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

 

[poke01]

They were for preserving the CUDA onion.

Nah.. CUDA is useless for gaming. Those Tensor cores in GeForce were for DLSS. Unlike AMD in 2018, Nvidia had a proper software roadmap.

 

[adroc_thurston]

Nah.. CUDA is useless for gaming. Those Tensor cores in GeForce were for DLSS. Unlike AMD in 2018, Nvidia had a proper software roadmap.

Okay I get it, you don't know how CUDA CC model works. Don't.

 

[poke01]

Okay I get it, you don't know how CUDA CC model works. Don't.

sigh, you don’t. Why can’t you admit that Tensor cores in GeForce were for DLSS?

No gamer isn’t going to compile CUDA code……..

 

[adroc_thurston]

sigh, you don’t. Why can’t you admit that Tensor cores in GeForce were for DLSS?

They were for making sure CUDA CC onion model is preserved.

 

[poke01]

They were for making sure CUDA CC onion model is preserved.

you are the first person I’m hearing that from. so your saying that DLSS is a byproduct of preservation?

 

[GodisanAtheist]

NERD FIGHT!

 

[adroc_thurston]

so your saying that DLSS is a byproduct of preservation?

yeah. they had matrix cores and had to invent something to use them.

 

[poke01]

yeah. they had matrix cores and had to invent something to use them.

otherwise they would be wasted silicon if they didn’t. So they preserved CUDA and added the DLSS lockin. Marketing must have had a field day.

 

[poke01]

I see nothing wrong with whatever power limit they (or aibs) impose

RDNA4 looks to be very efficient. I’m just very picky about GPU power limits as it’s get toasty in Australia during the summers.

It’s probably only the OC models going over 350 watts which I don’t care about. Standard XT model should be below <300 according to leaks.

 

[adroc_thurston]

Marketing must have had a field day.

Well if only. DLSS1 was godawful!

 

[CouncilorIrissa]

AMD should probably sell the GPU part of their company. To someone who either will spend the money to compete or knows what they are doing. Maybe resurrect ATI and setup shop back in Canada. Those were very good times, ATI was great. The conversion rate would result in very good GPU prices here is the USA.

Someone tell me when the last time an AMD GPU release went well or according to plan? I think it was back when they released the 7850/7950/7970 back in 2011. Those were Nvidia killer cards.

Ah yes, they should sell their DC money printer. Sure.

 

[ajsdkflsdjfio]

This doesn't include R&D spend.

Nvidia barely spends 2x the R&D of AMD, while having 9x the marketshare, higher ASP and operating margins, and generally better savings due to scale. I'm pretty sure that Nvidia pricing the xx90 at 25% less isn't going to come close to hurting their bottom line.

BTW AMD gaming department is still in the black for 2024 despite it being one of their worst years yet and with kneecapped semi-custom revenue. With that in mind, how well do you think Nvidia's gaming division is doing right now?

 

[DaaQ]

Nvidia barely spends 2x the R&D of AMD, while having 9x the marketshare, higher ASP and operating margins, and generally better savings due to scale. I'm pretty sure that Nvidia pricing the xx90 at 25% less isn't going to come close to hurting their bottom line.

BTW AMD gaming department is still in the black for 2024 despite it being one of their worst years yet and with kneecapped semi-custom revenue. With that in mind, how well do you think Nvidia's gaming division is doing right now?

Not all that great since it is AI focused.

 

[ajsdkflsdjfio]

Not all that great since it is AI focused.

Are you talking about profit or what because AI focused or not, Nvidia gaming is going to do well.

 

[DaaQ]

Are you talking about profit or what because AI focused or not, Nvidia gaming is going to do well.

Nvidia is all AI focused. EDIT: Gaming is an afterthought.

 

[poke01]

Well if only. DLSS1 was godawful!

True, it wasn’t till DLSS2 it was good. Was the same with FSR1, FSR2 was much better and FSR4 is basically on par with CNN DLSS.

EDIT: Didn’t help with DLSS1 you had to train the games. DLSS2 made ML upscaling usable and proved it was worth it

 

[poke01]

Nvidia is all AI focused. EDIT: Gaming is an afterthought.

Gaming still makes a dent on their financial reports. Nvidia is definitely AI first tho, that’s where the money is right now. It’s the same with AMD. DC and AI is where the R&D is mostly spent.

 

[ajsdkflsdjfio]

Nvidia is all AI focused. EDIT: Gaming is an afterthought.

I agree AI is their bread and butter, don't really know what if anything incentivizes them to care about gaming at all, but the market share and revenue numbers don't lie. Nvidia has only increased gaming market share at the same time as increasing prices/margins since they've started to focus on AI.

Maybe in your original statement you meant the gaming revenue is "not all that great" in comparison to datacenter?

I mean their gaming revenue for 2024 Q3 was 3.3 billion, almost as much as AMD's gaming revenue for the entirety of 2024. Their yearly gaming revenue for 2022, 2023, and probably 2024 matches or exceeds Nvidia's entire yearly revenue from 2018. I get that AI has out shadowed gaming, but doesn't mean their gaming division isn't also doing well.

 

[Hans Gruber]

Ah yes, they should sell their DC money printer. Sure.

AMD money machine is their server chip business.

 

[adroc_thurston]

AMD money machine is their server chip business.

Guess what, graphics is part of that.

 

[Hans Gruber]

Guess what, graphics is part of that.

Explain why AMD cannot compete with Nvidia effectively in the GPU segment?

 

[adroc_thurston]

Explain why AMD cannot compete with Nvidia effectively in the GPU segment?

Wdym can't compete, N48 is pretty damn good.
Just good GFX IP, down to the very basics.

 

[Keller_TT]

Wdym can't compete, N48 is pretty damn good.
Just good GFX IP, down to the very basics.

It better be, isn't it. That's what one must expect after a disaster gen last time?
I have my rough reference set w.r.t Navi 21 6800 at 250W accounting for full 16GB VRAM in die area and 3/4 the full spec otherwise.
Navi 32 was a disaster after +54% PPW claim and a full node transition, and a good 5nm node from TSMC. Far less transistors and clock bump than ideal.

This time, for N4P power efficiency, which is a good step from their 5nm for +6% density, N48 at 250-260W must comfortably best 7900 XT even including for heavily revised CUs for ML and BVH. It's well over 4 years since RDNA2.
Efficiency matters a great deal where I live.

 

[adroc_thurston]

That's what one must expect after a disaster gen last time?

Wasn't even that much of a disaster, just a miss plus the top config was undersized.

even including for heavily revised CUs for ML and BVH

They're revised for per-offchip bit performance.

 

[Keller_TT]

Wasn't even that much of a disaster, just a miss plus the top config was undersized.

Ironically enough, the top chip was the best of the lot that just didn't scale down for other configurations.
It was a big miss too as the 7900 XT MBA should've been the XTX in performance for a good architecture.
This time, looking at listed AIB specs and power levels, I'm like wth is happening?