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

 

[SolidQ]

I find only 40CU unlikely.

Maybe it's going 60CU N43? N22 to N32 40CU up to 60CU. N44 = 32CU?

 

[TESKATLIPOKA]

Maybe it's going 60CU N43? N22 to N32 40CU up to 60CU. N44 = 32CU?

I find only 32CU too little. Strix Point will have 1/2 of that, but maybe between RDNA3.5 vs RDNA4 there will be a significant difference in performance.

 

[DisEnchantment]

I find only 40CU unlikely. Just to be on par with N32, It would need 50% higher clock.

Well if the supposed 5700XT tier RDNA4 rumor is true then it wont be on par with N32.

If N43 is indeed a 40CU GPU, then it makes sense based on current product lineups. N43 is not designed for replacing current N32, it is going to replace 6700XT/6750XT 12GB and I believe it can only support 192-bit memory bus with 12GB GDDR6 as well. And I believe it will be made by TSMC's N4P process with monolithic design, not N3E as people hope so.

As for pricing, @adroc_thurston has hinted it would cost $399 at best which make sense with smaller die size and monolithic design. I am hoping AMD would go lower like $349 with OC version cost up to $399. There is going to have some overlapping between N43 (likely going to model as Radeon RX7700 non-XT series) and 7700XT especially with OC version of 7700; but I believe AMD would focus more on RX7800XT based on current pricing.

If N43 is indeed a 40CU GPU, then it makes sense based on current product lineups. N43 is not designed for replacing current N32, it is going to replace 6700XT/6750XT 12GB and I believe it can only support 192-bit memory bus with 12GB GDDR6 as well. And I believe it will be made by TSMC's N4P process with monolithic design, not N3E as people hope so.

I am just wondering why replace lower end/midrange parts with new architecture? I would expect current lower midrange 7700XT next year will be coming down to 399USD levels. Once NV adjust 4060 prices it will come down.
Which means they don't need RDNA4 at all for next year. They could just create a Strix successor and put RDNA4 inside for 16CUs and call it a day with RDNA4.

 

[Tigerick]

Well if the supposed 5700XT tier RDNA4 rumor is true then it wont be on par with N32.



I am just wondering why replace lower end/midrange parts with new architecture? I would expect current lower midrange 7700XT next year will be coming down to 399USD levels. Once NV adjust 4060 prices it will come down.
Which means they don't need RDNA4 at all for next year. They could just create a Strix successor and put RDNA4 inside for 16CUs and call it a day with RDNA4.

Good question, I believe there won't be much difference between RDNA3 and RDNA4 as people hopes; remember true RDNA4's architecture with chiplet design is cancelled and merged into RDNA5 which supposedly release by end of 2025. N43 and N44 sound like extension of RDNA3 rather than original plan of RDNA4.

Another reason might be BOM especially with RX7700XT, AMD is wasting one MCD and die area with much higher TDP. With monolithic design, AMD could use much lower power and bring down the cost as well.

With chiplet design, AMD seems like not able to bring the power consumption down, that's why we don't see N32 mobile GPU in sights...thus AMD needs to create monolithic version of GPU if they want tackle notebook platform especially with upcoming Zen5's platform. I am not talking about Strix Halo but Fire Range with 16-core and X3D as well...

 

[jpiniero]

I find only 40CU unlikely. Just to be on par with N32, It would need 50% higher clock.

I would think we are talking about N33 and N24 replacements.

 

[branch_suggestion]

I think there isn't enough public info about the GFX12 uArch changes to really elaborate much on config.
Still, if AMD can make a $399 part that obsoletes N32 and a $249 part that obsoletes N33, RDNA4 is easily worth it.

I am just wondering why replace lower end/midrange parts with new architecture?

Because perf/$, both for the consumer and AMD.

Which means they don't need RDNA4 at all for next year. They could just create a Strix successor and put RDNA4 inside for 16CUs and call it a day with RDNA4.

The IP has been developed, making some midrange parts that crush what NV currently has with a nice headstart over Blackwell is a very nice position for AMD.
Think RV770 vs G92/GT200.

Good question, I believe there won't be much difference between RDNA3 and RDNA4 as people hopes; remember true RDNA4's architecture with chiplet design is cancelled and merged into RDNA5 which supposedly to come out by end of 2025. N43 and N44 sound like extension of RDNA3 rather than original plan of RDNA4.

RDNA4 is GFX12, there are big changes regardless of the packaging.

Another reason might be BOM especially with RX7700XT, AMD is wasting one MCD and die area with much higher TDP. With monolithic design, AMD could use much lower power and bring down the cost as well.

Small enough parts are best mono, yeah. N32 is just big enough that the layout they went for was better, but it will never be seen again as it was mainly a test run for the new fanout that Zen6 and other stuff will be using.

With chiplet design, AMD seems like not able to bring the power consumption down, that's why we don't see N32 mobile GPU in sights...thus AMD needs to create monolithic version of GPU if they want tackle notebook platform especially with upcoming Zen5's platform. I am not talking about Strix Halo but Fire Range with 16-core and X3D as well...

RDNA3 power issues are at the core level, chiplet power hit is ~10w for N32 vs a mono fabric. AMD has been slow on getting mobile parts out for very different reasons.

 

[eek2121]

Which then confirm no high end right? because if the "next-gen" doesn't best the previous gen, that by definition means no high end. Releasing it of course still makes sense in terms of power user and features.

32-40CU is too little for a next gen chip even If they scrapped High-end, this could be the slower chip.
Considering they have strix Halo with 40CU IGP, I would expect 48CU full and 40CU as a cutdown for the weaker chip.
Stronger could be 72CU and 60CU cutdown.

I got ~171mm2 for a 48CU(3SE) RDNA3 5nm GCD, based on what Locuza posted some time ago.
If I applied N3E density improvement to It.

At worst, It could be 132mm2(+30% improvement) and at best 114mm2(+50%).
Of course, we don't know how much more transistors RDNA4 uses vs RDNA3 for comparable specs, so this estimate is inaccurate.

To your argument, If It's not better to just port RDNA3 N32 to N4P and save some money in the process. That's very hard to tell when we know absolutely nothing about RDNA4.
Considering that they plan to release It, It should still have some advantage compared to RDNA3.
I personally expect significantly higher clocks. With 48CU, It would need ~25% to be on par with 7800XT.

The problem I see is not this GCD per se, but MCD and Vram. For a 48 CU at higher clock you would need comparable BW to N32.
4 MCDs would have comparable size to that GCD, even If they cost less to produce.

Does anyone have anything bad to say about the 5700XT? Didn’t think so.

AMD could stick to 40CUs if they fix whatever issues are dogging RDNA3. I could easily see an increase of 50% perf/watt over RDNA3.

Would 40CUs compete at the top end? Of course not. Would it be competitive at midrange? yep. Very competitive. I guess we will find out when it launches…

EDIT: I personally think they will go bigger, FWIW. Probably 60CUs.

IMO they should rework RDNA3 into a single 60CU monolithic design on N4P, fixing bugs and optimizing the architecture.

 

[TESKATLIPOKA]

I would think we are talking about N33 and N24 replacements.

Instead of replacing higher performing GPUs they replace the bottom? What's the point in that?

 

[TESKATLIPOKA]

Does anyone have anything bad to say about the 5700XT? Didn’t think so.

AMD could stick to 40CUs if they fix whatever issues are dogging RDNA3. I could easily see an increase of 50% perf/watt over RDNA3.

Would 40CUs compete at the top end? Of course not. Would it be competitive at midrange? yep. Very competitive. I guess we will find out when it launches…

EDIT: I personally think they will go bigger, FWIW. Probably 60CUs.

IMO they should rework RDNA3 into a single 60CU monolithic design on N4P, fixing bugs and optimizing the architecture.

Your 50% perf/W improvement could mean many things, be more specific.

Not even sure what level of performance you consider as midrange.

 

[SolidQ]

I find only 32CU too little

Do you think N44 is 40CU? N41 was supposed go to 270CU?

Probably 60CUs.

I'm think same for N43 = 60CU. If N22 to 32 going 40 to 60CU, then possible N32 going to N43 with higher clock and much better RT

 

[TESKATLIPOKA]

Good question, I believe there won't be much difference between RDNA3 and RDNA4 as people hopes; remember true RDNA4's architecture with chiplet design is cancelled and merged into RDNA5 which supposedly release by end of 2025. N43 and N44 sound like extension of RDNA3 rather than original plan of RDNA4.

There is no difference between RDNA4 monolith or chiplet.
The same as N33 vs N32 or N31.

With chiplet design, AMD seems like not able to bring the power consumption down, that's why we don't see N32 mobile GPU in sights...thus AMD needs to create monolithic version of GPU if they want tackle notebook platform especially with upcoming Zen5's platform. I am not talking about Strix Halo but Fire Range with 16-core and X3D as well...

There is no proof that chiplet design is the cause of missing laptop N32.

No one would pair RDNA4 monolith with Fire Range, If It's not fast enough.
And AMD was always doing bad in laptops, I doubt they would especially make a monolith mainly for laptop use.

 

[DisEnchantment]

Because perf/$, both for the consumer and AMD.

This also applies to high end parts.

Anyway this whole RDNA4 at 5700XT tier started from twitter chatter. We shall see what comes out from the tieba rumor or from the alleged AMD employee debunking.

 

[TESKATLIPOKA]

Do you think N44 is 40CU? N41 was supposed go to 270CU?


I'm think same for N43 = 60CU. If N22 to 32 going 40 to 60CU, then possible N32 going to N43 with higher clock and much better RT

Who knows?
N4C(N41 or N4X) had 270CU in 9 SEDs. Or 30CU pre SED.
If you think about It, then even 40CU looks very little in comparison. That's only 14.8% of compute, ignoring frequency difference.
But N4C was a GPU using 9 Shader Engine chiplets, It could be scaled down easily.

edit:
What I want to know is what memory configuration was planned to be used with It.
If It needed the same BW requirement as N31, then It would need ~3x higher BW.
Even with 512-bit GDDR7 you would get only 2.1x better BW. So the rest would have to be compensated with more cache.

 

[PJVol]

There is no proof that chiplet design is the cause of missing laptop N32.

Yeah, in case of n33 vs n23 the improvement was so huge that they even had to omit the "xt" moniker in 7600, lol

 

[TESKATLIPOKA]

Yeah, in case of n33 vs n23 the improvement was so huge that they even had to omit the "xt" moniker in 7600, lol

With missing clocks, It's not surprising It ended badly.
We will see with Strix, If there is any gain in clock speed or not.

 

[PJVol]

With missing clocks

Oh... what a surprise for those thinking the chiplets are the culprit.
I even dare to assume they're among those who "silently predicted" 50% perf/w failure unlike most of the rumormongers rgt, mlid, et al condemned here in this forum, who claimed 2.5x performance for top to top

 

[TESKATLIPOKA]

Oh... what a surprise for those thinking the chiplets are the culprit

Chiplet at best is not as power efficient as monolith, but the difference shouldn't be that big in my opinion. This looks to me like a problem with WGPs.

 

[jpiniero]

Instead of replacing higher performing GPUs they replace the bottom? What's the point in that?

Because the high end chiplet models got cancelled. I am assuming that the N32 replacement would have been the single chiplet.

 

[TESKATLIPOKA]

Because the high end chiplet models got cancelled. I am assuming that the N32 replacement would have been the single chiplet.

If N32 replacement was already a chiplet, then the 2 surviving monoliths would perform worse than N32.
One thing would be replacing N33 with 20-30% faster chip, but what's the point of replacing N24?
Performance would be weak and I am pretty skeptical about price being low enough. And what about Vram? With anything less than 8GB Vram It would be DOA.

 

[maddie]

Because the high end chiplet models got cancelled. I am assuming that the N32 replacement would have been the single chiplet.

Now the question. Were the full range 1,2,3 chiplets or 1,2,4, bottom to top.

 

[TESKATLIPOKA]

Now the question. Were the full range 1,2,3 chiplets or 1,2,4, bottom to top.

Top was supposedly this.

9 SEDs with 270CU in total, at least It supposedly contained 30CU per SED.
I have to wonder where were ROPs and frontend placed?

 

[soresu]

Top was supposedly this.

9 SEDs with 270CU in total, at least It supposedly contained 30CU per SED.
I have to wonder where were ROPs and frontend placed?

Nearly 3x N31 CU count?!!

They must have needed to drop the clocks back to ≤2 Ghz to make that work 🔥⚡

 

[TESKATLIPOKA]

Nearly 3x N31 CU count?!!

They must have needed to drop the clocks back to ≤2 Ghz to make that work 🔥⚡

Or 600W TBP.
It would be a ridiculous amount, but It still should have a lot better perf/W than N31.

The size of a single SED would be interesting to know.

 

[PJVol]

They must have needed to drop the clocks back to ≤2 Ghz to make that work

Not really.
Rough math: assuming n31 gfx core 6SE ~= n41 3SE consuming 150W - ~ 20% process advantage gives 3x120 ~= 360W. Add to it 100-120W for the rest we have ~ 475W ASIC. This is not taking into account rdna3 architecture-wise lack of perf/w

 

[TESKATLIPOKA]

Not really.
Rough math: assuming n31 gfx core 6SE ~= n41 3SE consuming 150W - ~ 20% process advantage gives 3x120 ~= 360W. Add to it 100-120W for the rest we have ~ 475W ASIC. This is not taking into account rdna3 architecture-wise lack of perf/w

How did you get that 90CU 3SED N41 consumes 150W and performs as N31?