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

 

[PJVol]

If I understand It correctly a RX7900xtx has a max boost for shaders at 2.5GHz

May be, or may not.
The 6800XT specs stated clocks are 2015 / 2250. But after the Gfx engine is initialized, actual Fmax value is set way above those, 2464 in the overdrive table, and to 2577 in the powerplay table, with the latter prioritized.
Just curious of what is set in a N31 bios.

 

[amenx]

There is no way the performance target for the 7800xt is not the 6950xt.

It will perform like a 6900XT (+/- 10%) some way or some how.

If it can outperform the 6900 by significant margins then AMD will lock that crap down in HW (like they did with the 6800's OC potential) and save for the mid gen refresh.

That would be one of the most lack luster generational gains ever for a same tier GPU (6800xt - 7800xt). The 6900 is only about 10% faster than a 6800xt. So if a 7800xt is 10% faster than a 6900xt, it would put it roughly 20% faster than the 6800xt.

 

[GodisanAtheist]

That would be one of the most lack luster generational gains ever for a same tier GPU (6800xt - 7800xt). The 6900 is only about 10% faster than a 6800xt. So if a 7800xt is 10% faster than a 6900xt, it would put it roughly 20% faster than the 6800xt.

- There just isn't that much room between the 6950XT and the 7900XT to fit another card in there, certainly not enough for a full chip one tier down. And the 7800XT can't perform too close to the 7900XT cause there just isn't that much room until you get to the 7900XTX.

AMD's other option here is to offer 6950XT performance, but leave a huge hole in pricing and go right down to $500 MSRP given the N32 should be relatively inexpensive to produce. Market it like you're offering $1000 performance at $500 (never mind there was only ever a ~10% difference between the $650 6800XT and 6950XT in the first place).

I mean I just don't see how it plays out any other way, AMD's sort of backed themselves into a performance corner here with their heavily cut N31 7900XT die.

 

[Kaluan]

24gbps memory won't be cheap either.
This wouldn't be much help to either RDNA3 or Phoenix.
It could help If the bus width will be widened to 512 or 768bit, because realistically 384bit bus and 12 chips is the limit.

I interpreted the 22Gbps 512bit demonstration as using their new 24Gbps chips but downclocked due to thermal constraint of them being on top of each other.

TBH my mind also jumped to AMD's alleged upcoming stacked RDNA3 MCDs. And if the 64bit controllers in them can also be stacked, not just the SRAM. Or is the controller on the stacked die just dead silicon?

Not that they'd more raw bandwidth for the current crop of SKUs. But it tickles curiosity about the future. They did say they are not CURRENTLY targeting AD102. Tho I won't hold my breath yet.

Oh and I think I was the one who posted the impressive 8GB 6500 XT performance numbers, I hypothesized that the 4GB framebuffer was the real Achille's Heel of N24, not 4x lanes or the narrow bus. And I randomly came across someone reviewing Arc 380 who also benched Sapphire's 8GB 6500XT.

Or maybe they also posted it before idk 😛

- There just isn't that much room between the 6950XT and the 7900XT to fit another card in there, certainly not enough for a full chip one tier down. And the 7800XT can't perform too close to the 7900XT cause there just isn't that much room until you get to the 7900XTX.

AMD's other option here is to offer 6950XT performance, but leave a huge hole in pricing and go right down to $500 MSRP given the N32 should be relatively inexpensive to produce. Market it like you're offering $1000 performance at $500 (never mind there was only ever a ~10% difference between the $650 6800XT and 6950XT in the first place).

I mean I just don't see how it plays out any other way, AMD's sort of backed themselves into a performance corner here with their heavily cut N31 7900XT die.

~10%?
What?

6950XT is like ~21% faster than 6800XT (at 4K). Or ~15% at minimum, depending on your go-to review outlet.

 

[PJVol]

If the loss is within 10% then yes, N32 at 2.8GHz could be a few % faster.
What I want to point out is that 2.8GHz is just the shader clockspeed, the Frontend will be clocked higher, based on N31 It should be 3050 MHz.

If N32 SE clock is 2.8ghz then it should be on par with the 7900XT in 1080p and may be even in 1440p.
Just curious where N32 should fit in the rdna 3 stack if it doesn't inherit rumoured N31's "clocking issues"

6950XT is like ~21% faster than 6800XT (at 4K).

According to TPU 13.6% @4K (12% @QHD and and 10% @FHD resp.)

 

[Timorous]

If N32 SE clock is 2.8ghz then it should be on par with the 7900XT in 1080p and may be even in 1440p.
Just curious where N32 should fit in the rdna 3 stack if it doesn't inherit rumoured N31's "clocking issues"

According to TPU 13.6% @4K (12% @QHD and and 10% @FHD resp.)

AMD would love that.

Stop making 7900XT and sell a lot more 7800XTs with similar performance at a better price with better margin.

There has to be a reason why 7900XT is worse value than the XTX and the most logical to me is that AMD don't really want to sell many.

If AMD make the 7900XT overclockable then they can probably have a large enough market to sell into with their harvest dies but a small enough market that they don't need to down bin perfectly good dies that could be an XTX.

 

[beginner99]

newest rumor via reddit is there are reworked n31 cards in the works with up to 3.6 ghz clocks and 3d vcache.

 

[TESKATLIPOKA]

newest rumor via reddit is there are reworked n31 cards in the works with up to 3.6 ghz clocks and 3d vcache.

Finally! I feared I had to wait a few months before RDNA4 release to see another hype train, but RDNA3's hype train is still alive and kicking.
AMD can simply dissolve their marketing team, netizens do a better job than them.

 

[coercitiv]

Finally! I feared I had to wait a few months before RDNA4 release to see another hype train

Few months? This thread is 22 months old, and the hype was overwhelming from day 1.

 

[exquisitechar]

newest rumor via reddit is there are reworked n31 cards in the works with up to 3.6 ghz clocks and 3d vcache.

The source is RGT and I don't really believe this specific rumor at all.

Concerning the hype train, I'm not expecting anything anymore, but I wouldn't be surprised if the rumor that N31 is botched is true, either. The claims in the slides for RDNA 3 don't seem to match up with the reality that is the 7900 XTX.

 

[TESKATLIPOKA]

3.6GHz for reworked N31 looks like a total BS. That would be 44% higher clockspeed and no sane person should believe It.
It looks like RGT doesn't have anything better to write about, so they came up with this.

 

[PJVol]

Concerning the hype train, I'm not expecting anything anymore, but I wouldn't be surprised if the rumor that N31 is botched is true, either.

Yep, and even if it's true to some extent, given the current position of Zen 4 and supposedly Rdna 3, with the current PC configuration I can safely skip both upgrades.

 

[eek2121]

3.6GHz for reworked N31 looks like a total BS. That would be 44% higher clockspeed and no sane person should believe It.
It looks like RGT doesn't have anything better to write about, so they came up with this.

Not saying I believe it, especially since RGT (or any youtuber) mentioned it, but the original fmax in the vbios was 3.72ghz. The 3D VCache thing though, nah, I think it is fake.

I am surprised we've seen no leaked benchmarks as of yet.

 

[TESKATLIPOKA]

Not saying I believe it, especially since RGT (or any youtuber) mentioned it, but the original fmax in the vbios was 3.72ghz. The 3D VCache thing though, nah, I think it is fake.

I am surprised we've seen no leaked benchmarks as of yet.

If I am not mistaken, then N21 also has a lot higher Fmax in vbios than It's actual clockspeed is.

 

[Justinus]

N21 Fmax for the original 6900XT is 3GHz for all models. 6900XT "Ultimate" models with the XTXH chip received a 4.0 GHz Fmax, in response to demand from overclockers to allow them to go higher with exotic cooling. 6950XT Fmax is 5.0 GHz. On water, my original 6900XT Strix runs at 2.45GHz stock and overclocks stable to 2.65GHz (actual reported clocks, not set clock maximum) so 3.0 GHz isn't really that much higher than the AIB clocks and potential overclockability.

With regards to RDNA3, this means nothing without assuming they used the same rationale to determine Fmax. I am not so sure they did.

 

[Kaluan]

Yep, and even if it's true to some extent, given the current position of Zen 4 and supposedly Rdna 3, with the current PC configuration I can safely skip both upgrades.

What's your current config, if you don't mind me asking?

I try to avoid TPU data, something about them just rubs the wrong way.

Their GeForce reviews cover DLSS1/2 and now DLSS3 (since 4090 launch) since day 1, but somehow nearly 2 years on, they can't be bothered to have a FSR rubric for Radeon cards. It's pretty clear nVidia (and Intel in different regards) has their review methodology written for them.

Which is even more bizarre considering the usually laud FSR implementation in their game by game articles.

But I still use their GPU performance charts (4K data only for 2080Ti and up) from time to time when I'm lazy. 😂
Anyway, even their chart says 15%, dunno about other people but 13.6, 15, 16 and 21% (HU and Igor's, but I think they reviewed custom) have one thing in common, they're not "~10%" 😛


Back to the main subject for a sec, it's naive to think 7900XTX is AMD's last say with high-end RDNA3. But it's even more so ridiculous to believe these 3,6GHz "leaks"

Wanna see how easy it is to make stuff up? Scaling Navi31 clocks up by 44% means they will also scale the bandwidth up by 44%. Samsung just demonstrated GDDR6W w/ 512bit@22Gbps = 1,4TB/s (960GB/s + 44%).
7950XTX will use GDDR6W CONFIRMED! /s 🤣

 

[Mopetar]

Few months? This thread is 22 months old, and the hype was overwhelming from day 1.

After RDNA2 it was certainly understandable that expectations would be high. We started out there wondering if clock speeds would even hit higher than 1.8 GHz and it kept ramping up. AMD came within spitting distance of the best GeForce card available when in the past several generations they barely had anything beyond mid-range.

 

[PJVol]

What's your current config

5600X / 6800XT / 1440p monitor

 

[desrever]

- There just isn't that much room between the 6950XT and the 7900XT to fit another card in there, certainly not enough for a full chip one tier down. And the 7800XT can't perform too close to the 7900XT cause there just isn't that much room until you get to the 7900XTX.

The 7900XT should be 30% faster than the 6950XT by numbers given by AMD. 7800XT could easily slot between the 6950XT and 7900XT. Depending on clocks of course.

6700XT to 6800 is about ~20% if we compared last gen. If the 7900XT is 20% faster than the 7800XT, then the 7800XT will be faster than the 6950XT using AMD's released numbers.

Number depends on resolution tho. The 7800XT may not scale as well at 4k because of the number of shader engines. It will probably be better suited as a 1440p card.

 

[Saylick]

3.6GHz for reworked N31 looks like a total BS. That would be 44% higher clockspeed and no sane person should believe It.
It looks like RGT doesn't have anything better to write about, so they came up with this.

I honestly think the alleged 7950XTX looks even more suspect...

2.5 GHz to 3.3 GHz max clock (~32% increase) at the same TBP on the same node with the same architecture? I find that difficult to believe, but would be pleasantly surprised if it does happen.

 

[gdansk]

I have no special information but I can say
1. It isn't happening
2. You won't even see 3.1GHz "boost" clocks for N32

 

[Leeea]

The hype train is always good for clicks, views, and $$$.

It is all lies though. We have already seen everything that is going to be.

 

[Kaluan]

The hype train is always good for clicks, views, and $$$.

It is all lies though. We have already seen everything that is going to be.

Agree with the first part, it's how social media... scratch that, everything runs. Sadly.

Hard disagree with the second. We can see basically both linear TBP and linear TFLOP scaling to performance between 7900XT and 7900XTX. We didn't have this with RDNA2/Navi21 yet we still got a Navi21 halo SKU refresh.

And none of that even has anything to do leak theories or hopium about clock targets not being met and future tap-outs fixing it.

nVidia architecture probably pushing 500W+ TBP on a better node than AMD and AMD not exploiting that? Neah.

...not to mention we're talking about products not even released yet, but somehow you pass definitive judgement. Hilarious, really.

 

[TESKATLIPOKA]

I honestly think the alleged 7950XTX looks even more suspect...

2.5 GHz to 3.3 GHz max clock (~32% increase) at the same TBP on the same node with the same architecture? I find that difficult to believe, but would be pleasantly surprised if it does happen.

View attachment 71985

Frequency increase while keeping TBP is just ridiculous. BW increase compared to previous gen also doesn't match with Flops increase, even If you excluded dual issue.
I simply don't believe It, but If It turned out to be true, then AMD has to discontinue the original ones and price these new ones a lot higher. I don't think the early adopters would be happy If this turned out as true.
I have to say that these clocks would explain why AMD didn't make bigger GCDs, but It could also be that they didn't think Nvidia will release a 600mm2 chip.

 

[TESKATLIPOKA]

I have no special information but I can say
1. It isn't happening
2. You won't even see 3.1GHz "boost" clocks for N32

1. I also think so.
2. 6500XT ended at 2815 MHz, factory OC 6500XT clocks at 2906Mhz in Cyberpunk, Link.
7-10% higher clockspeed is not unrealistic for a new arch and better process.
Of course, I don't expect N32, which is 1.5x N22 to have 3.1GHz(~+20%) and still keep the same TBP.
With 300W TBP, I can see those clocks.