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

 

[Heartbreaker]

Folks, stop trying to come up with rebranding option better than Amedeon. Won't happen.

I dont' get it.

 

[IEC]

Should rebrand to Remedeon because they need to go back to Marketing 101 and remediate their failings since Vega.

Hopefully RDNA4 is more like Polaris but I’m not going to hold my breath.

 

[blackangus]

If they are going to re-brand they need to buy the 3DFX name and rebrand to that.
(Yes I know this would be difficult, but its a joke anyway =)

 

[SolidQ]

about 9070XT. 5k Yuan

 

[jpiniero]

about 9070XT. 5k Yuan

So $649 perhaps?

 

[exquisitechar]

So $649 perhaps?

That’s what I was expecting.

 

[gaav87]

So $649 perhaps?

Well rip mss il buy it thou xD

 

[SolidQ]

So $649 perhaps?

yeah and vanilla 550?

 

[Mopetar]

I think AMD missed a big free marketing opportunity with the delay of RDNA4. I would assume RDNA4 would have run circles around 7900XTX running local LLMs.

Pray tell how AMD was supposed to know about a Chinese developed LLM that caught the entire market off guard and how well it would perform on their hardware?

Between the low inventory and the driver problems, it's looking like Nvidia should have delayed until March as well.

 

[marees]

yeah and vanilla 550?

Probably $500 or $530

 

[Tup3x]

To be honest, if we are lucky 9070 XT costs 899 € here in EU VAT included. 799 € for non XT version probably. In extremely lucky situations -100 € off from both prices. I can't see them being any cheaper than that.

 

[adroc_thurston]

about 9070XT. 5k Yuan

Hell yeah please let it be $749.

 

[PJVol]

Sigh... I was hoping Nvidia would price the 5070ti at $600-650 so I could get the 9070xt for $550.

 

[beginner99]

So $649 perhaps?

Yeah even with mediocre blackwell that won't sell well at all. gain some in raster vs the 5070 Ti, lose in RT minus brand recognition and no vram advantage. that would be a hard sell, literally worth than 7900 xt price/performance wise.

 

[jpiniero]

Yeah even with mediocre blackwell that won't sell well at all. gain some in raster vs the 5070 Ti, lose in RT minus brand recognition and no vram advantage. that would be a hard sell, literally worth than 7900 xt price/performance wise.

At some point if they are going to stay in this, they need to show a little bit of chutzpah cuz the nodes aren't getting any cheaper.

I believe there is not going to be a FE for the 5070 Ti so the $749 is going to be even less real than the 5080 and 5090.

 

[reaperrr3]

Lets say n4p is 16000$ per 300mm wafer
And n4p is 13x27 ~351mm2 this means 160 dies per wafer
80% yield = 128 dies per wafer = 125$ per n48 (without dicing, testing, packaging, and binning)
Packaged and binned N48 ~150$
16GB gddr6 = spot pricing~36$
My total BOM est. is less than 300$
Lets say 300$ 1.5x gross profit margin = 450$ (chinese chiphell amd insiders still saying yesterday 479$ was the price shortly after ces)
(If the 300mm wafer is 20k = 500$, 24k = 525$)
So even at 24k$ per wafer and 80% gross margin 630$

- Re wafer prices: Not ruling out that's what AMD is paying due to contracts, but the standard price for N4P is allegedly rising from 18k$ in 2024 to 20k$ for this year.
- yield of chips per wafer that can handle XT specs is likely quite a bit below 80% (let's say 60%).
- I remember reading somewhere that the GPU makers usually have contracts with the mem makers that basically protect AMD/NV from price spikes, but also the mem makers for demand slumps, so AMD might actually be paying a bit more than the current spot prices.
- to leave some margin for AIBs (who handle PCBs, cooling, some testing, putting cards into boxes with some accessoirs) distributors and shops even at MSRP, AMD themselves probably sell the GPU+mem at significantly below MSRP.
Example: If the XT's MSRP were 649$, I don't see AMD themselves asking much more than 450-500$ for GPU+mem, leaving them with maybe 200-300$ profit (in this example).

And they still need to make back all the fixed cost for R&D, design, masks etc. (and fund next-gen's development), so that wouldn't be as much as it sounds, unless they move tons of volume at that price.

 

[Mopetar]

yield of chips per wafer that can handle XT specs is likely quite a bit below 80% (let's say 60%).

The specs are chosen based on yields (including parametric yields) to hit whatever product mix AMD thinks they can sell, which is largely based on historic data.

Most dies will not have any defective parts, but will still have some cores disabled because they can't achieve the minimum clock speed at target voltages for the top x% bin.

If the process node is mature or the design team didn't do anything that caused issues, that just means that the final product has better specs than it otherwise would have.

These days I think that AIBs are largely allowed to ignore MSRP and charge whatever they can get away with to account for their own costs. Reference models still need a third party manufacturer so AMD still pays for costs of the board and assembly.

What you left out is the retailers margin. Depending on the product retail margins have historically been as much as 30% of the sale price. For some products though companies will negotiate lower percentages or even just a flat amount.

 

[gdansk]

At some point if they are going to stay in this, they need to show a little bit of chutzpah cuz the nodes aren't getting any cheaper.

I believe there is not going to be a FE for the 5070 Ti so the $749 is going to be even less real than the 5080 and 5090.

Honestly they should simply copy Nvidia. Make some small amount of MBA cards that are never in stock. Give them a low, fictional MSRP. And tell AIBs to charge $100-$200 over that if that's what that market can support.

It gives them great lines in review charts like an $800 RTX 5080. Why not get a $400 9070 XT line in that same chart? 20% off both of them. He'll have to include it if they make up MSRPs too.

 

[coercitiv]

I just give up, the absurdity is reaching new heights every day.

What follows is a fragment from a TechPowerUp interview with David McAfee from AMD (GM of Client Channel Business)

TechPowerUp: Is RDNA 4 a monolithic design?
David McAfee: We haven't talked about that yet.

TechPowerUp: Why are you introducing a new Radeon naming scheme at this time and why?
David McAfee: We've been building momentum with Radeon. Our strategy is similar to Ryzen—focus on value, listening to the community, and providing features they care about. We want to ensure that Radeon graphics deliver excellent capabilities for gamers at reasonable price points.

TechPowerUp: I like it.
David McAfee: It's a good move. Transparency helps consumers understand our products better.

Interviewer: I like it.
Company: It's a good move.
Gamer: Yeah man, transparency!

 

[SolidQ]

gamers at reasonable price points.

750$ would be reasonable?

Gamer: Yeah man, transparency!

Another Gamer: I can also buy 5090 for 4k$!

 

[Hans Gruber]

AMD went from we are no longer trying to compete with Nvidia's high end GPU's 5080/5090. Suddenly Nvidia flops and everybody is getting excited about RDNA 4? I don't get it. A 7900xtx on N4P and GDDR7 would be something to get excited about.

 

[Mopetar]

I just give up, the absurdity is reaching new heights every day.

What follows is a fragment from a TechPowerUp interview with David McAfee from AMD (GM of Client Channel Business)

Those sort of people are trained on how to handle talking to the press and are specifically instructed to treat members of the press like mushrooms.

Once you understand that their corporate droid responses make sense, even if they're not any more informative. It may not be any more entertaining, but it is at least less maddening.

 

[Mopetar]

AMD went from we are no longer trying to compete with Nvidia's high end GPU's 5080/5090. Suddenly Nvidia flops and everybody is getting excited about RDNA 4? I don't get it. A 7900xtx on N4P and GDDR7 would be something to get excited about.

People are only excited because AMD technically hasn't disappointed them yet. The excitement will die when the product launches and it's the same old same old once again. When reality sets in and they realize that Intel is their only hope, it may dawn on them just how screwed they really are.

I'm just saying that if we made a mountain out of this molehill it would be called Mt. Doom.

 

[gdansk]

A 7900xtx on N4P and GDDR7 would be something to get excited about.

We have seen RDNA3.5 on N4P and it's merely adequate, not exciting. Barely beating BMG (N3B).

Some people are excited that RDNA4 might possibly not be total crap like RDNA3 and Blackwell. But most people are excited by the prospect AMD might force Nvidia to cut prices. They won't.

 

[adroc_thurston]

We have seen RDNA3.5 on N4P and it's merely adequate, not exciting. Barely beating BMG (N3B).

"merely adequate, not exciting, barely beating BMG (N3B)"