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

 

[PJVol]

I said consumes like n31 not petforms. And yeah, from the Hwinfo reported gfx core power

 

[TESKATLIPOKA]

I said consumes like n31 not petforms. And yeah, from the Hwinfo reported gfx core power

Are you saying that Hwinfo reports only 150W as gfx core power for N31? Isn't that wrong, considering It has 355W TBP?

 

[Frenetic Pony]

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?

This. A: Could be entirely trash, very obviously. B. Might be referring to which arch they were going with, and nothing to do with models at all. The arch shown is an arch, not a specific configuration, RDNA4 could very easily have an entire range of MCM parts that look like RDNA3 instead of being even more chiplet heavy (a good decision given non existent high end packaging supply at the moment anyway)

Regardless, looking at it again, if they didn't want to majorly swap out the cache heirarchy (the L2 cache utilization is good right now) doubling the L1 cache size would be a good idea either way. It has a pretty poor hit rate across the board, and clocking 20% higher should keep latency down even with a bigger cache size.

 

[TESKATLIPOKA]

This. A: Could be entirely trash, very obviously. B. Might be referring to which arch they were going with, and nothing to do with models at all. The arch shown is an arch, not a specific configuration, RDNA4 could very easily have an entire range of MCM parts that look like RDNA3 instead of being even more chiplet heavy (a good decision given non existent high end packaging supply at the moment anyway)

It looks pretty legit to me.
It could have been an actual product.

You can say that 270(9*30)CU is a serious overkill, but with RDNA3 they were losing pretty badly. To be on par with Full Ada102 they would have needed ~50% more CU, that's 144CU in total.
270CU for next generation would be 88% more. If they wanted to get the performance crown, then this amount is not that surprising.

Regardless, looking at it again, if they didn't want to majorly swap out the cache heirarchy (the L2 cache utilization is good right now) doubling the L1 cache size would be a good idea either way. It has a pretty poor hit rate across the board, and clocking 20% higher should keep latency down even with a bigger cache size.

L1 cache is part of CU(WGP), so you would have to clock them higher too, which would raise power consumption.

 

[adroc_thurston]

It could have been an actual product.

Oh it was.

You can say that 270(9*30)CU is a serious overkill

Kinda.

but with RDNA3 they were losing pretty badly

Because it clocks a whole lot less than it should.

If they wanted to get the performance crown, then this amount is not that surprising.

They really wanted to squash some bugs but not happening for obvious reasons.

L1 cache is part of CU(WGP)

A-ha.
Well you see.
(also he meant RDNA L1 aka SA-level read-only cache, not CU-private L0).

 

[TESKATLIPOKA]

Because it clocks a whole lot less than it should.

It's true that with 3.5GHz It would perform better than RTX 4090 and likely only loose to RTX 4090Ti by not that much in raster, but this didn't happen.
Let's hope RDNA3.5 shows us some nice frequency gains.

 

[PJVol]

Are you saying that Hwinfo reports only 150W as gfx core power for N31? Isn't that wrong, considering It has 355W TBP?

Yes (160w for the hypothetical 90cu). No.
The SoC power is too high IMO. They need something to do with it.

 

[TESKATLIPOKA]

Yes (160w to be correct). No.

Do you have a screenshot or link?

edit: does It also show the SoC power?
Considering AMD had to clock Shaders lower than the rest of the chip, I always thought It was the biggest consumer.

 

[beginner99]

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

it's what they have as the multi-gcd approach seems to have not been ready. A good bottom card can also be used for laptops if they finally manage to deliver good performance/watt.

 

[adroc_thurston]

always thought It was the biggest consumer

It is.

 

[PJVol]

Do you have a screenshot or link?

PCGH review?

 

[TESKATLIPOKA]

PCGH review?

I will check It out.

P.S. Why do so many of you like to link german sites. I don't speak german.

edit: Here It is.

Spoiler: RX 7900 XTX HWiNFO

 

[PJVol]

P.S. Why do so many of you like to link german sites. I don't speak german.

It's just that pcgh is the only site (from my memory) that publishes hwinfo screenshots, that is a big 👍 for me.
Also they recently started to put an aggregated performance index chart for the GPUs tested, that combines all resolutions and gaming scenarios including RT.
Google translates from german nice, btw.

 

[TESKATLIPOKA]

It's just that pcgh is the only site (from my memory) that publishes hwinfo screenshots, that is a big 👍 for me.
Google translates from german nice, btw.

I wasn't being serious. Take It as a joke from someone who is working for a German company, but is just too lazy to learn the language.

And...?

And I don't know which one is what GPU part.
You have:
GPU Core Power (VDDCR_GFX)
GPU Memory Power (VDDIO)
GPU SoC Power (VDDCR_SOC)
GPU Memory Power (VDDCI_MEM)
GPU USR Power (VddCR_USR)
What I know is TGP and TBP power.

 

[PJVol]

And I don't know which one is what GPU part.

Don't worry, it's ok

 

[SolidQ]

Isn't RDNA5 is weak for 2028 year?

 

[TESKATLIPOKA]

Isn't RDNA5 is weak for 2028 year?

RDNA5 should be 2025.
Why do you think next gen Xbox will be released in 2028?

 

[SolidQ]

Why do you think next gen Xbox will be released in 2028?

RDNA5 should be 2025.

2027 RDNA6, so one year before Xbox arrived

 

[TESKATLIPOKA]

2027 RDNA6, so one year before Xbox arrived

I see, I expected next gen sooner.

To your question If It's weak or not, I would say It could be pretty strong, but RDNA6 should be better. (IPC, frequency, features, perf/W, etc)

 

[SolidQ]

but RDNA6 should be better. (IPC, frequency, features, perf/W, etc)

Yeah, they use RDNA2 in XSX, despite it's out same year as RDNA2 release

 

[Tigerick]

@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%

 

[TESKATLIPOKA]

@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 ?	?	RDNA4 N44	N4P	?	?	GFX12 ?	3rd Gen ?	32	?	128-bit GDDR6	?	?
RX 6750 XT	$549	RDNA2 N22	N7	250 W	335	GFX1031	1st Gen	40	96MB	192-bit GDDR6	432 GB/s	+ 19%
RX 7700 ?	?	RDNA4 N43	N4P	?	?	GFX12 ?	3rd Gen ?	40	?	192-bit GDDR6	?	?
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%

Those specs are not very likely. If we say the smaller one is 32CU, then the bigger one should be 48CU or there is no point in giving It 50% more BW.

Then there is still the question about performance. It has to have significantly higher performance than N33 and N22, or what's the point in releasing It? Just to sell It dirt cheap so someone will buy It?

I wouldn't be surprised for RDNA4 to use N3E, If even Strix Halo IOD will use It.

Raytracing even If on par with Nvidia with these specs won't be that useful. RTX 4060TI 16GB manages only 45FPS and even that only at 1080p, at 1440p It's only 27fps in Cyberpunk.

 

[Tigerick]

Those specs are not very likely. If we say the smaller one is 32CU, then the bigger one should be at least 48CU or there is no point in giving It 50% more BW.

Then there is still the question about performance. It has to have significantly higher performance than N33 and N22, or what's the point in releasing It? Just to sell It dirt cheap so someone will buy It?

I wouldn't be surprised for RDNA4 to use N3E, If even Strix Halo IOD will use It.

Raytracing even If on par with Nvidia with these specs won't be that useful. RTX 4060TI 16GB manages only 45FPS and even that only at 1080p, at 1440p It's only 27fps in Cyberpunk.

Ha, I know you won't believe it. Let me ask you:

1. Why did AMD name full die of N33 as RX7600, not 7600XT?

2. Why did AMD name cut-down version of N32 as 7700XT, not 7800 or 7700?

Something to think about....

 

[TESKATLIPOKA]

Ha, I know you won't believe it. Let me ask you:

1. Why did AMD name full die of N33 as RX7600, not 7600XT?

2. Why did AMD name cut-down version of N32 as 7700XT, not 7800 or 7700?

Something to think about....

Do you really think AMD will release two RDNA4 chips and then call It as RX 7600XT and 7700 or 7800? What about the cut down versions of them?

BTW, AMD already uses 7600M XT and 7700S in mobile for N33, so using a very similar name for a different architecture looks even less likely.

Just something to think about....

 

[Tuna-Fish]

Here comes the table with AMD GPU lineup for 2024:- Pending confirmation.

Two notes:

1. The model names will be 8000-series (or some new system), even if that means lots of rebranded gpu models. The model names are not driven by common sense or by AMD, but by OEMs, and they want a shiny new model name for their new products, regardless of what it's slapped on.
2. 24Gbit modules should be commonplace next year, all the memory fabs are sampling them. So, 128-bit bus can support up to 12GB, and 192bit bus can support up to 18GB.