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

 

[DisEnchantment]

Phoenix doesn't have the bug?

Phoenix does not have the SGPR bug but it has the export conflict bug.
Seems all current RDNA3 chips have this export conflict bug. I understood it can impact performance in no insignificant amount. By how much? we will never know.

There definitely something wrong with clocks, just compare RDNA3 vs Zen4

Obviously Zen 4 is on a performance optimized library.
Zen 4 density is 92 MTr/mm2. N31 GCD is 132.4 MTr/mm2.
For comparison, RDNA2 and Zen 3/Vermeer have performance optimized libraries, both are at ~51 MTr/mm2.

I am doubtful if there is really an issue with the usage of high density libraries. It is a conscious decision which means they know what they are going to get if they use HD libs.

RDNA3 performance is fine, relative to Lovelace.
They just don't have enough silicon on N31 to compete at the very top. It is a conscious choice.
And if they really put "Architected to exceed 3 GHz" then it has to be so, btw this is also the same thing which the press have been briefed.
You can see Jarred Walton slipping on this NDA'd info in someone's livestream.




BTW, it is not like they are making the patches right now and discovered oopsies, lots of patches were from last year if you check the initial commits which were not squashed.
Drivers development usually starts long before the chip is even taped out. They directly push the RTL code to the emulator and start writing drivers. The physical implementation goes in parallel.

Some HW bugs are just not deemed significant enough to warrant an architecture rework.
Minor physical implementation bugs sometimes can be fixed with metal layers. Many new blocks have unused silicon around critical areas to change logic in case of bugs. The metal layers can be adapted to form a corrective logic with the additional unused transistors.
Then there are chicken bits, which gets fused off in case things don't work the way they are intended. GPUs have tons of chicken bits.

 

[Joe NYC]

N31 can't compete with Ada102. Saving chips capable of higher clocks is pretty pointless for AMD.
I also don't believe AMD will release N31 with more cache.

Navi 32 might actually be the sweet spot of the architecture.

Drops 2 MCD, their power overhead and the overhead of their connected memory. These were not exactly necessary for level of performance and market segment. That's 33% power saving from the memory subsystem

2/3 of the die means similar 33% drop on GCD side as well

That would leave the power consumption at the same clock to be 2/3 of 355 = 236 Watts.

Now, let's say AMD uses the same power budget for Navi 32, of up to 355 Watts. From 2.5 GHz starting point, and 236 Watt power base power consumption base. Up 50% in power from that bases, applied to GCD alone. Can it hit 3 GHz range?

Depends on where the chip is in scaling, but if it did, it would offer ~83% performance of 7900 XTX with 66% BOM of 7900 XTX. Or around $600 when translated to MSRP.

83% performance of 7900 XTX would be about the same or above 4080 16 GB, 320 Watts in raster, which lists at $1,200

 

[Glo.]

Phoenix does not have the SGPR bug but it has the export conflict bug.
Seems all current RDNA3 chips have this export conflict bug. I understood it can impact performance in no insignificant amount. By how much? we will never know.

Yes, I heard about this bug, thats why I am surprised that PHX doesn't have it.

It actually may change a few things about Phoenix in relation to performance potential...

I am GENUINELY surprised, DisEnchantment.

 

[Panino Manino]

Linus is really on a good mood about AMD this week:

 

[Glo.]

N31 can't compete with Ada102. Saving chips capable of higher clocks is pretty pointless for AMD.
I also don't believe AMD will release N31 with more cache.

Mate, you are the one of best analytics on this forum.

How can you say that N31 can't compete with AD102, if N31, reference will be within 10% of raster performance of 4090 in 4K?

 

[tajoh111]

I think what this hype train was built on tribalism and the fact that people do not want to think. Rather they would rather just feel good, get on the hype train and if all things fail, cope by agreeing with like minded people and not activate their brains again.

Just think about this.

Did people really think AMD figured out a way for the Navi 33 which is a 206mm 6nm die to perform similarly as a 520mm 7nm die?

This would require 4096 shader navi 3 shaders to have an IPC greater than the 6900xt's 5120 shaders while using less than 40% of the die area using a node process with a similar transistor density.

This is literally impossible unless Navi 21 was super botched to begin with. But this was actually a very common thought on this forum. Navi 33 was going to match a 6900xt on what is essentially the same node.

Most people brains should have moved a trickle and realized this is impossible but the opposite happen. Rather than think, they just joined with similar beliefs without thinking.

This is how we got to Navi 31 was going to be 2.5x to 3x faster than navi 21 even though AMD was promising somewhere around 50% performance per watt increase. How did people misconstrue this when AMD is giving them this information is simply spellbounding and shows the power of tribalism. Don't use logic just agree with the rest of the tribe.

It should have been obvious something had to happen to IPC and shaders when the shaders were shrinking so much from the rumors and that something similar with what happened with Amphere was happening to Navi 3.

But again the brain drain and death of logic made people believe the impossible.

I.E

Shaders half the size somehow had higher IPC and/or vastly greater clocks which are typically contradictory in nature. Higher IPC/clocks usually increase shader size through increase cache/pipeline.

People that opposed my view on clocks and performance did not use logic to oppose it. Rather a simple I don't believe it or what you wrote is wrong with an actual justification which demonstrates the lack of thinking.

 

[jpiniero]

How can you say that N31 can't compete with AD102, if N31, reference will be within 10% of raster performance of 4090 in 4K?

Because it won't be. I think AMD would have at least tried higher pricing if it was.

 

[Geddagod]

N31 GCD is 132.4 MTr/mm2.

Where did you get the N31 GCD density from? IIRC, AMD gave us the total number of transistors on the chip, but didn't separate it into MCD/GCD transistor counts.

Zen 4 density is 92 MTr/mm2. N31 GCD is 132.4 MTr/mm2

I don't think you can compare directly like that, because wouldn't the zen 4 CCD density be lower regardless because it has a higher percentage of cache/SRAM in the die compared to the GCD of N31 (since the cache got moved off to the N6 MCDs)?

And if they really put "Architected to exceed 3 GHz" then it has to be so, btw this is also the same thing which the press have been briefed.

Couldn't this apply to the mid range of RDNA 3 where it would be easier to hit those clocks, esp without the power cost of the MCDs

 

[Geddagod]

Mate, you are the one of best analytics on this forum.

How can you say that N31 can't compete with AD102, if N31, reference will be within 10% of raster performance of 4090 in 4K?

AMDs own benchmarks say that. I wouldn't be surprised if the real raster diff is ~15% less or even ~20%.
Also the 4090 is really just a 4080 ti in disguise IIRC lol. It has a decent section of it's die disabled.
Sure N31 could also have 3d-stacked cache, but didn't angstronomics also add that it didn't add that much perf relative to the cost regardless?
But even disregarding that, even if top N31 beats the 4090ti/Titan by 10% in raster, how many people do you think, who are willing to spend 2000 dollars on a GPU, are not going to mind about the absolute demolition in RT?

 

[Joe NYC]

I think what this hype train was built on tribalism and the fact that people do not want to think. Rather they would rather just feel good, get on the hype train and if all things fail, cope by agreeing with like minded people and not activate their brains again.

Last GPU die AMD released prior to Navi 31 was Navi 24, on N6, and it had a base clock of 2.6 GHz and boost clock of 2.8 GHz.

Was it unrealistic or tribal to expect same or higher, after AMD was able to get a good clock speed uplifts in Zen 4 with N5?

 

[Geddagod]

AMD wouldn't have called it 7900 if they had any ability of "fixing" the issue that exists quickly. They'd release this as a 7800. I really dislike the whole smugness of some of the so-called leakers, but I assume that something really went wrong somewhere though. The clocks are just too low.

I also don't think that AMD has anything bigger. This is what AMD has. It's not terrible, but it's not really great either. Better to beat the 4080 handily in what they can with decent perf/watt and better pricing than lose to a 4090 anyway with terrible power usage.

AMD themselves said they were competing with the 4080, not the 4090.

 

[Rekluse]

AMD

I'm pretty positive RDNA1 had a hardware bug that limited clocks as well. It's a shame all the engineers not getting to see their works potential, hope they get things right next round.

Thanks for the explanation Kepler

 

[jpiniero]

If they caught it soon enough to fix it in N32/N33 or at least delay the production of those cards until they could fix it or get replacement masks, it should mean really good midrange products next year.

They might cancel the desktop parts over it but I imagine they would still continue with production. It's likely not much of an issue with mobile.

 

[Glo.]

But even disregarding that, even if top N31 beats the 4090ti/Titan by 10% in raster, how many people do you think, who are willing to spend 2000 dollars on a GPU, are not going to mind about the absolute demolition in RT?

Who said that N31 will beat 4090 by 10%?

I said that it will be within 10% of 4090 performance. Sometimes on par, sometimes 10% slower.

 

[Karnak]

Because it won't be. I think AMD would have at least tried higher pricing if it was.

The thing is the 7900XTX probably will be like ~15-20% faster in raster than the 4080 16GiB which means it'll probably be closer to the 4090. Given that AMDs pricing seems odd I agree...

... but then on the other hand when it comes to RT it'll be way slower than the 4080. Often even slower than the 3090Ti since there are a lot of games out there where the 3090Ti is already 60-90% ahead of an OC 6900XT / 6950XT. Sometimes in really heavy rt scenarios it's even a >=2x difference. And that's the real problem why AMD can't charge more for their cards.

I mean tbh I don't care if RT performance on N33 is not that great. That's mid-range stuff at best. But for one grand you don't want to have a 50:50 GPU with one really good part and one which 'sucks'. If I spend a thousand dollars on a GPU I want a GPU which is good (not the best) at everything. Enough memory, good raster and good RT performance.

That's only my opinion of course but looking at what AMD want's you to pay for their GPUs I guess theyre thinking the same - which means they think that's the market.

 

[Geddagod]

Who said that N31 will beat 4090 by 10%?

I said that it will be within 10% of 4090 performance. Sometimes on par, sometimes 10% slower.

No I meant it would be ~10-20% less.

 

[TESKATLIPOKA]

Mate, you are the one of best analytics on this forum.

How can you say that N31 can't compete with AD102, if N31, reference will be within 10% of raster performance of 4090 in 4K?

If It competed against Ada102, It wouldn't cost less than Ada103, and we are still talking about a cutdown Ada102, the full Ada102 has a headroom for ~15% more.
We are only talking about raster, you know very well RT is much worse.
BTW didn't someone from AMD say that N31 actually competes against Ada103? Not sure If It was here or on beyond3d forum, and I am too lazy to search for It.

Geddagod: thanks for finding It for me.

 

[Geddagod]

If It competed against Ada102, It wouldn't cost less than Ada103, and we are still talking about a cutdown Ada102, the full Ada102 has a headroom for ~15% more.
BTW we are only talking about raster, you know very well RT is much worse.
Didn't someone from AMD say that N31 actually competes against Ada103. Not sure If It was here or on beyond3d forum and I am too lazy to search for It.

"
[Radeon RX 7900 XTX] is designed to go against 4080 and we don’t have benchmarks numbers on 4080. That’s the primary reason why you didnt see any NVIDIA compares. […] $999 card is not a 4090 competitor, which costs 60% more, this is a 4080 competitor.
— Frank Azor to PCWorld
"

 

[gdansk]

Not expecting anyone is planning to use RT on an APU anyways lol

Exactly. So being behind in RT won't matter.

 

[Ranulf]

Compared to the 6950 XT, it's a big step forward. More performance for less power. I don't know why people are so spoiled.

High prices usually lead to high expectations. Especially when we used to get far better price/perf. All with the PC component that has the highest depreciation value.

 

[linkgoron]

"
[Radeon RX 7900 XTX] is designed to go against 4080 and we don’t have benchmarks numbers on 4080. That’s the primary reason why you didnt see any NVIDIA compares. […] $999 card is not a 4090 competitor, which costs 60% more, this is a 4080 competitor.
— Frank Azor to PCWorld
"

What can they say? It's not close to a 4090, which itself is just a cut AD102. Given the performance, this is the only realistic solution. Companies don't usually say "yeah, we messed up this was supposed to have 25% better clocks and destroy the 4090".

 

[Yosar]

Something did: crypto crashed.

AMD this time was not so strong on crypto market. Even miners preferred nVidia. GCN was much better for crypto than RDNA. But yeah AMD cards also benefited from crypto boom but nowhere near as nVidia. I doubt designing chips they were taking it into consideration prioritizing again high end cards (and that's not a fact even).

 

[H T C]

AMD this time was not so strong on crypto market.

RDNA 2 yes, RDNA 3 ... dunno ...

 

[GodisanAtheist]

What can they say? It's not close to a 4090, which itself is just a cut AD102. Given the performance, this is the only realistic solution. Companies don't usually say "yeah, we messed up this was supposed to have 25% better clocks and destroy the 4090".

- Shoulda reworked the stack naming and had N31 launch as the 7800XTX instead. Would nip a lot of perception issues in the bud right there. As it stands, AMD's 9 series part is competing with NV's 8 series part, which again just reinforces the "value brand" perception of AMD.

They were smart to launch N10 as the 5700XT when it rightly competed with NV's 7 series Turing cards.

 

[Yosar]

Not sure why you are disappointed, RDNA 3 seems like a good scalable architecture. Suitable for everything.
The fact that they did not, lets say, use 8 SE on N31 does not mean the architecture is weak. The other guy totally ballooned his die size relatively.

I'm not disappointed in RDNA3 actually, but yeah due to 'leaks' I expected more. But it's not like AMD promised me something and they didn't keep a promise. Far from it.
That 'leaks' had influence on my expectations is only my work, not AMD.
Like I said I was never customer for 1200 card even if it would beat 4090 by landslide. So 7900 XTX seems very appealing to me.

They probably lost the perf/watt gains to MCD interconnects. If you cut out 30-50W due to IF energy usage from 7900XTX the perf/W numbers will be much better.
Now on monolithic it is going to be totally different.

But RDNA 2 CU vs RDNA 3 CU, should be around 1.25x perf/clock gain.
1.25x perf/CU * 1.2x CUs * 1.15x frequency = 1.725x, round about the advertised gains by AMD (You can shave off some percentage gains due to CU scaling issues)

So they made some significant gains all around with a significantly less footprint. This bodes well for the integration of the architecture in all form factors.

From VCZ slides, seems @Kepler_L2 is right about the L0 and L1 sizes getting doubled and the frequency. We have to wait for the partner cards. He will be vindicated after all.

That's why I said they laid powerful groundwork with re-architectured CU and chiplets.
I agree it's very promising what they did. And they will benefit from this probably bigger and quicker than some people think now.
I'm far from doom and gloom. And people calling it failure simply don't know what they talk about.