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

 

[marees]

That from new patch RE4. What is real resolution don't know

The PS5 pro has PSSR. Thay scales better than DLSS (even though it is less stable)

Effectively PS5 pro looks like a 4070 super with DLSS due to PSSR

Not sure how RDNA 4 will scale with FSR 4

Without scaling n48xtx RT would be around 4070 super

 

[soresu]

The PS5 pro has PSSR. Thay scales better than DLSS (even though it is less stable)

PSSR will become more stable with time to tune it, just as early DLSS did.

 

[eek2121]

It wouldn't shock me if AMD leaned on FSR4 to compete next gen. Sure, NVIDIA cards will ultimately be faster, but AMD cards will punch above their weight in terms of perf/$ thanks to upscaling. I don't necessarily have an issue with that, provided they focus on a quality image. The few games where I've used DLSS on my 4090 looked perfectly fine to me. Sure, if you looked close, you may see an artifact, but very few gamers will be doing that. Just look at how AA evolved in the past.

I also think NVIDIA and AMD will both be investing heavily in upscaling tech moving forward as a way to keep die sizes down.

 

[jpiniero]

Finally I saw some power consumption on the PS5 Pro and it is only like 10-30 W more. So you are talking 220-240 W at the wall.

The performance increase looks like a big flop though, esp given that it's almost double the price (albeit with 2 TB versus 1)

 

[Arctic Islands]

What I heard was very similar to yours, RT = AD104/4070TI?, rasterization/nonRT = 7900XTX, but I don't know the naming scheme. TSMC N4.

Or are you just speculating?

Also N48 is still using a GCD+MCD design. Total die sie is much larger than 250mm^2 while N44 is smaller.

 

[poke01]

Also N48 is still using a GCD+MCD design. Total die sie is much larger than 250mm^2 while N44 is smaller.

wait is it not monolithic??
Ugh, my hopes for idle power consumption etc getting better dropped a bit now.

 

[Kepler_L2]

wait is it not monolithic??
Ugh, my hopes for idle power consumption etc getting better dropped a bit now.

It is monolithic.

 

[poke01]

It is monolithic.

Good. MCD design is bad for power consumption when doing light tasks.

 

[SolidQ]

MCD design is bad for power consumption when doing light tasks.

i think it's going solve in future, because chiplet design for now is future.

 

[soresu]

Good. MCD design is bad for power consumption when doing light tasks.

Not intrinsically, it's simply a problem of numbers.

Namely the density of metal wires inside a monolithic die vs the density of pins/microbumps/TSVs/hybrid bonds etc that connect 2 separate dies together.

It's a similar problem with bringing L2$ off the die.

As die interconnect density increases the negatives of MCD/stacking will decrease.

 

[ToTTenTranz]

It is monolithic.

https://twitter.com/x/status/1854795942131908986

🤷‍♂️

 

[poke01]

https://twitter.com/x/status/1854795942131908986

🤷‍♂️

Anything is possible these days

 

[beginner99]

If N48 isn't monolithic, the thrown around performance numbers in conjunction with the die size would make a whole lot more sense as the die size like is then only the gcd.

 

[eek2121]

This rumor has popped up before. I suspect it is bad info. The original implied that AMD finally got chiplets working in a scalable way.

 

[marees]

This rumor has popped up before. I suspect it is bad info. The original implied that AMD finally got chiplets working in a scalable way.

If N48 is a chiplet then why were N43 & N42 scrapped

AMD could have made a viable successor to N31

I don't buy it all. My understanding is that originally N43 was chiplet. But AMD redesigned it as monolithic N48. The clue is in the sequence of numbers. 41, 42, 43, 44 was original design then first 2 are scrapped & 43 redesigned as (monolithic?) 48

 

[blckgrffn]

If N48 is a chiplet then why were N43 & N42 scrapped

AMD could have made a viable successor to N31

I don't buy it all. My understanding is that originally N43 was chiplet. But AMD redesigned it as monolithic N48. The clue is in the sequence of numbers. 41, 42, 43, 44 was original design then first 2 are scrapped & 43 redesigned as (monolithic?) 48

Exactly my thoughts.

I mean, it would be epic if AMD rolled in and a highly functional chiplet design working and offered a range of options not covered by the leaks.

That seems pretty unlikely though.

 

[static shock]

Nevermind.



Disclosure: I have no NDA to break, i don't work for AMD and any of its AIB partners and my sources are surely heavy readings on forums around the web.

Please stop spamming and trolling, it is not permitted here. You have 10 days off to reevaluate your posting style.

Mod DAPUNISHER

 

[jpiniero]

and my sources are surely heavy readings on forums around the web.

Voices in your head?

 

[static shock]

Voices in your head?

Aliens.

 

[marees]

Mobile lineup

https://twitter.com/x/status/1855655299484643721

AMD Radeon RX 8000 “RDNA 4” Mobile GPUs Include 16, 12 & 8 GB Variants, Up To 175W TGPs​

AMD Radeon RX 8000 "RDNA 4" Mobile GPUs Include 16, 12 & 8 GB Variants, Up To 175W TGPs

The first specifications of AMD's upcoming Radeon RX 8000 "RDNA 4" Mobile GPU lineup have been revealed by Golden Pig Upgrade Pack.

wccftech.com

So what are the SKUs ?

N48xtx 16gb 64cu — $500-$600
N48xt 16gb 56cu — $400-$500
N48xl 12gb 48cu — $350-$400
N44xt 8gb 32cu — $250-$300
N44xl 8gb 28cu <$250 ?

 

[SolidQ]

So Blackwell have more efficiency, than RDNA4?

 

[gaav87]

I heard some rummors of RDNA3+ high end cards (faster then rdna4) or is this bs anyone ?

 

[leoneazzurro]

So Blackwell have more efficiency, than RDNA4?

Just like current 4090M is more efficient than 4080M, you trade chip die size for performance. Efficiency of an architecture is depending on many factors. We could speak about efficiency if comparing dies with similar size and resources at the same power budget. That said, Nvidia has the lead currently and my guess is that it will likely keep an advantage with the next gen, then the unknown will be how good or bad RDNA4 will be compared to the green side's next gen.

 

[Tuna-Fish]

So Blackwell have more efficiency, than RDNA4?

I would be surprised if it wasn't. NV has an architecture advantage this generation, and RDNA4 doesn't seem to be changing that much in the computation side itself.

 

[poke01]

RDNA4 will be more efficient than RDNA3 but more efficient than Nvidia we have to wait for reviews