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

 

[Joe NYC]

Zen 4 vcache is outside of AMD's control.

Edit: and I just saw this...

https://twitter.com/x/status/1583934983332249600

Don't understand how it can't be ready yet since N32 and N31 shares arch?

There was an article of how NVidia is planning on artificially creating shortage by limiting volume of 4090 chips.
NVIDIA focuses on H100 graphics cards The news said that the RTX 4090 is out of stock, and the price will rise until next year - fast technology - technology changes the future (mydrivers.com)

With Zen 4 V-Cache, it would be a natural shortage, if AMD releases it as soon as the first batch is packaged. Instead of waiting for a day when these chips could be launched in full volume.

What AMD could have learned (but did not) from the Zen 3 V-Cache fiasco is that for every day the chip is released after its competition (which was ADL), its irrelevance increases. When it finally launched in March - April, Zen 3 V-Cache achieved nearly full irrelevance.

PS: Low volume (with shortages) of Zen 3 V-Cache was possible in last December. But AMD, instead, sent those chips to Microsoft Azure cloud.

 

[Joe NYC]

If they started N32 production in Q4, March is a perfectly reasonable release date. You're probably underestimating production timelines.

True. But also, AMD did not go from 0 to 60(k wafers) in one quarter.

Zen 4 likely entered full production in Q2, and then there is whole Q3 between the Q2 and 60k wafers in Q4.

 

[Joe NYC]

Maybe. But if that were the case, why are prices here in Europe still way too high? 6900xt if it even is in stock is still $1000, 6800 xt is $750. They could very easily sell atom of n21 here but not at these prices.

Did you check AMD.com?

Here in the US:


Also, AMD updated the CPU prices, down to reality.

5800x: $289 (down from $449)
5800x3d: $399 (down from $449)

 

[biostud]

Did you check AMD.com?

Here in the US:
View attachment 69738

Also, AMD updated the CPU prices, down to reality.

5800x: $289 (down from $449)
5800x3d: $399 (down from $449)

Also some people in Europe forget we usually have a sales tax between 18-25%, which is why graphics cards are more expensive than in US.

 

[maddie]

There was an article of how NVidia is planning on artificially creating shortage by limiting volume of 4090 chips.
NVIDIA focuses on H100 graphics cards The news said that the RTX 4090 is out of stock, and the price will rise until next year - fast technology - technology changes the future (mydrivers.com)

With Zen 4 V-Cache, it would be a natural shortage, if AMD releases it as soon as the first batch is packaged. Instead of waiting for a day when these chips could be launched in full volume.

What AMD could have learned (but did not) from the Zen 3 V-Cache fiasco is that for every day the chip is released after its competition (which was ADL), its irrelevance increases. When it finally launched in March - April, Zen 3 V-Cache achieved nearly full irrelevance.

PS: Low volume (with shortages) of Zen 3 V-Cache was possible in last December. But AMD, instead, sent those chips to Microsoft Azure cloud.

The prioritizing H100 has more to do with the coming China sanctions. They won't be able to sell anymore after the cut-off date.

 

[maddie]

Also some people in Europe forget we usually have a sales tax between 18-25%, which is why graphics cards are more expensive than in US.

Yep. Euro falling + higher sales tax = rising or stagnant prices relative to US.

 

[H T C]

Also some people in Europe forget we usually have a sales tax between 18-25%, which is why graphics cards are more expensive than in US.

From AMD's Portuguese site: out of stock, though.

Portugal's tax is 23%, which would up the $799 price to $982.77. However, and as seen in the pic ... it's a slight more than that ... and it's not the $ to € conversion since they are virtually the same ...

 

[Joe NYC]

View attachment 69739

From AMD's Portuguese site: out of stock, though.

Portugal's tax is 23%, which would up the $799 price to $982.77. However, and as seen in the pic ... it's a slight more than that ... and it's not the $ to € conversion since they are virtually the same ...

Wow, that's a huge difference from the $799 price in the US.

But AMD has been updating the US prices very recently. Last few days. So, it is possible AMD will adjust European prices as well, and get more cards in stock, if AMD really still has a large number of N21 chips in stock.

 

[Joe NYC]

The prioritizing H100 has more to do with the coming China sanctions. They won't be able to sell anymore after the cut-off date.

NVidia does not have a shortage of wafers, the opposite. NVidia was negotiating with TSMC to delay its N5 wafer allocation.

So, there is not trade-off between the two segments of products (consumer, datacenter). NVidia has more than enough wafers for both.

 

[Joe NYC]

Speaking of amd.com website, I just noticed some nice bundle discounts, going up to $450 discount for buying CPU+GPU bundle.

I think I will be watching the amd.com when I am ready to build my new system, which will likely be 7800x V-Cache + 7800xt + B650E

 

[amenx]

There was an article of how NVidia is planning on artificially creating shortage by limiting volume of 4090 chips.
NVIDIA focuses on H100 graphics cards The news said that the RTX 4090 is out of stock, and the price will rise until next year - fast technology - technology changes the future (mydrivers.com)

AMD would be ecstatic if that were to happen. Nvidia would score an own goal by giving AMD free rein over the high tier segment.

 

[linkgoron]

Companies will always release the best product first, unless that is a total dud (Hey Intel).

Dumb Nvidia, releasing the 4090 and 4080 first.

IMO, the most annoying thing that AMD's marketing is doing is selecting stupid dates for releases - for example 7/7 for the 7nm Ryzen 2 and RDNA1, or November 3rd because of RDNA3. Just release it!!!

I miss the olden days when launches were launches, and people would discuss if it's a paper launch or hard launch. Now everything is just announcements of announcements. You wait for November 3rd but the cards are only really available weeks later. These days with AMD and Nvidia selling directly to consumers through websites - it should have been easier to hard launch stuff.

 

[maddie]

NVidia does not have a shortage of wafers, the opposite. NVidia was negotiating with TSMC to delay its N5 wafer allocation.

So, there is not trade-off between the two segments of products (consumer, datacenter). NVidia has more than enough wafers for both.

Throughput?

When Nvidia wanted to reduce their wafer commitment, do we know over what period of time? To sell these H100 die they must finish starting the last wafers several months before the sanctions begin and they would want to sell the full Chinese orders for the lifespan of the generation, so probably bringing forward late 2023 & 2024 orders.

 

[Yosar]

You just nailed It.

No. Simply no. 4090 was not held back by anything. It's a 4K. Any modern processor is enough.
If processor is slower it's slower for both cards. Not for one.

You can even compare data for the same games on both sides. Let's look at them.

Game	TPU benchmark - 6950 XT/4090 (5800X)	Computerbase benchmark - 6950 XT/4090 (12900K)	Difference for TPU / Difference for Computerbase (between cards with different processors)
CP2077	39 / 71.2	37 / 67.5	82.5% / 82.4 % (perfect)
Deathloop	72.5 / 143.1	60.6 / 121.6	97,3% / 100,6% (again almost perfect)
Doom	168.5 / 305	163.2 / 317.2	81,0% / 94,3%
Dying Light 2	58.1 / 100.1	55.7 / 100.3	72,3% / 80,00%
F1	140.9 / 253	130.4 / 217.7	79,5% / 66, 95%
Forza Horizon	98.9 / 142.2	71.9 / 125.3	43,8% / 74,26%

And so on, so on. Look at FH numbers, 5800X was able to generate 142.2 frames, so more than 12900K 125.3 (probably difference in settings but it matters for graphic card, not processor).
Some of the data are clearly outliers, so who's right? Well let's look at PCGH then because they had 20 games tested (with 12900K) and also some of those above.

Nvidia Geforce RTX 4090 im Test: So powerful, it's kinda ridiculous! [Update]

Nvidias Geforce RTX 4090 im Spieletest mit zahlreichen Benchmarks und Platzierung in der Bestenliste.

www.pcgameshardware.de

Doom - 95% (more Computerbase)
Dying Light - 67% (definitely more TPU)
Forza Horizon - 79% (more Computerbase)

So nobody's right.
And in the end in PCGH tests 4090 is 56% faster than 6950 XT, so again definitely more TPU (53%) than Computerbase (82%). Actually with bigger number of games 4090 seems 'slower'.
Computerbase seems more and more outlier than not (I'm not surprised, they chose games like Shadow Warrior 3 instead of Resident Evil Village or Elden Ring, you can not take it seriously) .

Conclusion is, there is more variance because of selected place of tests than because of processor in 4k. And that's why cards/processors should be tested with as many games as it's possible.
With more games any outliers in results have less and less impact on the end result.
17 is clearly not enough.

 

[eek2121]

Of course they can. There's plenty of market share to take from Intel and Nvidia even in a retracting market, and AMD has cost advantages. They just need to stop clinging onto the fantasy of pandemic margins and get into the mindset of growing marketshare, even if it isn't immediately that profitable to do so. The hesitation to go down this path is understandable given that it has done little for AMD in the past, but the context is very different this time. With the recession the market will reward companies that provide better value, as opposed to punishing them with a value brand stigma, and given the probable relative strength of Zen 5 and RDNA4 against the competition, AMD should be able to hold onto any gains for a reasonable amount of time. The goal for AMD should be to greatly increase margins when the economy recovers, while keeping share it captured. To do that, AMD needs competitive products when that happens.

I have Joe on ignore since he tends to spout garbage, but I saw your comment and wanted to reply.

You won't see AMD lower prices (except possibly more sales). In a normal economy, products 'selling out' can either be a marketing gimmick or a lack of supply. I bring this up because on the AMD subreddit one of the top posts was showing the availability of the 7950x vs the 13900k at Micro Center. The 7950x was widely available, while the 13900k was sold out. The user used this to claim that prices were too high. Guess what, all things being equal, if I ship a store 100 widgets and my competitor ships 10, who sells out first? This type of tactic has long been used by marketers to hype a product. I am not saying that is the case here, but let us move on. Next, we have folks showing screenshots of AMD's enthusiast level chips not topping the sales charts. What they tend to avoid talking about and/or showing is that the chips that are outperforming Zen 4 are Zen 3 chips. Why? Zen 4 is a new product requiring a new platform investment. AMD also released the top end chips first. Of course Zen 4 isn't going to top the sales charts. I am willing to bet Intel marketing is part of this messaging as well. Note that I'm not singling out Intel here, AMD does it also. How does this relate to RDNA3? AMD wants to hit 55% margins. They aren't going to lower prices. They don't have to. Their chips ARE selling. Ask me. Ask anyone in the Zen 4 builders thread. Ask folks who bought Zen 4 on reddit. Ask folks on overclock.net. The chip hasn't even been out a month and some components (like DDR5 with AMD EXPO) have had shortages. If nobody was buying Zen 4, we wouldn't be seeing sellouts of AMD EXPO RAM and motherboards. The supply chain is much healthier this round, which further paints a false narrative.

RDNA3 will hopefully be cheaper because AMD's costs are lower than NVIDIA's costs. However, they may choose to charge similar pricing to NVIDIA for healthier margins. There is nothing wrong with that.

Oh and the 4090 selling out instantly is laughable. NVIDIA always tweaks stuff until the last minute, and then ships batches to partners at the last minute. It isn't surprising that the 4090 is sold out, and it says nothing about overall sales volume. The 4090 is actually a decent card, but NVIDIA's earnings report will show you the real story. If it is selling out due to demand and not something else, earnings will blow past the 3xxx series launch. If it isn't, NVIDIA will suffer another quarter of falling revenue. My bet is on the latter. All the folks driving volume with the 3xxx series won't see a need to upgrade again, especially since consoles are targeting Zen2/RDNA2.

I don't know how well the community will accept RDNA3, or how competitive it will be, but I know that if AMD puts out a good product, it will sell. I will personally upgrade my GPU if they can make something smaller and faster than my 3090 at 4k resolutions for gaming, and also something that can run reasonable production workloads.

EDIT: Two more quick thoughts here: first, at this point in time you cannot compare DDR4 pricing to DDR5 pricing anymore than you can compare Zen 3 to Zen 4 prices or Alder Lake to Raptor Lake. DDR4 is on clearance. The majority of future sales will be DDR5. Second: Inflation is a thing folks. It isn’t some imaginary boogeyman. Component prices have gone up. It sucks, but unfortunately that is what happens when inflation gets out of control. Prices don’t typically fall either, so inflation is permanent. Competition between Intel and AMD may help a tiny bit, but you will very likely never see a $150 X670e motherboard short of clearance.

 

[biostud]

View attachment 69739

From AMD's Portuguese site: out of stock, though.

Portugal's tax is 23%, which would up the $799 price to $982.77. However, and as seen in the pic ... it's a slight more than that ... and it's not the $ to € conversion since they are virtually the same ...

Here in Denmark with 25% sales tax, they start at 842 euro or $830, they are readily in stock and you can get a liquid cooled card for around $860.

rx 6900 xt - PriceRunner

PriceRunner

www.pricerunner.dk

 

[H T C]

Here in Denmark with 25% sales tax, they start at 842 euro or $830, they are readily in stock and you can get a liquid cooled card for around $860.

rx 6900 xt - PriceRunner

PriceRunner

www.pricerunner.dk

So ... the exact same price ...

It seems Portuguese AMD's site hasn't updated their price yet, judging from the fact that there are cheaper 6900X available:



Never thought i'd see the day when an ASUS card is the cheapest ... they are usually 100€+ more expensive than the rest ...

 

[GodisanAtheist]

I think the the 4090 was held back by the 5800X they used @TPU which is why the do not have so much improvement.

- It's wild a year ago I was considering picking up a relatively cheap 10700k and was directed to all the data showing AMD demolishing that chip with the 5800x.

Here we are now a year later and the 5800x is holding back next gen GPUs.

 

[Kepler_L2]

https://twitter.com/x/status/1584414085784879105

Three N31 SKUs, two XTX models (same 96 CUs on X3D and non-X3D)

 

[biostud]

So ... the exact same price ...

It seems Portuguese AMD's site hasn't updated their price yet, judging from the fact that there are cheaper 6900X available:

View attachment 69764

Never thought i'd see the day when an ASUS card is the cheapest ... they are usually 100€+ more expensive than the rest ...

Except the price in Denmark is 842€ included sales tax. The cards are generally 150-200€ more expensive in Portugal than Denmark, weird.

 

[TESKATLIPOKA]

No. Simply no. 4090 was not held back by anything. It's a 4K. Any modern processor is enough.
If processor is slower it's slower for both cards. Not for one.

You can even compare data for the same games on both sides. Let's look at them.

Game	TPU benchmark - 6950 XT/4090 (5800X)	Computerbase benchmark - 6950 XT/4090 (12900K)	Difference for TPU / Difference for Computerbase (between cards with different processors)
CP2077	39 / 71.2	37 / 67.5	82.5% / 82.4 % (perfect)
Deathloop	72.5 / 143.1	60.6 / 121.6	97,3% / 100,6% (again almost perfect)
Doom	168.5 / 305	163.2 / 317.2	81,0% / 94,3%
Dying Light 2	58.1 / 100.1	55.7 / 100.3	72,3% / 80,00%
F1	140.9 / 253	130.4 / 217.7	79,5% / 66, 95%
Forza Horizon	98.9 / 142.2	71.9 / 125.3	43,8% / 74,26%

And so on, so on. Look at FH numbers, 5800X was able to generate 142.2 frames, so more than 12900K 125.3 (probably difference in settings but it matters for graphic card, not processor).
Some of the data are clearly outliers, so who's right? Well let's look at PCGH then because they had 20 games tested (with 12900K) and also some of those above.

Nvidia Geforce RTX 4090 im Test: So powerful, it's kinda ridiculous! [Update]

Nvidias Geforce RTX 4090 im Spieletest mit zahlreichen Benchmarks und Platzierung in der Bestenliste.

www.pcgameshardware.de

Doom - 95% (more Computerbase)
Dying Light - 67% (definitely more TPU)
Forza Horizon - 79% (more Computerbase)

So nobody's right.
And in the end in PCGH tests 4090 is 56% faster than 6950 XT, so again definitely more TPU (53%) than Computerbase (82%). Actually with bigger number of games 4090 seems 'slower'.
Computerbase seems more and more outlier than not (I'm not surprised, they chose games like Shadow Warrior 3 instead of Resident Evil Village or Elden Ring, you can not take it seriously) .

Conclusion is, there is more variance because of selected place of tests than because of processor in 4k. And that's why cards/processors should be tested with as many games as it's possible.
With more games any outliers in results have less and less impact on the end result.
17 is clearly not enough.

That TPU chart(Link) clearly shown that RTX 4090 was held back even at 4K in some games.

I am looking at that comparison table of yours, and I must highlight some things.
Only 2 games had comparable difference between reviews and those are CP2077 and Deathloop, but those games are not limited by 5800x.
Actually, from those 6 games only Forza Horizon was limited. Just check out my comparison to see what is the difference in performance by using 12900k in those games. It's from the long TPU chart I linked, and you quoted.

Resolution: 4K	CyberPunk 2077	Deathloop	Doom Eternal	Dying Light 2	F1 22	Forza Horizon 5
5800x -> 12900K	+1.8%	+0.3%	+1.4%	+0.3%	-0.6%	+10.6%

Your claim that If the CPU is slower, then It's slower for both GPUs is also not true. If one GPU is ~1.8x faster than the other, then It's not possible for It to be limited by the CPU to the same extent as the weaker GPU is.
Here is the proof, although at different resolutions:

RTX 4090 lost only 5 FPS or 4% by increasing the resolution from 1080p -> 1440p.
RX 6900XT lost ~53 FPS or 38% by increasing the resolution from 1080p -> 1440p.
RTX 3090ti lost ~38 FPS or 27.5% by increasing the resolution from 1080p -> 1440p.
Why did RTX 4090 lose so little FPS compared to other cards? The only explanation is because It was held back by the CPU a lot more than the weaker GPUs.

Just because you don't like ComputerBase using Shadow Warrior 3 doesn't mean we shouldn't take It seriously.
I agree that the higher number of games tested the better, but then TPU would be the best out of the 3 reviews, because they test 25 games, but as shown they used a weak CPU.

The best thing would be If they reviewed the upcoming N31 with the fastest gaming CPU and that one should be 13900k.

 

[Leeea]

The only explanation is because It was held back by the CPU a lot more than the weaker GPUs.

It is also possible the software was simply not coded to run that quickly. Wait timers, sync locks, frame caps, etc.

It is possible it could not transfer the frame data to the GPU quickly enough.

A GPU bottleneck not related to the CPU, like the scheduler.

A GPU bottleneck not related to the CPU, like the memory controller.


In short, it could be the CPU, but their are many other possibilities. Some more likely then others.


Example:
In some ways, Intel ARC shows similar scaling, and nobody is suggesting that is CPU limited. Many people blame the GPU's memory controller.

 

[TESKATLIPOKA]

It is also possible the software was simply not coded to run that quickly. Wait timers, sync locks, frame caps, etc.

It is possible it could not transfer the frame data to the GPU quickly enough.

A GPU bottleneck not related to the CPU, like the scheduler.

A GPU bottleneck not related to the CPU, like the memory controller.


In short, it could be the CPU, but their are many other possibilities. Some more likely then others.


Example:
In some ways, Intel ARC shows similar scaling, and nobody is suggesting that is CPU limited. Many people blame the GPU's memory controller.

That is certainly not the case with CyberPunk 2077.
By using 12900K instead of 5800x the RTX 4090 saw 31.1% increase in FPS in CP2077 at 1080p. TPU
That would mean this:

 

[Timorous]

That is certainly not the case with CyberPunk 2077.
By using 12900K instead of 5800x the RTX 4090 saw 31.1% increase in FPS in CP2077 at 1080p. TPU
That would mean this:
View attachment 69766

You are pointing at a single game though.

Given this is a speculation thread the data we have on hand is either vague (>50% perf/watt) or rumour so all we can do is try and use some napkin maths + historic definitions to get a ballpark for how the new products will perform 'on average' and what kind of TBP we think that can be done in. Given the accuracy / precision of the data we have trying to delve into the nitty gritty of a singular gaming example or trying to get error bars smaller than +/- 10% is just not going to happen. It should be assumed any estimates given in here have pretty huge error bars and are heavily caveated for this very reason.

 

[Aapje]

@eek2121

Allegedly, the Ryzen 7950X and 7900X are actually selling relatively well, but sales of 7700X and 7600X is very low. This indicates that those products are overpriced, or just as likely, the platform costs for those chips is too high (and the higher component costs of DDR5 and AM5 mobo's are relatively more significant for those builds).

I actually consider both 13th gen and Ryzen 7000 to be relatively poor generations, right now. 13th gen intel is basically an overclock, combined with more generous SKUs (more e-cores per tier). The main saving grace for 13th gen Intel is the latter and no increase in platform costs. I see neither as a particularly good generation to upgrade to for gamers right now if you have a decent system and want to upgrade when you get a really good benefit from it.

RDNA3 will hopefully be cheaper because AMD's costs are lower than NVIDIA's costs. However, they may choose to charge similar pricing to NVIDIA for healthier margins.

There is no way they can charge the same for similar performance. Radeon cards simply don't have the same inherent value as a platform. My expectation is lower prices than Nvidia, but still prices that are too high compared to their actual value. Then prices will fall to the level where AMD can find enough buyers.

The main question is how many cards AMD wants to sell, which mainly depends on how many chips they need for other products. Once they start having excess production volume, it makes sense to lower prices and/or introduce lower end SKUs.

EDIT: Two more quick thoughts here: first, at this point in time you cannot compare DDR4 pricing to DDR5 pricing anymore than you can compare Zen 3 to Zen 4 prices or Alder Lake to Raptor Lake. DDR4 is on clearance. The majority of future sales will be DDR5.

That's mostly irrelevant for someone who is contemplating buying now and not in the future. The benefit for current-gen DDR5 is small, much smaller than the price increase. Clearance prices are good, premium prices for products that will be superseded anyway are bad.