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

 

[adroc_thurston]

how about this "Not relevant"? Under 35% and 150-300$ price

3060, 3060ti, 3070 are all higher ASP parts.
You should do real channel checks instead of quoting gabeN survey.

 

[SolidQ]

You should do real channel checks instead of quoting gabeN survey.

I know about price, but alot people still get 150-300 range.

 

[adroc_thurston]

I know about price, but alot people still get 150-300 range.

Channel checks say otherwise.

 

[Mahboi]

Channel checks say otherwise.

That's incredibly irrelevant, people who can only go below channel prices will buy used.
You need a proper feel for the market and channel is a godawful indicator.

 

[adroc_thurston]

That's incredibly irrelevant, people who can only go below channel prices will buy used.

No it's a good metric.

You need a proper feel for the market and channel is a godawful indicator.

Used GPU market is pretty tiny.

 

[Mahboi]

No it's a good metric.

Top quality argument as usual

Used GPU market is pretty tiny.

For which you have no metric since nobody cares.

 

[adroc_thurston]

Top quality argument as usual

?

For which you have no metric since nobody cares.

It's just tiny.

 

[Mahboi]

?

?

 

[maddie]

Before, people "hoped" that the new AMD card with the next generation architecture would compete with Nvidia's best for the top.
Now people are "hoping" that that in the generation after that AMD will decide to compete with Nvidia's best for the top.

Interesting drive-bys.

 

[adroc_thurston]

Before, people "hoped" that the new AMD card with the next generation architecture would compete with Nvidia's best for the top.
Now people are "hoping" that that in the generation after that AMD will decide to compete with Nvidia's best for the top.

That's the easy part.
Doing it at lower power/area, now that's a trick.

 

[Mahboi]

That's the easy part.
Doing it at lower power/area, now that's a trick.

A trick which NV doesn't seem to care to do much. Doesn't seem to bother their sales either.

 

[Aapje]

A trick which NV doesn't seem to care to do much. Doesn't seem to bother their sales either.

Ada is very power efficient. I'm hoping that RDNA 4 catches up in this regard.

 

[Mahboi]

I'm not too worried about the power draw being reined in. RDNA 1/2 were fine, RDNA 3 just ate electrons for 2 years due to its design issue. I think RDNA 4 will fix it all decently and on TSMC 4 there's no reason to worry, I'm sure it'll be a good gen.
The better question for me is how much better will the power draw be on RDNA 3.5 vs RDNA 3, I'm really expecting near-miracles at this point, like 35% less power draw for the same FPS.

My honest disappointment with RDNA 4 is with the size of the biggest die, which is still utterly ridiculous.
At least we'll see an interesting RT development for what it's worth.

 

[igor_kavinski]

At least we'll see an interesting RT development for what it's worth.

What's not clear is if it will require developers to add support for the RDNA4 RT enhancements to their game engine or if it will benefit all RT titles retro-actively without any development cost. I'm hoping it's the latter but given how AMD has so far preferred doing things their way (not using dedicated RT units like Intel and Nvidia), I'm skeptical.

 

[adroc_thurston]

A trick which NV doesn't seem to care to do much.

Oh but they do.
Gutted the hell out of DPFP MULs on B100 to hit that area target.

 

[Mahboi]

Oh but they do.
Gutted the hell out of DPFP MULs on B100 to hit that area target.

Oh lord, what a loss, what infamy.
Meanwhile AMD plans any die above 300mm² as "too expensive without chiplets".

 

[marees]

Oh lord, what a loss, what infamy.
Meanwhile AMD plans any die above 300mm² as "too expensive without chiplets".

Which is why I say that Nvidia is (or became long ago) a vertically integrated software company (like apple)

Whereas AMD is still a hardware company.

What Nvdia (& Apple) do will work only for them and not for others

Nvidia was amongst the first to go for unified/programmable shaders and Taiwanese/TSMC outsourced manufacturing. They've been able to distort the market enough to make it consume their supply.

 

[igor_kavinski]

Meanwhile AMD plans any die above 300mm² as "too expensive without chiplets".

Why would they risk decreasing their wafer yield for parts that most gamers consider the budget option?

We know that Raja Koduri screwed them over with his ambitions and then did the same thing to Intel's GPU effort too by overhyping and over-promising. Maybe he's a double agent working for Nvidia. Screwed over the only two viable competitors. If we want AMD to start taking risks, first we need to start taking risks ourselves by plopping down full price for their GPUs (good or bad). I still have an unopened ASUS 6800 XT LC card (coasting by on Gigabyte RX 6800). I believe DrMrLordX also has an unopened 7900 XTX. We need more of us to support AMD.

 

[ToTTenTranz]

N48 is two N44 dies glued together?

https://twitter.com/x/status/1795047238076350468

 

[soresu]

N48 is two N44 dies glued together?

https://twitter.com/x/status/1795047238076350468

I'll believe it when I see AMD's announcement slide deck, and not a moment sooner.

After the incessant back and forth over the chiplet config of RDNA3 hi end only to find out it was IO/♾️$ only I have no confidence in chiplet rumors at all.

For RDNA5 I would pretty much expect shader chiplets for at least 1-2 segments, if not all - until then rumors on that score are just wind.

 

[leoneazzurro]

If that was true, then a 3-chiplet (or even 4) variant should have been relatively simple, and it would have in part satisfied the high end market. So I am quite skeptic about that.

 

[soresu]

I still have an unopened ASUS 6800 XT LC card (coasting by on Gigabyte RX 6800)

Why you have 2 GPUs of the same segment?

 

[soresu]

If that was true, then a 3-chiplet (or even 4) variant should have been relatively simple, and it would have in part satisfied the high end market. So I am quite skeptic about that.

More chiplets = more IO connections = more overhead = larger IO die.

 

[igor_kavinski]

Why you have 2 GPUs of the same segment?

Got the LC first. Then saw someone selling their 6800 for $258. Couldn't resist. Found out it offers more than enough performance for my needs. 6800 XT is in reserve for emergency need (maybe years later ).

 

[poke01]

The reasoning that AMD is doing MCM because of Nvidia is wild. MCM existed before, Apple did it.

If AMD uses MCM it’s because it’s the easiest way to design a newer die. Until more people report on this I have my doubts.