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

 

[Schmide]

The irony. The place where you quoted all these graphs just did a followup basically agreeing with Steve. Steve will not have to do an investigation.

https://www.techspot.com/review/3008-radeon-9070-xt-amd-finewine/

Edit: In a strange turn of fate, the techspot review is done by a Steven. Well at least we can tell him appart from the other Steves. Until we watch Rick Steve's europe then we get all possessive.

 

[mostwanted002]

Any hints/leaks/clues about Radeon PRO AI R9700 ???

(Also, I clearly remember announcements and news headlines quoting the card as W9700. When did it become R9700 )

 

[igor_kavinski]

Any hints/leaks/clues about Radeon PRO AI R9700 ???

Should hit the shelves in July: https://www.fudzilla.com/news/graphics/61129-powercolor-announces-radeon-ai-pro-r9700-graphics-card

 

[igor_kavinski]

Is @adroc_thurston excited about the 9080 XT 32GB GDDR7?

 

[adroc_thurston]

Is @adroc_thurston excited about the 9080 XT 32GB GDDR7?

Client graphics is a ghetto for poor people so who cares about that.

 

[igor_kavinski]

I am committing to Bazzite today on my 5600X3D+RX 6800 system.

Another guy's positive experience: https://www.overclock.net/posts/29482450/

Ran Fallout New Vegas using DXVK earlier today on Win 11, applied a lot of performance fixes and anti crash handling optimizations, game works but is a stuttery mess yet these mods were supposed to fix that.

Under Linux, you install the game, run the game, smooth as silk like we had never truly experienced even 60FPS smooth before.

 

[soresu]

Another guy's positive experience: https://www.overclock.net/posts/29482450/

I'd like to say that MS has utterly and completely dropped the ball.

Unfortunately AMD likely owns no small part in this as their proprietary (gaming) GPU drivers are being deprecated on Linux for good reasons.

That being said, both DXVK and Zink dev teams are basically laser focused on Linux perf and compatibility, with the Windows use case as only a secondary consideration.

They are made with Mesa/Gallium in mind, and I doubt they will ever work truly 100% on Windows drivers.

If I'm talking out of the crack of my posterior and anyone here knows any better please feel free to correct me on this, I'd love to hear more about DXVK and Zink on Windows.

Edit: Salient point, these are the specs of that machine:

14600K + PTM 7950| Alphacool Eisbaer 280mm| Asrock Z790 Livemixer| 32GB DDR5 7000mhz C32 Patriot Viper| XFX MERC 310 RX 7900 XTX + PTM 7950 + TG putty| Crucial P5 2TB| ACER FA200 2TB| Samsung 980 1TB| Silicon Power A55 1TB| Toshiba 3TB HDD| MSI MPG A850GF PSU| Antec NX416 case| AULA F75 keyboard| Razer Viper Pro V3 8K mouse| MSI 325CQRF QD E2 32" QHD monitor| Gullikit King Kong 3 MAX controler

 

[moinmoin]

That being said, both DXVK and Zink dev teams are basically laser focused on Linux perf and compatibility, with the Windows use case as only a secondary consideration.

They are made with Mesa/Gallium in mind, and I doubt they will ever work truly 100% on Windows drivers.

I mean how could it be any different? These are open source projects, with people working with open source code. They are being tested against open source drivers which are often tweaked as well. This is what makes the whole stuff feasible to begin with.

Closed source operating systems and closed source drivers are black boxes. There it's completely by chance whether something works as intended or not, and nobody working with open source can change anything about that.

 

[igor_kavinski]

There it's completely by chance whether something works as intended or not

Not true though. Collusion (funny way to put it I guess) between the OS developer and hardware developer and further collusion between hardware developer and software developer ensures that things work, most of the time. Substitute "collusion" with "partnership" if you want to be politically correct

Microsoft has long been accused (and somewhat proven to be guilty of with the "undocumented" series of books) of having undocumented APIs in its operating systems for helping with their preferred "partnerships".

 

[soresu]

Closed source operating systems and closed source drivers are black boxes

Yes and no.

By design Vulkan is structured so that the vast majority of coding is the onus of the application programmers rather than the driver dev team.

That's a big part of why DXVK and Zink are possible, but of course it carries the impact of being a huge investment to code with of course as a result.

Recent PR and new extensions from Khronos show that they are trying to mitigate this though and make Vulkan more accessible, or at the very least significantly less burdensome to code for.

 

[moinmoin]

most of the time

Exactly right, most of the time. What word aside "chance" do you like to choose for something sometimes randomly not working as intended?

And if it doesn't, anybody sticking to open source won't be able to "help" closed source even if they wanted to. There is a reason why Nvidia is so universally "loved" in open source lands. There is also a reason why Nvidia in now over a decade seems incapable to deliver e.g. decent Wayland support.

By design Vulkan is structured so that the vast majority of coding is the onus of the application programmers rather than the driver dev team.

All the parts still need to fit together regarding calls and callbacks. Which starts to be possible to check and ensure once aside the game the whole stack is open.

 

[Thunder 57]

You didn't address the counter strike issue. I will show you why the results are so suspect from their own data concretely. Below are the original 5070 review.

Here is the 9600xt review. You can't say different scene because of numbers from cards like the 4070 and 7800xt(they are the same, different scene would be different numbers). The 9060xt techspot review specifically mentions gains from blackwell.

View attachment 126567

These updated 5070 ti results are worse than a 5070 from their 9060xt review. If this isn't explicitly suspicious, I really can't help you. You cannot use the different scene excuse they mention the 479 fps in initial review.

And if you are going to throw our results because they are buggy, why did they keep the spiderman ones? The RTX 5070 ti numbers are the same as an RTX 5070.

More issues can be seen in this if we cross reference the 9600xt review with same the original RTX 5070 review.

Look at cyberpunks results. Launch results with a RTX 5070.

Now look at the 9060xt review with updated 5070 results.

There is a substantial gain of 12%. More than the gains from updated 9070 results(7%). Considering how much faster the 5070 ti is than the RTX 5070 ti, the results make no sense.

Stalker also shows this. Launch review for the RTX 5070 which has the same results in this updated comparison for the RTX 5070 ti(so no different scene excuse).

Now here are the results for the 5070 in the 9060xt review. Again significant gains from driver updates from 52 to 58 FPS. you can't use the difference scene excuse the numbers for the 7800xt and 4070 are the same in both reviews. Where are these gains for the RTX 5070 ti in this comparison? Do you think the RTX 5070 ti is only 10% faster than a 5070 is an accurate result?

They are only testing two cards, buggy results is a half ass excuse to not include 2 popular games still played now. Warthunder and marvels rival have around 50k active players today compared to games to starfield which has 2000 players or even hogwarts legacy which has around 10k players.

Too many weird results affecting too high a percentage of the sample size of games. If Steve is not being paid by AMD, he should have retested with a different system because too many buggy results to accept this comparison as valid.

I started to read this post until I just started scrolling down because it was verbal diarrhea. The fact you spent this amount of time to promote one company and call out HUB is concerning. Why not go enjoy that time instead playing your favorite games on your Nvidia GPU and then everyone wins.

 

[jpiniero]

https://videocardz.com/newz/aida64-adds-support-for-zen6-cpus-and-unannounced-radeon-rx-9060-non-xt-gpu

Speculation about the 9060 Non-XT. I think this is a mistake, but assuming AMD isn't getting any OEM deals, they have to do something with the busted dies.

 

[maddie]

https://videocardz.com/newz/aida64-adds-support-for-zen6-cpus-and-unannounced-radeon-rx-9060-non-xt-gpu

Speculation about the 9060 Non-XT. I think this is a mistake, but assuming AMD isn't getting any OEM deals, they have to do something with the busted dies.

Where did the "busted" 7600 go?

 

[jpiniero]

Where did the "busted" 7600 go?

Mobile. (In theory, anyway)

7600M and 7600S have the CUs cut from 32 to 28.

 

[Timorous]

https://videocardz.com/newz/aida64-adds-support-for-zen6-cpus-and-unannounced-radeon-rx-9060-non-xt-gpu

Speculation about the 9060 Non-XT. I think this is a mistake, but assuming AMD isn't getting any OEM deals, they have to do something with the busted dies.

The play here is what I said a long time ago. 96bit 12GB 28CU or 24CU and sell it for $250. EOL the 9060XT 8GB and just have a 12GB 9060 and 16GB 9060XT.

It instantly destroys the 5050 at the same price.

Performance should be in the ballpark of the 6700XT and it would be a very solid card.

 

[Heartbreaker]

The play here is what I said a long time ago. 96bit 12GB 28CU or 24CU and sell it for $250. EOL the 9060XT 8GB and just have a 12GB 9060 and 16GB 9060XT.

It instantly destroys the 5050 at the same price.

Performance should be in the ballpark of the 6700XT and it would be a very solid card.

So your brilliant plan for AMD is they should make a more expensive to manufacture card (more VRAM), and sell it for less??

Did you consider that perhaps that the 9060 XT is already near the minimum margin they find acceptable?

 

[Timorous]

So your brilliant plan for AMD is they should make a more expensive to manufacture card (more VRAM), and sell it for less??

Did you consider that perhaps that the 9060 XT is already near the minimum margin they find acceptable?

Pairing a fully working die with 8GB of VRAM is just a waste, give it 16GB and the increase in ASP is far more than the incremental increase in BOM.

For the cut part, a cut part is just salvage, 12GB and 96bit has more market value than an 8GB 128bit part yet the number of defective dies that would pass the binning to be 128bit 28CU is lower than it is with a 96bit version simply because a few defects will hit the memory PHY.

Give it lower power and lower clocks to go with it and you have a cheaper heatsink design, smaller box, etc etc.

An 8GB cut part just won't sell, even at $220-$230. Slicing the bus and giving it 12GB means it will sell like hot cakes even at $250-270.

 

[jpiniero]

An 8GB cut part just won't sell, even at $220-$230.

Yields are so good at TSMC that it doesn't need to sell much.

My guess is that it will be $250 and market it as being faster than the 5050.

 

[ToTTenTranz]

So your brilliant plan for AMD is they should make a more expensive to manufacture card (more VRAM), and sell it for less??

Taking away SKUs that bring down brand value and apparently don't even sell enough units to justify the production line is simply an obvious plan and not a particularly brilliant one.

As for the "more expensive to manufacture" part, at the moment we're looking at $2.5 per GByte average spot price on GDDR6. Weekly minimum is less than $1.5.
12GB instead of 8GB brings an additional $6 to $10. Pushing that onto the consumer means $230 instead of $220, or $260 instead of $250. It's not that much of a problem.

 

[Thunder 57]

Pairing a fully working die with 8GB of VRAM is just a waste, give it 16GB and the increase in ASP is far more than the incremental increase in BOM.

For the cut part, a cut part is just salvage, 12GB and 96bit has more market value than an 8GB 128bit part yet the number of defective dies that would pass the binning to be 128bit 28CU is lower than it is with a 96bit version simply because a few defects will hit the memory PHY.

Give it lower power and lower clocks to go with it and you have a cheaper heatsink design, smaller box, etc etc.

An 8GB cut part just won't sell, even at $220-$230. Slicing the bus and giving it 12GB means it will sell like hot cakes even at $250-270.

You really ought to give AMD a call and tell them how stupid they are.

 

[Heartbreaker]

Pairing a fully working die with 8GB of VRAM is just a waste, give it 16GB and the increase in ASP is far more than the incremental increase in BOM.

They do have a full die with 16GB, it's called the 9600 XT 16GB.

For the cut part, a cut part is just salvage, 12GB and 96bit has more market value than an 8GB 128bit part yet the number of defective dies that would pass the binning to be 128bit 28CU is lower than it is with a 96bit version simply because a few defects will hit the memory PHY.

Cut parts are mostly NOT salvage, because yields are generally quite good when mainstream parts hit mass production. So while there is some salvage in the cut down parts, they are mostly just disabled for product segmentation, and they could have sold as fully enabled parts.

For this die they don't even have a cut down part at all, so yields must be extremely good, and AMD is really loath to lower their margins with product segmentation.

 

[jpiniero]

Cut parts are mostly NOT salvage

I would think in this case it would be. Otherwise I don't think AMD would bother.

 

[marees]

I would think in this case it would be. Otherwise I don't think AMD would bother.

There was always an n44xl in plan in addition to n44xt

However launch window is not clear / unknown

 

[mostwanted002]

Day 9 of looking for Radeon AI PRO R9700 on shelves. ⌚