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

 

[TESKATLIPOKA]

Yes AMD relies on media to know the tech specs of the competition lol

It's funny how he believes everything Kepler is saying.
Kepler lost all credibility with his RDNA3 hype. End of story.

 

[Unreal123]

Yes AMD relies on media to know the tech specs of the competition lol

No. These are very big companies and they watchdogs in each companies. These companies does not run assumptions or etc. They run factual data, and research.

Kepler said in June that those rumors are false and AMD is wrong foot to underestimated rtx 4090.

Kepler said this in Jume 2022 on twitter.

 

[Unreal123]

It's funny how he believes everything Kepler is saying.
Kepler lost any credibility with his RDNA3 hype. End of story.

He said on June 2022 that Rdna 3 is being over hyped on twitter.

 

[TESKATLIPOKA]

He said on June 2022 that Rdna 3 is being over hyped on twitter.

And he was one of those who hyped It untill launch. High clocks and all, you can still find his predictions about ADA vs RDNA3 and other stuff he said. But you can believe him all you want, I don't care.

 

[AtenRa]

He said on June 2022 that Rdna 3 is being over hyped on twitter.

The one that overhypes is the one that says it is overhyped lol

Anyway, I believe AMD had their reasons they only opted to compete up to $999 with a middle die size chip instead of competing against 600mm2 AD102 die.

 

[biostud]

If it was supposed to be as fast as the 4090, it would also have required more power.

 

[NTMBK]

Right now, nVidia has A TON of 3000 series inventory to get rid of due to crypto crash, which is why they priced 4090 as they did. Don't kid yourselves: nVidia's plan was to price the upcoming 4090 Ti @ LEAST in the $2000 range.

By pricing the 7900 cards aggressively like this, they make it A LOT MORE difficult for nVidia to get rid of those cards, which is why stuff like this is already happening.

AMD is in a position to make it VERY HARD for nVidia to get rid of that inventory: if they make their 6000 series cards EVEN CHEAPER than they are now and NOT BY A LITTLE, they MAY be able to clear out THEIR OWN inventory of 6000 series cards which they too have (though nowhere near to the extent of nVidia's).

Should AMD opt to go this route, it won't be cheap for them and would MOST CERTAINLY greatly reduce their profits BUT, it would also substantially increase their market share in the process WHILE forcing nVidia to resort to SERIOUS price cuts themselves in order to get rid of their current inventories, but this would cost A LOT MORE to nVidia than it would to AMD (my guess would be around 10 times more).

An approach like this would be MUCH MORE EFFECTIVE if the 7900 series were to match (or more) the 4090 @ least in raster but that appears not to be the case and, as expected, they still trail in RT.

What's with all the RANDOM SHOUTING in this post?

 

[Unreal123]

Can anybody guide why both Amd and Nvidia are going for fp32 x 2. What is the benefit of this?

 

[Timorous]

If it was supposed to be as fast as the 4090, it would also have required more power.

Not if AMD expected a better v/f curve which just didn't pan out.

 

[DisEnchantment]

It's funny how he believes everything Kepler is saying.
Kepler lost all credibility with his RDNA3 hype. End of story.

This is why I prefer manuals, patches and patents. I don't really trust leaks.

From patents we knew HW traversal and sorting will be coming, and they did.
From patches we knew 1.5x VGPRs will be coming, and they did.
From patches we knew Dual issue will be coming, and they did.
So on and so forth.

However, RDNA3 as an architecture is great.
N31 as a high end product unlikely because they spent less resources on it. What are they thinking? They did not anticipate NV to push the limits?

But the RDNA3 design is a well rounded design for Mobile.
Strong on compute and raster with a very tiny CU design.
Media capabilities are great.
Maybe that's where their key requirements came from.

Look at this chart prepared from updated Locuza annotation. CU is 40% smaller. CP is 60% smaller.


They are using iPhone SoC level Transistor density on the GCD. N21 has ~52 MTr/mm2. This value from AMD means N31 GCD has 135+ MTr/mm2


This means for N31, this is as good as it gets

• 3.0+ GHz clocks is zero possibility without exotic cooling/overclock
• Heat density will be insane

To me it seems they design this architecture for something else.

But what it means for Mobile, is that
For Phoenix, CPU Cores + Multimedia + GPU + PCIE + DDR5 MC etc. can be done within 140-150 mm2 compared to 180+ mm2 on Ryzen 5000 Mobile and 210+ mm2 on Ryzen 6000 Mobile.

Bean counters likely decided the fate of N31

 

[DAPUNISHER]

Hmm, that may be it, that the reference cards are BIOS locked to spec for maximum power draw. This should change on models with 3 8pin connectors.

It will according to what Jarred Walton says AMD told the media at the event. And AMD did say themselves it is designed to scale as high as 3GHz.

It would be a fairly savvy play by AMD, if some of the cards do hit near 3GHz. Price and power draw were the Belle of the Ball. Despite not taking the performance crown in raster, the reception by the media has been largely positive. Partner cards with massive coolers, 3 power connectors, drawing over 400w, and hitting close to 3GHz, with another $100-$200 price boost, would have dampened the enthusiasm. If they can pull off 10% performance over stock, for what a reference 4080 costs? It is going to look good in reviews.

 

[exquisitechar]

This is why I prefer manuals, patches and patents. I don't really trust leaks.

From patents we knew HW traversal and sorting will be coming, and they did.
From patches we knew 1.5x VGPRs will be coming, and they did.
From patches we knew Dual issue will be coming, and they did.
So on and so forth.

However, RDNA3 as an architecture is great.
N31 as a high end product unlikely because they spent less resources on it. What are they thinking? They did not anticipate NV to push the limits?

But the RDNA3 design is a well rounded design for Mobile.
Strong on compute and raster with a very tiny CU design.
Media capabilities are great.
Maybe that's where their key requirements came from.

Look at this chart prepared from updated Locuza annotation. CU is 40% smaller. CP is 60% smaller.
View attachment 70470

They are using iPhone SoC level Transistor density on the GCD. N21 has ~52 MTr/mm2. This value from AMD means N31 GCD has 135+ MTr/mm2
View attachment 70471

This means for N31, this is as good as it gets

• 3.0+ GHz clocks is zero possibility without exotic cooling/overclock
• Heat density will be insane

To me it seems they design this architecture for something else.

But what it means for Mobile, is that
For Phoenix, CPU Cores + Multimedia + GPU + PCIE + DDR5 MC etc. can be done within 140-150 mm2 compared to 180+ mm2 on Ryzen 5000 Mobile and 210+ mm2 on Ryzen 6000 Mobile.

Bean counters likely decided the fate of N31

Excellent analysis. I had similar thoughts, that RDNA 3 will be significantly better in iGPUs. When you lay it out like that, it all becomes clear. N31 is an afterthought and its issues mostly stem from the fact that RDNA 3 is very focused on maximizing raster performance in as little area as possible (such as in iGPUs), as well as excellent media capabilities, with everything else being secondary.

Considering how the current AMD operates, this comes off as no surprise at all. Leakers just pulled the wool over many people’s eyes, myself included. Next time around, I’m certainly ignoring everything but patches and Angstronomics, if they keep making articles of that nature.

 

[Karnak]

Well something with RDNA3 doesn't seem right going with the benchmarks from AMD and looking at this one.

"architected to exceed 3GHz" -> the 7900XTX only sits around 2.3-2.5GHz
"up to 1.8x RT" -> only 1.6x according to AMDs numbers in Doom Eternal (so the 7900XTX is prob running with lower clock speeds than 2.5GHz?)

But now we probably know were the "more than 3GHz" rumors came from.

 

[leoneazzurro]

"architected to exceed 3GHz" -> the 7900XTX only sits around 2.3-2.5GHz

Well, these are the rated clocks but we know from RDNA2 that these clocks can be exceeded. For sure they are not 3+GHz, but I will not be surprised to see in the reviews that actial clocks can reach 2,6+GHz in boost mode. As said previously, it's probably the power consumption that keeps it to go higher more than an intrinsic architectural limit. I wonder if this is also a result of going for maximum area reduction (high density libraries instead of high performance?).

"up to 1.8x RT" -> only 1.6x according to AMDs numbers in Doom Eternal (so the 7900XTX is prob running with lower clock speeds than 2.5GHz?)

1.8x should be at same clocks. But RT is only a part of the graphics pipeline. You'll se maybe that improvement in titles like Minecraft RT, but in other games where there are other things to do you will see the improvements go more along the raster performance mprovement.

 

[Joe NYC]

Zen 4 3D is definitely going to be a great gaming CPU, but so is Raptor Lake and it's already available. Also, one thing I like about Raptor Lake compared to Zen 4 is that it can utilize ultra high speed DDR5. No reviewer has yet to benchmark Raptor Lake with DDR5 7200 and above memory speeds using an RTX 4090. Should be interesting!

Another way you can look at it is that with Zen 4 V-Cache, you don't have to buy expensive DDR5 (and endure random instability). DDR5 -7200 cost double of the more standard DDR5, and with 32 or 64 GB, it starts to add up

 

[TESKATLIPOKA]

This is why I prefer manuals, patches and patents. I don't really trust leaks.

From patents we knew HW traversal and sorting will be coming, and they did.
From patches we knew 1.5x VGPRs will be coming, and they did.
From patches we knew Dual issue will be coming, and they did.
So on and so forth.

However, RDNA3 as an architecture is great.
N31 as a high end product unlikely because they spent less resources on it. What are they thinking? They did not anticipate NV to push the limits?

But the RDNA3 design is a well rounded design for Mobile.
Strong on compute and raster with a very tiny CU design.
Media capabilities are great.
Maybe that's where their key requirements came from.

Look at this chart prepared from updated Locuza annotation. CU is 40% smaller. CP is 60% smaller.
View attachment 70470

They are using iPhone SoC level Transistor density on the GCD. N21 has ~52 MTr/mm2. This value from AMD means N31 GCD has 135+ MTr/mm2
View attachment 70471

This means for N31, this is as good as it gets

• 3.0+ GHz clocks is zero possibility without exotic cooling/overclock
• Heat density will be insane

To me it seems they design this architecture for something else.

But what it means for Mobile, is that
For Phoenix, CPU Cores + Multimedia + GPU + PCIE + DDR5 MC etc. can be done within 140-150 mm2 compared to 180+ mm2 on Ryzen 5000 Mobile and 210+ mm2 on Ryzen 6000 Mobile.

Bean counters likely decided the fate of N31

Nice analysis, but I don't think the GCD has 135+ MTr/mm2.
They were comparing RDNA3 vs RDNA2 WGP when they were talking about 165% increase.


If what Locuza wrote is correct, then IF+interconnect+xbar use up 1/4 of GCD.
AMD really should have made a 400mm2 GCD instead.

Honestly speaking, Phoenix won't improve performance as much as Rembrandt. Number of WGP(CU) is still the same and frequency won't see noticeable gains.

 

[H T C]

What's with all the RANDOM SHOUTING in this post?

It's the points i was TRYING to make: emphasizing THOSE to better illustrate the points i was making.

 

[PJVol]

Looking at Zen4 default PPT it's not hard to guess the price for the XT(X) 3ghz+ in terms of power.

 

[DisEnchantment]

Nice analysis, but I don't think the GCD has 135+ MTr/mm2.

Lets say I ran out of my paper napkin to consider the additional decimal points during multiplication.
It is 132.4 MTr/mm2., but I did say similar to iPhone SoC.
So I learnt from the big leakers to guard my statements with a non quantifiable value.

 

[igor_kavinski]

Has it occurred to someone here that maybe AMD is saving the best dies for their flagship to be revealed at a later date?

 

[TESKATLIPOKA]

Looking at Zen4 default PPT it's not hard to guess the price for the XT(X) 3ghz+ in terms of power.

If they could save up to 25% of power just by downclocking shaders 2.5->2.3GHz, then I have to question If It's possible to clock N31 anywhere close to 3GHz without attacking 600W.

 

[H T C]

Has it occurred to someone here that maybe AMD is saving the best dies for their flagship to be revealed at a later date?

Isn't the XTX model a full GPU? If it is, unless they come up with a model with more cache, then the only way that's true is with a different chip.

If that's not the full chip, then that makes perfect sense.

 

[TESKATLIPOKA]

Has it occurred to someone here that maybe AMD is saving the best dies for their flagship to be revealed at a later date?

7900XTX is the full die. If that is not a flagship, then I don't know what is. Higher clock will gain maybe ~10%, but TBP will suffer a lot.

Isn't the XTX model a full GPU? If it is, unless they come up with a model with more cache, then the only way that's true is with a different chip.

If that's not the full chip, then that makes perfect sense.

More cache won't help much. The problem is the low number of WGPs and clockspeed.

 

[DisEnchantment]

Well something with RDNA3 doesn't seem right going with the benchmarks from AMD and looking at this one.

"architected to exceed 3GHz" -> the 7900XTX only sits around 2.3-2.5GHz
"up to 1.8x RT" -> only 1.6x according to AMDs numbers in Doom Eternal (so the 7900XTX is prob running with lower clock speeds than 2.5GHz?)

But now we probably know were the "more than 3GHz" rumors came from.

Has it occurred to someone here that maybe AMD is saving the best dies for their flagship to be revealed at a later date?

Noo way, you want to resurrect the leaks again?
Hype train derailed, now you want to put it back on the tracks again to be derailed once more on December? Very sadistic.

 

[igor_kavinski]

Yes, full die but maybe worse binned die? The good dies that can reach higher speeds are possibly being saved because both AMD and Nvidia are keeping their cards close to their chest? Putting out their best dies right now when the market situation is not ideal would only mean having to sell them cheap on sale later. So first put out the slightly slower versions. When sufficient inventory is sold, you show your best card and suddenly, all the people with the big wallets start selling their existing cards to get the newer ones. They know the psychology of people who want the best more than anyone else.