Question Speculation: RDNA3 + CDNA2 Architectures Thread

[uzzi38]

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 is much bigger than from RDNA1 to RDNA2. We should expect many big improvements in GFX11. 🤔" / Twitter

 

[Heartbreaker]

For example, the 4080 has a much smaller chip than the 3080, so even if they pay twice as much to TSMC as they paid to Samsung, these prices are pretty much impossible to explain. Especially since other parts of the card certainly won't have had a 200% increase in cost (VRAM, cooling, PCB, etc).

You have to get away from looking at die size and focus on transistor count.

3080 was bottom cut of a 28 Billion Transistor chip. Margin in this usage would be MUCH less than using them for 3090.

4080 is the top cut of the 45 Billion transistor chip, it doesn't have a higher margin outlet, so they kind of need higher margins here.

3080 also only had 10GB of VRAM vs 16GB on 4080.

So there is likely significant cost growth in the two most expensive components, GPU chip and VRAM, plus some margin increase.

It's not one or the other. It's likely both higher input costs and higher margins.

 

[Aapje]

You have to get away from looking at die size and focus on transistor count.

A partially disabled chip costs just as much as a full chip. So unless you want to argue that they sold the 3080 at a loss, they greatly increased the margins.

So there is likely significant cost growth in the two most expensive components, GPU chip and VRAM, plus some margin increase.

And yet somehow Sony can deliver an entire computer for $500 with 16 GB. That extra 6 GB doesn't really cost that much.

But you agree with me that they increased margins, so what are we even arguing about.

 

[Heartbreaker]

A partially disabled chip costs just as much as a full chip. So unless you want to argue that they sold the 3080 at a loss, they greatly increased the margins.

It's not that simple. When you are making a business plan for a new chip, what matters is the ASP, or average selling price. NVidia will be selling 3080 chips for a lower price than 3090 chips and margins will be lower on 3080 chips. If they were only going to sell it as 3080 prices, they probably wouldn't have had a business case and even with 3090 raising ASP they likely were NOT happy with 3080 margins.

And yet somehow Sony can deliver an entire computer for $500 with 16 GB. That extra 6 GB doesn't really cost that much.

Yes by having essentially no margin at all, and making the money on the game sales.

But you agree with me that they increased margins, so what are we even arguing about.

You act like increasing margins is a crime, but it's likely there were selling 3080 at relatively lower margins, and those margins would have already been propped up by 3090 margins. This time they aren't mitigated by 4090 margins and they want closer to their average margins, than a lower margin. Nothing nefarious going on.

 

[gdansk]

The RTX 4080 is cheaper to make the RTX 3080. The die itself anyway. The rumored estimates for N5 wafer costs are highly overstated.
And memory prices are down a lot since 2020-2021 too.
The $500 price increase is 80% margin seeking, 10% inflation and 10% oversized coolers.

 

[Aapje]

@guidryp

You are reasoning yourself into a narrative where Nvidia should ask higher prices for full dies of cheaper chips, than for cut dies of expensive chips. By that reasoning, we should demand that they put AD102 into the 4050, so the price can be lower...

You act like increasing margins is a crime

Excessive margins are a sign of a poorly functioning market. It is actually a reason for the government to intervene, if they care.

In any case, as consumers, we have the power to say: no.

 

[jpiniero]

The RTX 4080 is cheaper to make the RTX 3080. The die itself anyway. The rumored estimates for N5 wafer costs are highly overstated.

N7/N6 prices may tank at some point if demand doesn't recover... N5/N4 no lol.

 

[gdansk]

N7/N6 prices may tank at some point if demand doesn't recover... N5/N4 no lol.

Based on AMD's now current pricing of a 300mm2 part at $800 including 5 other dies and new packaging techniques yes lol. Nvidia could have targeted $800 AD103 with the same margin as the RTX 3080.

And based on TSMC's >10% drop in average wafer price last quarter. All they have to achieve to be cheaper than the RTX 3080 is 0.66x cost per transistor of Samsung's 8nm in 2020-2022. At the height of the "Global Silicon Shortage".

 

[jpiniero]

Based on AMD's now current pricing of a 300mm2 part at $800 including 5 other dies and new packaging techniques yes lol.

That was because AMD/retailers had no choice if they wanted to move any 7900 XT. And that's why N32 is MIA, and won't be released unless mining returns real soon.

The 7900 XT's been under $800 btw.

 

[gdansk]

That was because AMD/retailers had no choice if they wanted to move any 7900 XT. And that's why N32 is MIA, and won't be released unless mining returns real soon.

That doesn't mean that the RTX 4080 is priced based on costs of N5 at all. AMD targets at least 50-60% margin at $900 for the 7900XT. The implication for the RTX 4080 margin? Absurd.

AD103 pricing is clearly wrong. AD102 is 1.6x as large and only 1.33x more. AD104 is 0.8x as large but goes for 0.5x the cost.
The pricing isn't based on TSMC's costs. All I can say.

 

[insertcarehere]

And based on TSMC's >10% drop in average wafer price last quarter. All they have to achieve to be cheaper than the RTX 3080 is 0.66x cost per transistor of Samsung's 8nm in 2020-2022. At the height of the "Global Silicon Shortage".

I doubt that they have much of a cost per transistor advantage at all vs SS8. SS8 was an older process which didn't have nearly the demand that TSMC's processes attracted, in terms of external high performance customers for the process Nvidia was the only one. Proportedly wafer pricing reflected this with at least 30% haircut off what N7 wafers fetched for, with other extra concessions to boot.

And besides, RTX 3080 was what they did with the GA102 dies that couldn't go to 3090s, 3090 Tis, and other cuts. For Nvidia, ideally they're not putting the best dies into $699 3080s but $1499+ 3090s.

 

[gdansk]

I doubt that they have much of a cost per transistor advantage at all vs SS8. SS8 was an older process which didn't have nearly the demand that TSMC's processes attracted, in terms of external high performance customers for the process Nvidia was the only one. Proportedly wafer pricing reflected this with at least 30% haircut off what N7 wafers fetched for, with other extra concessions to boot.

And besides, RTX 3080 was what they did with the GA102 dies that couldn't go to 3090s, 3090 Tis, and other cuts. For Nvidia, ideally they're not putting the best dies into $699 3080s but $1499+ 3090s.

We're working from the same rumors. 30% off N7 would be about 1/2 the price of N5. N5 AD103 has nearly 3x density of SS8 GA102.
That's already close to 0.66x cheaper per transistor. Meaning AD103 is basically already cheaper to make. Except as you say all the extra cut downs. But AD103 is small with less wafer wastage and TSMC has better yields and it is cut down in all its incarnations so far.

 

[PJVol]

It may very well be that the actual 7900xtx was destined to be a 2-chiplet card that was supposed to compete with NVIDIA’s best

This doesn't seem to be provided for the current arch, rather for the next rdna iteration and just in the form of GCD partitions.
On the other side, given the gfx11 core supports up to 8 SE's, the 128 CU part would be really competitive, although way beyond reasonable efficacy.

 

[insertcarehere]

We're working from the same rumors. 30% off N7 would be about 1/2 the price of N5. N5 AD103 has nearly 3x density of SS8 GA102.
That's already close to 0.66x cheaper per transistor. Meaning AD103 is basically already cheaper to make. Except as you say all the extra cut downs. But AD103 is small with less wafer wastage and TSMC has better yields and it is cut down in all its incarnations so far.

If SS8 is half the price of N5 and 30% off N7, this implies N5 is only like 30% more expensive than N7.

Which doesn't jive with what we see AMD doing, I. E moving everything bar the absolute essentials that they can get away with off cutting edge nodes.

RDNA3 as it stands is almost certainly a less performant and less efficient architecture that takes more transistors than it would've been under monolithic TSMC N5, simply due to the space and energy needed to move data off/on the different dies. Not to mention that the specialized packaging needed for chiplets isn't itself free. Why would AMD go to all this trouble and comprimise if moving significant chunks of the die area into trailing edge nodes doesn't even yield significant cost decreases in wafer costs, which is what you're implying here.

 

[Aapje]

Samsung also reportedly had very poor yields, which is almost certainly why they put GA102 in the 3080, resulting in relatively low margins for that chip. You can see indications for poor yields in the entire line-up, with them using a lot of deeply cut chips. Using GA102 even for 3070 Ti's, using GA103 for 3060 Ti's, using GA104 for 3060's and using GA106 for 3050's.

So even if TSMC is way more expensive than Samsung, part of that will be offset by much better yields, allowing them to sell more full (or less cut) dies. All else being equal, selling full or less cut dies results in better margins.

 

[jpiniero]

AMD Radeon RX 7600 expected to cost €349 in France - VideoCardz.com

AMD Radeon RX 7600 now has a rumored price According to Cowcotland the price for the upcoming mid-range RX 7600 GPU is €349. The French site, citing their sources, claims that the price for custom Radeon RX 7600 has been settled at €349. A quick conversion from EUR to USD tells us that the card...

videocardz.com

Videocardz has a claim that the official MSRP in France will be 349 Euros. Which as Videocardz notes means that the MSRP is probably $300-330.

You can still get a 6700 XT for not much more than that. That's the problem with releasing it now instead of whenever RDNA 2 is wiped out.

 

[Aapje]

That would be DOA. So the prices are then going to drop.

 

[KompuKare]

That would be DOA. So the prices are then going to drop.

You meant a repeat of most recent Radeon launches then!?

 

[jpiniero]

That would be DOA. So the prices are then going to drop.

That's why I'm saying that this is just an exercise to dump the N33 dies because they overestimated N33 laptop demand. The shipments are probably not that much.

 

[coercitiv]

That's why I'm saying that this is just an exercise to dump the N33 dies because they overestimated N33 laptop demand. The shipments are probably not that much.

So AMD is, at the same time, having too much unsold N33 inventory resulting in them dumping chips in the desktop market but also having "not that much" excess inventory, resulting in them dumping at high prices. Fascinating.

 

[Aapje]

That's why I'm saying that this is just an exercise to dump the N33 dies because they overestimated N33 laptop demand. The shipments are probably not that much.

If they really wanted to get rid of chips, you'd think that they start with low prices. AMD is also very conservative and as coercitiv said, they have very low inventory. Nvidia seems to have overproduced a lot and so did Intel, for Alder Lake, but AMD never seems to have big problems, even for initial flops, like AM5.

 

[Dribble]

Remember that Nvidia sees itself more as a software company, so it's natural it has a advantage there. Also that's an advantage built over decades like with CUDA. AMD until recently didn't have the financial power to even pay an equal work force. So being a follower is safe and not as costly.

This likely changes completely now that Xilinx people are taking over everything software.


Except for the 7000 launch (where something obviously went awry considering AMD didn't meet its own promises from shortly before) I'm actually quite happy with their results as a follower so far: Hardware was keeping up, and in software AMD managed to counter Nvidia's proprietary solutions with adequate open source counterparts.

Of course AMD's ambition should be to be more than a follower. Let's see what change Xilinx being in the mix makes in the next couple years.

It's because of x86, and it's not just money - Intel are in the same boat. Making money with x86 is easy - you provide hardware and pretty well everything else is provided on a plate. No need for big software dev teams, or spending lots of money trying to break into new markets, you can just ride the x86 gravy train. It's an easy life as long as x86 stays at the centre of everything....

 

[insertcarehere]

So much for the whole "you need low prices to move volume to amortize fixed costs" theory some were touting just a page ago lol.

 

[coercitiv]

Making money with x86 is easy - you provide hardware and pretty well everything else is provided on a plate. No need for big software dev teams, or spending lots of money trying to break into new markets, you can just ride the x86 gravy train.

You are aware Intel is currently operating at a loss, right?

 

[Aapje]

You meant a repeat of most recent Radeon launches then!?

Yes, but you'd think that they learn at some point.

It's actually pretty funny how some reviewers are now putting out videos telling AMD at what prices they will definitely give a negative review, since they are so frustrated at having to pan AMD, only for the prices to drop to more sensible levels afterwards. I really don't get why AMD does this to themselves.

If I was them, I would not just use the reviewers to sell product, but rope them into the pricing decisions as well. Just make a little survey where you present them with a bunch of different theoretical cards and let them give a price range they think is reasonable. It's not like AMD have innovative new features each new release that make a difference, so just giving a performance estimate plus VRAM amount is enough to allow reviewers to give a pricing suggestion.

Aside from better prices, this kind of thing also just flatters reviewers, so it will probably result in slightly better reviews as well.

 

[moinmoin]

It's actually pretty funny how some reviewers are now putting out videos telling AMD at what prices they will definitely give a negative review

Should have done so years ago with Nvidia, then we wouldn't even be in the current pricing situation.