10 and 7nm Vega 20 were both gfx9.

Both gfx10 and gfx10.3 silicon (RDNA1 + RDNA2 dGPUs) were on 7nm.

There is no correlation between ISA and the process node. It does not make sense whatsoever.

 

[SolidQ]

New WGP?

https://twitter.com/x/status/1727630408081420430

 

[branch_suggestion]

This is exactly what I don't get.

We have the canned super complex multi compute part at the halo end. We have the small monolithic dies at the low and middle end. What happened to the upper middle and high end parts? Why were they canned?

Every part above the mono parts uses the same tech as the halo part, so similar validation.

Could make sense if three months after RDNA4 launch, we get the RDNA5 Halo part, several months before Geforce 5090. That would be a masterful stroke by AMD. Should translate to some good high margin sales.

No, RDNA4 is competitive against Blackwell midrange and comes out first, RDNA5 obliterates Blackwell and comes out way after.

so as long as RDNA5 is realistic for like Mid 25 it's NBD.

H1 2026, nearly 2 years after RDNA4

Guess we will know more when AMD make the changes in github library. Do you think AMD will launch RDNA4 cards during CES2024?

Computex most likely.

 

[jpiniero]

No, RDNA4 is competitive against Blackwell midrange

Depends on your definition of midrange these days, lol.

And not much has leaked on Blackwell. So while I am very pessimistic about the gen-on-gen perf improvement outside of the $$$ stuff it is too early to be definitive about it.