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

 

[Tuna-Fish]

Cache and memory controllers does not shrink well

I'd like to note that based on preliminary photos, AMD seems to have successfully shrunk the L3 quite a bit for Zen5 by general optimization of their design, despite the densest memory cells themselves not shrinking much at all. I think they can fit a lot of stuff on the die just from doing similar stuff to the 64MB of cache.

 

[Ajay]

I'd like to note that based on preliminary photos, AMD seems to have successfully shrunk the L3 quite a bit for Zen5 by general optimization of their design, despite the densest memory cells themselves not shrinking much at all. I think they can fit a lot of stuff on the die just from doing similar stuff to the 64MB of cache.

Yeah, that was an interesting surprise. Can’t air for an article about it.

 

[PJVol]

I'd like to note that based on preliminary photos, AMD seems to have successfully shrunk the L3 quite a bit for Zen5 by general optimization of their design, despite the densest memory cells themselves not shrinking much at all.

4T?

 

[Ajay]

4T?

No way. 6T minimum. Probably still 8T, but with some smanchy layout. IMHO.

 

[Kepler_L2]

No way. 6T minimum. Probably still 8T, but with some smanchy layout. IMHO.

L3 has been 6T since Zen3.

 

[Hans Gruber]

What will the highest level RDNA4 GPU perform like? Are we talking 7800xt or 7900GRE better/worse? There should be efficiency gains from the silicon alone. RDNA4 should be 20% or more efficient.

 

[Ajay]

L3 has been 6T since Zen3.

So 8T L1/L2? High speed cache used to be 8T (except a way back when is wasn’t really fast).

 

[adroc_thurston]

4T?

AMD SRAM is 6t HD for L3 since Zen3.

I'd like to note that based on preliminary photos, AMD seems to have successfully shrunk the L3 quite a bit for Zen5 by general optimization of their design, despite the densest memory cells themselves not shrinking much at all. I think they can fit a lot of stuff on the die just from doing similar stuff to the 64MB of cache.

Mind that MALL was already fair bit denser in Mb@mm^2 due to way lower freq targets.

 

[maddie]

On navi22, the infinity cache along with the memory controllers and infinity fabric take about a third of die.... and make note this is only 192bit bus.

N22 has a 256b bus. The 7700XT only uses 192b, but the space has been used for the full 256b layout.

 

[PJVol]

AMD SRAM is 6t HD for L3 since Zen3.

Yes, indeed

So 8T L1/L2?

L2 seems to use 6T except for the state macro which is 8T cells

 

[marees]

What will the highest level RDNA4 GPU perform like? Are we talking 7800xt or 7900GRE better/worse? There should be efficiency gains from the silicon alone. RDNA4 should be 20% or more efficient.

5% more than 7900xt in raster, if not memory constrained
5% less than 7900xt in raster, if memory constrained (for non-over clocked cards)

 

[marees]

Using NBD Trade Data can help the users comprehensively analyze the main trade regions of M S ADVANCED MICRO DEVICES INC. , check the customs import and export records of this company in NBD Trade Data System till now, master the upstream and downstream procurers and suppliers of this company, find its new commodities procured or supplied, search the contact information of M S ADVANCED MICRO DEVICES INC. and the procurement decision maker's E-mail address. NBD Trade Data System is updated once every three days. At present, the latest trade data of this company have been updated until 2024-02-28.

Recent customs import and export records of M S ADVANCED MICRO DEVICES INC. are as follows:

2024-04-18

Export

84733099

GRAPHIC CARD: NAVI48 G28201 DT XTX REVB-PRE-CORRELATION AO PLATSI TT(SAMSUNG)-Q2 2024-3A - 102-G28201-00 REV 13

INDIA

A***D

https://en.nbd.ltd/trader/info/NBDD2Y322830385

 

[marees]

Using NBD Trade Data can help the users comprehensively analyze the main trade regions of M S ADVANCED MICRO DEVICES INC. , check the customs import and export records of this company in NBD Trade Data System till now, master the upstream and downstream procurers and suppliers of this company, find its new commodities procured or supplied, search the contact information of M S ADVANCED MICRO DEVICES INC. and the procurement decision maker's E-mail address. NBD Trade Data System is updated once every three days. At present, the latest trade data of this company have been updated until 2024-02-28.

Recent customs import and export records of M S ADVANCED MICRO DEVICES INC. are as follows:

2024-04-18

Export

84733099

GRAPHIC CARD: NAVI48 G28201 DT XTX REVB-PRE-CORRELATION AO PLATSI TT(SAMSUNG)-Q2 2024-3A - 102-G28201-00 REV 13

INDIA

A***D

https://en.nbd.ltd/trader/info/NBDD2Y322830385

102-G28211
102-G28501
102-G28201 (navi 48 xtx)

https://www.nbd.ltd/customs-data?t=0&v=VIDEO%2FGRAHICS%20CARD

 

[MoogleW]

I'd like to note that based on preliminary photos, AMD seems to have successfully shrunk the L3 quite a bit for Zen5 by general optimization of their design, despite the densest memory cells themselves not shrinking much at all. I think they can fit a lot of stuff on the die just from doing similar stuff to the 64MB of cache.

Surely that sacrifices clocks? Could they really do that and clock above 3-3.2ghz?

 

[Kepler_L2]

Surely that sacrifices clocks? Could they really do that and clock above 3-3.2ghz?

It doesn't seem like Zen5's L3 sacrificed anything.

 

[TESKATLIPOKA]

I think the specs of navi 48 are going to similar to the specs of a 7700xt. 240mm2 of die space does not allow for 60 CU when you add back the memory controller and the infinity cache. On navi22, the infinity cache along with the memory controllers and infinity fabric take about a third of die.... and make note this is only 192bit bus. If navi 48 is using a 256bit bus which seems to be the case with the 16GB memory leaks, something has to take a hit. I think Navi 48 is going to have 48CU. Cache and memory controllers does not shrink well and if AMD is indeed changing the cores for RDNA4 and increasing the the ray tracing performance, a toll has to paid somewhere.

You cannot add back all the components of the MCD back into the die for only 40mm2 along with make improvements to the rest of RDNA4 architecture when N4P only provides a 6% improvement to transistor density. Only a positive note, I would not be surprised to see RDNA4 clock at 3.3 or 3.4ghz.

The leak said 32WGPs, 256bit GDDR6.

https://twitter.com/x/status/1784561456694046744

posted here by @CouncilorIrissa

 

[marees]

102-G28211
102-G28501
102-G28201 (navi 48 xtx)

https://www.nbd.ltd/customs-data?t=0&v=VIDEO%2FGRAHICS%20CARD

View attachment 100943

There were 2 more, but the dates were from last year. So I am guessing they are not RDNA 4

102-C48701-00
102-D74911-00

2023-04-03	84733030	PRINTED CIRCUIT BOARD ASSEMBLY FOR PERSONAL COMPUTER (VIDEO/GRAHICS CARD) - 102-C48701-00	A***ED	INDIA	M***D.	***	更多
2023-02-23	84733030	PRINTED CIRCUIT BOARD ASSEMBLY FOR PERSONAL COMPUTER (VIDEO/GRAHICS CARD) - 102-D74911-00	A***ED	INDIA	M***D.	***	更多

 

[SolidQ]

I wouldn't suprise if we're going to see same situation Radeon 7 vs Radeon 5700XT(7900XTX vs 8700XT), but with much faster RT
AMD is unpredictable in last time

 

[marees]

I wouldn't suprise if we're going to see same situation Radeon 7 vs Radeon 5700XT(7900XTX vs 8700XT), but with much faster RT
AMD is unpredictable in last time

You mean 7900 xt vs 8700 xtx (with much faster RT) ??

 

[SolidQ]

7900 xt

XTX

it's possible between XT and XTX, but in some games beat XTX, like was 5700XT

 

[ToTTenTranz]

To be honest, more than 7900XTX rasterization performance isn't really needed almost anywhere, at the moment.

That GPU is already much faster than a PS5 (~3x or more?), so with that baseline in mind and much faster CPUs available, we're not getting games that demand a lot more anytime soon.

So a GPU that performs close to the 7900XTX in raster performance and significantly higher raytracing performance could be enough to make me upgrade from my 6900XT.

 

[Ajay]

To be honest, more than 7900XTX rasterization performance isn't really needed almost anywhere, at the moment.

4k 240 Hz+ would like a word (game dependent).

 

[ToTTenTranz]

4k 240 Hz+ would like a word (game dependent).

Yes, and I'm all for people having the freedom to buy super expensive hardware that allows them to play games at 4K with 500FPS on 500Hz monitors.

It's just that those games would be identically enjoyable at 55-80FPS VRR by >99.9% of the people with much cheaper and accessible hardware.

 

[Mopetar]

I can drive to work in a no frills Toyota perfectly fine, but if I had the money I could do it in a Ferrari.

Expensive GPUs and CPUs are a considerably inexpensive hobby by comparison to alternatives.

 

[dr1337]

Saw this just now, N48 boards seemingly have been in pre-production for a few months already.