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

 

[ToTTenTranz]

They could always do that? GEMM is GEMM.
A 'tensor core' is just a hardwired unprogrammable GEMM accelerator.

They could, but I guess they can do it faster on the shaders now, maybe akin to RDNA3 > RDNA2.

They very much can't which is why 5090 price is right in line with the pretty massive BOM bump.

They're asking for exactly the same price-per-FPS as the old top-end model, which in GPU terms is pretty much asking as much as you want (unless there's a crypto craze going on). More than that and they're pretty much competing unfavorably against their own previous generation.

 

[adroc_thurston]

They could, but I guess they can do it faster on the shaders now, maybe akin to RDNA3 > RDNA2.

They also always could do that? They had double rate FP16 ALUs in H100 I think.

They're asking for exactly the same price-per-FPS as the old top-end model,

Because costs are not going down.

 

[ToTTenTranz]

Because costs are not going down.

Costs may be higher, but the RTX4090 had a gigantic markup and so does the RTX 5090 -> because it lacks competition.

 

[ajsdkflsdjfio]

Because costs are not going down.

Cost's aren't going down sure, so the increase of pricing from 4090->5090 might be justified since the costs also jumped a proportionate amount. But the 4090 pricing was already ludicrious to begin with so you're going from inflated price to inflated price. If RDNA3/4 provided competition at the high end you would not see this pricing from Nvidia.

In the first place aren't graphics cards supposed to get cheaper at the end of their market cycle? I see this happening in the mid-low end Nvidia cards and pretty much all the cards for AMD, but not for the 4090/4080. I wonder why?

 

[adroc_thurston]

but the RTX4090 had a gigantic markup

no? it's a chungus n4 die.

 

[ajsdkflsdjfio]

no? it's a chungus n4 die.

Whatever you think the die costs, it's not 700-900 more than the navi31 package.

Oh yea also rtx 4080S die size 379mm2 charging 1k+ even 2 years after it's release, and supposedly the 9070xt is ~350mm2 ish up to 380(same vram, similar TDP and cooler sizes), and people expect it to be around 530-550. Saying Nvidia isn't up-charging the high-end due to lack of competition is an absolute cope.

 

[DavidC1]

Well GPU-Z leaked clocks match powercolor oc bios
View attachment 114825

Well there goes my theory that they can use clocks for differences in CU count. If it's to match 7900XT it would have to have a huge per CU bump.

no? it's a chungus n4 die.

This logic only makes sense from caring % margin perspective. From an absolute revenue and net revenue standpoint it doesn't, because at certain point silicon business is like a money printing business.

So if it costs $200 to make and you sell for $2000, and the next gen costs $300 to make, you need to sell it for $3000 for the same margins, but this is a very poor excuse for bumping prices by 50%.

 

[jpiniero]

Saying Nvidia isn't up-charging the high-end due to lack of competition is an absolute cope.

The problem is that even competitive products isn't enough to get people to buy.

 

[Hans Gruber]

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.

 

[ajsdkflsdjfio]

The problem is that even competitive products isn't enough to get people to buy.

What is your definition of getting people to buy. Ofc there will always be a portion of people that just buy Nvidia because of mindshare, marketing, and Nvidia inventing new metrics to unfairly judge their own products vs competition, but competitive products will still help get more people to buy. Look at RDNA1/2/3 market share if you think competitive products won't get people to buy AMD.

Either way this is separate to the discussion of the lack of competition at high-end. AMD being competitive in mid-low end is good and will help stave off rising Nvidia prices, but the High-end will remain unaffected for the most part by competition in the mid-low end.

 

[Hail The Brain Slug]

Someone tell me when the last time an AMD GPU release went well or according to plan?.

Are we just going to pretend RDNA2 wasn't an incredibly well executed and successful product line, at least with regards to N21 and N22? What about that generation would you describe as "didn't go well" or "didn't go according to plan"?

 

[Novacius]

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.

That would be suicide for their mobile CPU business, server GPU business (which made over $5 billion in 2024) and for their semi custom business. Not everything revolves around PC gaming.

 

[branch_suggestion]

All_The_Watts (formerly rogane I think) says 9070XT is on par or slightly faster than the 4070 Ti Super but not quite the 4080. It also hits >3 GHz with power above 3000W.

9070 non-XT is faster than 4070 Super, but slower than 4070 Ti Super. Clocks in the 2 GHz range with power also in the 200s W.

https://twitter.com/x/status/1878490567014777077

They were wrong 6 months ago, no reason to believe them now.
Missing die size by 100mm^2, fraudulent.

 

[adroc_thurston]

at least with regards to N21 and N22

N23 too. 6600 was such a value banger.

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.

they do not exist to serve the toddler googoo gaagaa market you live in. GFX IP is important for like a billion other usecases. Get real or get gone.

Someone tell me when the last time an AMD GPU release went well or according to plan?

RDNA2 4 years ago was flawless execution is a market hollowed out by mining. Too bad!

 

[Hail The Brain Slug]

N23 too. 6600 was such a value banger.

You're right N24 is where it really fell off.

 

[ToTTenTranz]

no? it's a chungus n4 die.

I think this argument doesn't add up. You can fit up to 100x AD102 chips in a 300mm wafer. Even if yields are mediocre for the cutdown AD102, say 70%, you're still getting 70 chips on a wafer that costs $15k. That means each chip costs ~$215 from the wafer.

No matter how much you put in from RAM, PCB, voltage regulation, packaging etc. nothing is going to make that card's BOM+assembly+distribution costs above $600, and then Nvidia put a 170% markup over it.

It's a gigantic markup no matter what angle you take, probably akin to luxury and sports cars and it's only possible because it's lacking competition.

Someone tell me when the last time an AMD GPU release went well or according to plan?

RDNA2 went pretty well.

 

[adroc_thurston]

No matter how much you put in from RAM, PCB, voltage regulation, packaging etc. nothing is going to make that card's BOM+assembly+distribution costs above $600, and then Nvidia put a 170% markup over it.

This doesn't include R&D spend.
Again, AD102/GB202 pricing is fine. It's TU102/GA102 that were the outliers.

 

[ToTTenTranz]

This doesn't include R&D spend.

Why not include software development costs while you're at it? Because those are always diluted throughout the generation. Always have been.

Again, AD102/GB202 pricing is fine. It's TU102/GA102 that were the outliers.

Lol AD102/GB202 pricing "is fine". It's:

TU102
GA102
GP102
GM200
GK110
GF110
GF100

that were the outliers.

The only thing that sets the latest two from all the others is... lack of competition in their performance bracket. Just like everyone is saying.

 

[poke01]

hopefully AMD didn’t increase power consumption too much and kept power consumption in check. These >3GHz clocks sound like it’s going to be over 350 watts.

 

[poke01]

You're reading too much into puff marketing for a very very very underwhelming uarch.

Ehh, Blackwell will age like Turing. It’s a good architecture just not for pure raster.

I remember when everybody crapped on Turing but come 2025, if you have >RTX2070 card it’s still usable with DLSS4.

 

[adroc_thurston]

Blackwell will age like Turing

Badly?

It’s a good architecture just not for pure raster.

It's good because it's not worse than Ada.

if you have >RTX2070 card it’s still usable with DLSS4.

The selling point of Turing was le raytracing and unfortunately you're not gonna run any on Turing

 

[Hans Gruber]

N23 too. 6600 was such a value banger.

they do not exist to serve the toddler googoo gaagaa market you live in. GFX IP is important for like a billion other usecases. Get real or get gone.

RDNA2 4 years ago was flawless execution is a market hollowed out by mining. Too bad!

RDNA 1 had no fanfare and actually released really well. RDNA 2 fell a bit flat from the start and recovered nicely. The problem with RDNA 2 was Covid and the mystery disappearance of a ton of GPU's.

 

[poke01]

Badly?

Yep, at least the cards below 5090. The super refresh will be fine due to high VRAM.

The selling point of Turing was le raytracing and unfortunately you're not gonna run any on Turing

that was the marketing, but those Tensor cores were for DLSS in mind. Although it took DLSS 2 years to be usable.

 

[adroc_thurston]

RDNA 2 fell a bit flat from the start and recovered nicely.

It did not.

RDNA 1 had no fanfare and actually released really well

It did not, Navi10 had nasty display core bugs.

The super refresh will be fine due to high VRAM.

Moar VRAM doesn't solve the baseline perf uplift being what it is.

Tensor cores were for DLSS in mind.

They were for preserving the CUDA onion.

 

[PJVol]

hopefully AMD didn’t increase power consumption too much and kept power consumption in check. These >3GHz clocks sound like it’s going to be over 350 watts.

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

Navi10 had nasty display core bugs.

Issues like "failed to resume from suspend/s3" iirc have been around since GCN and are still not fully resolved.

This doesn't include R&D spend.

And the whole media machine, including their lapdog tech outlets.