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

 

[gdansk]

You have it backwards. NVidia has a $1000 in their lineup because they know AMD doesn't have anything to fit into it.

It is circular in any case. AMD has nothing to slot into it because they expected a GB202 cut. But Nvidia didn't bother making that GB202 cut because AMD has nothing there. It takes two to avoid a tango. 🤷‍♂️

 

[adroc_thurston]

AMD has nothing to slot into it because they expected a GB202 cut

No, it's because the kilobuck target was covered by the canned navi42.

 

[eek2121]

Regarding area performance efficiency, it actually sounds like NVIDIA and AMD are neck and neck here, but we will see. It will depend on the final numbers from AMD.

GB203 used in the 5070ti is 378mm2 while GB205 is 263mm2. Navi48 should be around 330-340mm2 and should perform somewhere in between the two, if rumors are even within the ballpark of being accurate.

I guess we will find out.

 

[Heartbreaker]

You have it backwards. NVidia has a $1000 in their lineup because they know AMD doesn't have anything to fit into it.

Chip designs were locked in ages ago when neither side knew what the other was doing, and they were just guessing at best.

Given that both decided to significantly limit progress this generation, is a sign that the underlying economics of forward progress are extremely challenging.

 

[MrTeal]

Don’t forget: NVIDIA also has 2 consoles, the switch and the switch 2. PS5/Xbox sales are winding down, while Switch 2 production is likely ramping up. NVIDIA also has higher revenue SKUs like the 4090 (and soon, 5090) and some other stuff going on.

NVIDIA can get away with selling a $1,600-$2,000 card where AMD’s price ceiling is half that at most, and likely most of the product they sell sells for a third of that. With how margins are added on, AMD makes substantially less money for the same amount of units sold.

I am not disagreeing with anything you are saying, just that the numbers aren’t an apples to apples. There is no doubt in my mind NVIDIA sells more units than AMD.

Fair enough, and I agree. We don't have nearly enough information to get a good ratio on GPUs sold just from the financials, it's just a very obvious sanity check that the disparity is so large that there's no credible way to twist it that AMD sells more GPUs than Nvidia no matter what data comes out from a single German retailer.

In those 462M there s no semicustom and Xbox chips since they are working on a new gen, both Sony and MSFT are living on their remaining inventories, so we can conclude that this amount to 3M GPUs at 150$ ASP, to compare to JPR s number who state that there s about 8M DT GPUs sold per quarter, and AMD sold almost exclusively DT GPUs.

What? That's nothing but a bunch of conjecture. AMD explicitly includes semicustom in their gaming revenue.

• Gaming segment revenue was $462 million, down 69% year-over-year and 29% sequentially primarily due to a decrease in semi-custom revenue.

We don't know what the split is, but it's certainly not 0 since the PS5 Pro was just recently released and Sony sold more than 4M units in November alone.

Even then, let's roll with your assumptions and say AMD had no semi custom or mobile revenue and all it was all from 3M dGPUs in Q3. Nvidia must be pretty happy that their remaining share of 5M GPUs brought in $3+B in revenue.

 

[adroc_thurston]

Chip designs were locked in ages ago when neither side knew what the other was doing, and they were just guessing at best.

?

Given that both decided to significantly limit progress this generation

? What.

is a sign that the underlying economics of forward progress are extremely challenging

It's a sign of NV having to try to make a better SM. Next.

 

[Abwx]

Even then, let's roll with your assumptions and say AMD had no semi custom or mobile revenue and all it was all from 3M dGPUs in Q3. Nvidia must be pretty happy that their remaining share of 5M GPUs brought in $3+B in revenue.

The difference is that Nvidia sell much more in house FE whole cards while AMD mainly rely on OEMs, this way GPU sales numbers are much inflated, FI a 600$ card has a GPU that could be sold 150$ to an OEM, so by selling whole cards you can increase this number by 4x, guess that you can gladly cut those $3+B by a 2 factor.

 

[gaav87]

In other news AIB's are pissed at AMD xD

 

[coercitiv]

In other news AIB's are pissed at AMD xD

I am not an AIB!

 

[gdansk]

In other news AIB's are pissed at AMD xD

And I saw a report that a rice bag fell over too.

 

[Mopetar]

Chip designs were locked in ages ago when neither side knew what the other was doing, and they were just guessing at best.

It's much easier to crate a new bin or adjust your prices than it is to create a new die. If AMD had a larger die, Nvidia would have done things differently. There's no reason that the 5080 had to be GB203.

Look back at Ampere where the 3080 was the same GA102 die as the 3090 because RDNA2 was far more competitive and had a full product stack. If AMD had something above the 9070 XT you can bet Nvidia would have a different lineup with various price adjustments.

 

[Heartbreaker]

It's much easier to crate a new bin or adjust your prices than it is to create a new die. If AMD had a larger die, Nvidia would have done things differently. There's no reason that the 5080 had to be GB203.

Look back at Ampere where the 3080 was the same GA102 die as the 3090 because RDNA2 was far more competitive and had a full product stack. If AMD had something above the 9070 XT you can bet Nvidia would have a different lineup with various price adjustments.

Ampere GA102, was a significantly smaller die, on a much cheaper process. That makes it much more tenable, if not outright necessity, to use the biggest die for 080 cards.

GB202 OTOH is an insane size on a much more expensive process, too expensive or 080, maybe they could do a 080 Ti later though.

 

[Mopetar]

GA102 was still large (over 600 mm^2) and NVidia would sell a cut down GB202 die if they had to. They will eventually, but there's no need to release what will eventually be a 5080 Super or Ti (or a 5090 D) now when there's nothing competing at that performance level.

The die is what it is, but products and pricing can be changed to adapt to the realities of the market when the die ships. No doubt AMD would have charged more for RDNA3 products if they were more capable and could go toe to toe with Nvidia cards like RDNA2 was able to.

 

[CastleBravo]

GA102 was still large (over 600 mm^2) and NVidia would sell a cut down GB202 die if they had to. They will eventually, but there's no need to release what will eventually be a 5080 Super or Ti (or a 5090 D) now when there's nothing competing at that performance level.

RTX 5090 is already a cut down GB202, so most dies with a defect can still be used for a $2k 5090. Even a defect in the memory controller could still be used in a future 448 or 384 bit 5090 variant with faster vram clocks. With 3GB 37gbps DDR7 modules, they could do a 36GB 384bit bus "5090 Super" and sell it as an upgrade over the base 5090. I don't see a future GB202 based 5080 variant happening, at least not without a 90-class price tag.

 

[Keller_TT]

NVIDIA can get away with selling a $1,600-$2,000 card where AMD’s price ceiling is half that at most, and likely most of the product they sell sells for a third of that. With how margins are added on, AMD makes substantially less money for the same amount of units sold.

AMD has to pay its dues before it can think of getting in by slightly undercutting NV. The problems with Radeon are well known to those in AMD. NV for Windows just works while the scope of a GPU has increased. In a software defined world, AMD has to translate promise to execution and consistency.

They bid their time knowing their rights and wrongs with CPUs and finally it came together for them just as Intel had fab issues.
NV aren't going to have Fab issues and are so much more well-oiled than Intel was. AMD has to settle for No. 2 and a strong No. 2 at that. Don't let dGPU consumer market share drop below 30%, play the DC game well, and polish ROCm to close the gap to CUDA in training.

 

[adroc_thurston]

AMD has to translate promise to execution and consistency.

They need to translate it into a 300-something CU chungus.

while the scope of a GPU has increased.

it did not. it plays video games™.

AMD has to settle for No. 2 and a strong No. 2 at that

Lmao

Don't let dGPU consumer market share drop below 30%

lmao

 

[Keller_TT]

Lmao

Schritt für Schritt, Alter Schwede!

What's your idea? Try to fly before you can accelerate, or just not try at all?

 

[SolidQ]

Return Flagship?

https://twitter.com/x/status/1880118882985463887

 

[Keller_TT]

Return Flagship?

https://twitter.com/x/status/1880118882985463887

N3E for Zen6 and UDNA seems very suspicious and unlikely. N3P would already be 2 years in HVM by the time they arrive, and the price adjusted node for TSMC as they ramp up N2.

First let's see RDNA4 before even thinking about the next hype cycle. lol.

 

[Bryo4321]

Return Flagship?

https://twitter.com/x/status/1880118882985463887

I expect that simply means a bigger die targeting the ~$1000-ish price range. Not some giant monstrosity. Basically what adroc has already alluded to.

 

[adroc_thurston]

Try to fly before you can accelerate, or just not try at all?

You don't fly. You go straight to space.

 

[del42sa]

GPU: UDNA N3E - Return of Flagship?

sure, return of disappointment ...

 

[poke01]

sure, return of disappointment ...

I actually think think AMD will beat Nvidia architecture wise next gen. AMD has better IP for gaming

 

[adroc_thurston]

sure, return of disappointment ...

Cope. They make da best™ IP.
Will they make an interesting config with it? Hell no.

 

[coercitiv]

Cope. They make da best™ IP.
Will they make an interesting config with it? Hell no.

Another Medium Die