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

 

[TESKATLIPOKA]

I'm not following. If the successor to 7800 XT is 3500MHz in games, then a higher speed GDDR6 will likely be sufficient. You think it won't be?

So you meant that 24gbps for the successor, I thought you meant It for standard 7800XT, which didn't make sense to me.

24gbps + 64MB IC should be enough for 3.5GHz 60CU.

 

[CouncilorIrissa]

Wake up babe, another copemon tweet just dropped

https://twitter.com/x/status/1774060521077260557

It appears to suggest that N48 < 7900XTX, but faster than 7900XT (?), and N44 is inbetween Wheat Nas (N32) and N33 (Hotpink Bonefish)

 

[jpiniero]

^^^^

That'd be a nice improvement, even assuming that N48 is more of a N32 replacement. 7900 XT is like 30% faster than a 7800 XT.

 

[SolidQ]

It appears to suggest that N48 < 7900XTX, but faster than 7900XT

Mlid was also saying between 7900XT and XTX, but RT faster than N31

 

[Hitman928]

Wake up, babe, another copemon tweet just dropped

https://twitter.com/x/status/1774060521077260557

It appears to suggest that N48 < 7900XTX, but faster than 7900XT (?), and N44 is inbetween Wheat Nas (N32) and N33 (Hotpink Bonefish)

Wouldn't the lowest N31 be the 7900GRE?

 

[SolidQ]

Wouldn't the lowest N31 be the 7900GRE?

Yes it's GRE. Maybe it's 7900XT level can be

 

[Timorous]

So I would guess N48 is 7900XT perf approx and N44 is 7700XT perf approx.

64 CUs at 3.3 GHz would probably manage that with fast enough vram.

32 CUs at 3.5Ghz or so could probably manage the later so I can see N44 being half of N48.

I would also guess that die size is going to be around 250mm and 125mm give or take.

If that is the rough performance then I think $600 for the top N48 part, 480 for the inevitable cut part. 350 for the top N44 part and probably 16GB of vram. A cut N44 part could be 250 ISH and probably 8GB, maybe 12 if they cut the bus to 96bit.

 

[SolidQ]

So I would guess N48 is 7900XT perf approx and N44 is 7700XT perf approx.

Fixed
Also possible N44 between 7700XT and 7800XT, because it's show >>, not >

 

[Timorous]

Fixed
Also possible N44 between 7700XT and 7800XT, because it's show >>, not >

Maybe but I would say the 7700XT is quite a bit faster than the 7600XT.

 

[Tuna-Fish]

Have there been any more recent info on what ram is used? 24Gbps GDDR6 seems still to be a paper product ("sampling"), while the only rumors about GDDR7 shipping seem to be for nV.

It might just be 256bit 20Gbps GDDR6 again.

It would be cheap, apparently most GDDR6 reaches that nowadays, to the point where new products that supposedly use slower grades use modules that can reach 20Gbps just fine. 7900GRE can be overclocked to be faster than a 7900XT when the memory is overclocked to ~20Gbps, even though that's only 4/5ths of it's bandwidth, meaning that the ~640GB/s isn't horribly limiting performance.

Could be that for a cost-optimized midrange part, they just built the fastest card they could with 256-bit 20Gbps GDDR6.

Also that would probably mean 16GB. I was hoping (and expecting, mea culpa) that 24gbit modules would be ready, but if this launches this year we should be hearing more about those already, and starting to see them in parts catalogs.

 

[Aapje]

RDNA4 should last only a year, so I can see them sticking with GDDR6 and then switching to GDDR7 for RDNA5. RDNA4 is just a stopgap product anyway.

And the weird part is that 24Gb was already defined for GDDR6, but they never implemented it. With the AI boom, would be really nice for Nvidia to have, so I assume that they are putting big pressure on the RAM makers to release it, but I guess that this is very hard to make.

 

[jpiniero]

With the AI boom, would be really nice for Nvidia to have, so I assume that they are putting big pressure on the RAM makers to release it, but I guess that this is very hard to make.

4080 Super already is using 23 gbps GDDR6X. May simply be 24 gbps and downclocked for some reason.

AFAIK AMD has no interest in GDDR6X although perhaps since GDDR7 also uses PAM, maybe they will?

 

[jpiniero]

Also that would probably mean 16GB. I was hoping (and expecting, mea culpa) that 24gbit modules would be ready, but if this launches this year we should be hearing more about those already, and starting to see them in parts catalogs.

Seems that the initial GDDR7 will be 2 GB.

 

[Aapje]

AFAIK AMD has no interest in GDDR6X although perhaps since GDDR7 also uses PAM, maybe they will?

GDDR6X was codeveloped by Nvidia and I'm pretty sure that Nvidia made a deal with Micron to get exclusive access to GDDR6X.

They even had to make a separate SKU for the 3060 Ti to dump excess slow GDDR6X into.

 

[TESKATLIPOKA]

Fixed
Also possible N44 between 7700XT and 7800XT, because it's show >>, not >

Even 7700XT is 54% faster than RX 7600 at FullHD.
It's unlikely for N44 to be faster than 7700XT in my opinion.
The question is the amount of Vram.

BTW Strix Halo which should have 40CU RDNA3.5 IGP is also supposedly >>N33.

 

[SolidQ]

Even 7700XT is 54% faster than RX 7600 at FullHD.

That from mlid(february) Called "4060ti killer"

 

[TESKATLIPOKA]

That from mlid(february) Called "4060ti killer"

That doesn't mean N44 will be faster than 7700XT, especially If he wrote performance above 7600XT but below 7800XT.

7700XT 12GB is 17%->19%->20% faster in raster than 4060Ti 16GB. TPU
4060Ti 16GB is 8%->8%->29% faster in RT than 7700XT 12GB. TPU

So even If N44 is weaker in raster than 7700XT but faster than 4060Ti, then It still can be considered as "4060ti killer", If It's not weaker in RT performance.

 

[TESKATLIPOKA]

Other tweet(Can't share more than 2 media):

https://x.com/All_The_Watts/status/1774075247886839954?s=20

"The Strix Halo iGPU > has 20 WGPs using the RDNA3+ architecture running at 3000 MHz."


These are more or less in-line with what Kepler and Adroc haven been saying. Interesting...

Something interesting reposted by @Ghostsonplanets.
3GHz turbo 40CU RDNA3.5 IGP.
That looks pretty good considering Strix Halo should be 125W part.

I think >3GHz for RDNA4 shouldn't be a surprise.

 

[SolidQ]

So even If N44 is weaker in raster than 7700XT but faster than 4060Ti, then It still can be considered as "4060ti killer", If It's not weaker in RT performance.

Only big question is how much CU in N44, 32 or 40, if 40 it's possible achive 7700XT perfomance.
Don't forget WGP is reworked + should be RT hardware accel

I think it's possible with higher clock and reworked wgp. If there 40CU
Difference usually 12-15fps
.

 

[eek2121]

Any rumors that suggest an x700 class part are wrong fyi. They have chips binned that are fast enough to be x800 class. I only bring that up because it seems some people think we are only getting something like a 8700xt.

Now as far as naming I have no clue…

 

[Tuna-Fish]

If they only have 2 dies, I would expect them to be binned/disabled so that they have at least a x800, x700 and a x600. x900 might be a stretch, depending on how fast the top part is, but AMD has shipped generations where the same "hundred" of next gen is worse than the previous gen, see Barts/Cypress. IMO that was terrible marketing, though, and I hope they won't repeat it.

 

[SolidQ]

They have chips binned that are fast enough to be x800 class

Because people looking at NVIDIA, their xx7x/ti class like prevoius 80ti/90class
Problem is AMD failed with RDNA3 naming 6600XT ~7600, 7800XT +- 6800XT, 6900XT/50XT ~ 7900Gre.
So that why people think about 8700XT should be like 7900XT

 

[jpiniero]

Problem is AMD failed with RDNA3 naming 6600XT ~7600, 7800XT +- 6800XT, 6900XT/50XT ~ 7900Gre.

Figure the branding is more what they want to charge for it. (Or try to anyway)

 

[GodisanAtheist]

Figure the branding is more what they want to charge for it. (Or try to anyway)

- Good take IMO.

I figure we'll get a x7xx level part if only because without a part that exceeds the previous gen top end performance of their own stack they will be competing with their own used parts and leftover inventory across the board, let alone NV.

 

[eek2121]

Because people looking at NVIDIA, their xx7x/ti class like prevoius 80ti/90class
Problem is AMD failed with RDNA3 naming 6600XT ~7600, 7800XT +- 6800XT, 6900XT/50XT ~ 7900Gre.
So that why people think about 8700XT should be like 7900XT

IMO AMD is just horrible at naming and a big part of that is pricing as another user stated. Quite often they simply follow NVIDIA’s lead. I do wish this would change, or if not, at least adjust model numbers down.

Other small rants:
1) AMD needs to set an example with reference cards. Reference cards should be the smallest/quietest cards possible. I went with a 4090 FE because they don’t do this.
2) OEMS need to focus on the looks of a card as well. Why is it hard to get a card that is a color other than black, gray, or white? It would be great if the industry could standardize removable shrouds in some way.