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

 

[gaav87]

529$ for 9070xt gigabyte gaming oc
IF this is even better than that is crazy

 

[eek2121]

Powercolor said, they have 20% less perf driver
Just for fun you can add 20% to leaked result

Nobody knows the final number, so if they stated this, they will be wrong unless they made a lucky guess. AMD hasn't released an official driver to anybody, even reviewers, despite shipping samples. That will change soon, of course.

I'm really surprised AMD hasn't thought of this before now, and they really aren't going hardcore enough. Just make the part artificially slow while consuming tons of GPU cycles so that board makers can design coolers and let the rest take care of itself. They can even send different drivers to each board maker to find leaks.

 

[SolidQ]

Nobody knows the final number,

I know, that why i'm writing "just for fun" We can have fun before event

529$ for 9070xt gigabyte gaming oc

Also Gpu Z screen was show better bandwidth, looks like 20gb, instead 19.5 in Gigabyte

 

[gaav87]

I know, that why i'm writing "just for fun" We can have fun before event


Also Gpu Z screen was show better bandwidth, looks like 20gb, instead 19.5 in Gigabyte

Specs are placeholder if u talking about the philiphines leaked price.

 

[jpiniero]

529$ for 9070xt gigabyte gaming oc
IF this is even better than that is crazy
View attachment 114733

5080/5070 Ti/5070 numbers are too optimistic...

 

[basix]

Frank azor just confirmed ALL the leaks are WORSE than RL performance. Too the moon. HOO HOO

I suspect Frank Azor does not know ALL the leaks, just the most prominent and little older ones

 

[eek2121]

5080/5070 Ti/5070 numbers are too optimistic...

Some folks are in for a rude awakening. The part is competitive with others in its class.

 

[eek2121]

I suspect Frank Azor does not know ALL the leaks, just the most prominent and little older ones

He reads this forum. 🤣

 

[lixlax]

Or the new shiny thing won't be in stock because "we knew the product was great but did not anticipate our competition to be quiet so terrible"...

I wouldn't be surprised at all. The more I think about the more certain I am that RTX 5000 is going to have very small uplift outside of framegen x3. Otherise they would've talked about non DLSS perf a lot more. Thats probably also the reason why they lowered the MSRPs compared to last gen (to minimize the backlash). The SKU that has reasonable amout more compute resources is the 5090 and that is also the only one that had its MSRP increased.

 

[gaav87]

I suspect Frank Azor does not know ALL the leaks, just the most prominent and little older ones

So all the watts 7900xt so its faster period.

 

[Sunaiac]

You guys are once again setting yourselves up for disappointment.
I'd rather set myself up for appointment.
So I expect 799$ card with 7900XT raster and 7600XT RT perf

 

[TESKATLIPOKA]

I will wait for reviews, this looks like another hype.
First the size is ~240mm2 and now 380mm2?
Then what about N44? What size does It have?

 

[marees]

I will wait for reviews, this looks like another hype.
First the size is ~240mm2 and now 380mm2?
Then what about N44? What size does It have?

Knowing AMD, N44 will not have all features of N48 just like N33 cut some features from N31/N32

 

[Josh128]

Stupid question: Why is top to bottom on RDNA2 and 3 21/31>23/33, but on RDNA4, N48 is supposed to be > N44???

 

[PJVol]

Stupid question: Why is top to bottom on RDNA2 and 3 21/31>23/33, but on RDNA4, N48 is supposed to be > N44???

Rather, what's up with navi 45, 46 and 47?) (assuming 40-43 were supposedly been cancelled)

 

[GTracing]

Stupid question: Why is top to bottom on RDNA2 and 3 21/31>23/33, but on RDNA4, N48 is supposed to be > N44???

Those codenames are given in the order that AMD starts to design them. N21 was designed first, then N22, then N23.

For RDNA4 it seems that AMD cancelled N41, N42, N43, N45, N46, and N47. That lines up with the rumors that AMD had some high end chiplet GPUs that they cancelled in favor of a monolithic chip (N48, AKA 9070/XT).

 

[Tuna-Fish]

And the reason they moved into this system was that their previous code name system gave out too much information about which design is what before it was out. So they publicly moved to a system that only tells you the generation and in which order the designs were started.

 

[Tuna-Fish]

Rather, what's up with navi 45, 46 and 47?) (assuming 40-43 were supposedly been cancelled)

They could have been just paper studies. You get the codename when a configuration is first proposed, before actual work goes into implementing it. Presumably, first they had a chiplet design, decided for some reason they didn't want it, then proposed a bunch of monolithic alternative designs to replace them, and chose 48.

 

[marees]

Rather, what's up with navi 45, 46 and 47?) (assuming 40-43 were supposedly been cancelled)

45, 46, & 47 must have been the monolithic replacements for 40, 41, 42

 

[gaav87]

Imagine beeing AMD design team lead and seeing this
RTX5080 only 17.5% faster than RTX4080 in cyberpunk2077

And mama SU thrown ur 384bit unified into the trash bin

3DCenter Forum - nVidia - RTX 5080: *Reale* Performance-Daten sowie Vergleich zu RTX 4080

nVidia - RTX 5080: *Reale* Performance-Daten sowie Vergleich zu RTX 4080 Spekulationen

www.forum-3dcenter.org

 

[gdansk]

Frank azor just confirmed ALL the leaks are WORSE than RL performance. Too the moon. HOO HOO

Regarding branding: "Gotta keep everyone on their toes" 🤣
Thanks AMD.

 

[Win2012R2]

their previous code name system gave out too much information about which design is what before it was out

4D chess ... or they could have just designed the best silicon to kick fake frames into the curb...

 

[reaperrr3]

Those codenames are given in the order that AMD starts to design them. N21 was designed first, then N22, then N23.

For RDNA4 it seems that AMD cancelled N41, N42, N43, N45, N46, and N47. That lines up with the rumors that AMD had some high end chiplet GPUs that they cancelled in favor of a monolithic chip (N48, AKA 9070/XT).

Possibly right for N40-43, but I bet there was never any N45-47, they just called N48 that because they used N44 as starting point and then doubled the design (4x2=8).

 

[GTracing]

They could have been just paper studies. You get the codename when a configuration is first proposed, before actual work goes into implementing it. Presumably, first they had a chiplet design, decided for some reason they didn't want it, then proposed a bunch of monolithic alternative designs to replace them, and chose 48.

n45-47 being projects that didn't make it past initial planning makes a lot of sense to me.

They could have been RDNA3-like chiplet GPUs. It would make sense if after canning their super-complex chiplet designs, AMD went back to what they did with RDNA3. 384bit, 256bit, and maybe 192bit designs that each have a single compute die with multiple cache/memory dies. And then they canned that too because a monolithic GPU 256bit GPU is more cost effective and 384bit models don't sell well.

 

[Win2012R2]

384bit models don't sell well

They did not sell well because -
1) fallen short of performance targets
2) had poor RT
3) cheapened out on design with too small GCD - which exaggerated #1 and #2

If 7900 XT actually did hit 4090 levels (in RT too) then it would have sold just fine for $1k.