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

 

[maddie]

If there is an issue that means it clocks lower than desired at a given voltage and it is fixable the best I can see is a refresh mid next year with 7950XT and 7950XTX adding maybe 20% more performance at the top end.

It feels like AMD were gunning for just over 4090 raster perf at a more sensible price point but got close enough to not want to bother delaying knowing their refresh products will be that little bit better instead.

If the rumor of the follow on products being fixed is true, this means that the flaw was found and corrected months ago on N32 & N33. It might be a lot quicker than your estimate.

 

[amenx]

If the rumor of the follow on products being fixed is true, this means that the flaw was found and corrected months ago on N32 & N33. It might be a lot quicker than your estimate.

If it was found and corrected with a refresh soon to follow, I hope they stick to the same pricing.

 

[biostud]

Reminds a bit of GTX 480, except that AMD has chosen not to turn into toaster for a tiny improvement in performance.

 

[lixlax]

Reminds a bit of GTX 480, except that AMD has chosen not to turn into toaster for a tiny improvement in performance.

With all that talk going on of potential bug thats limiting clocks and efficiency on RDNA3, I thought when before something similar (may) have happened and I came up with Fermi.
The respin of the silicon was released ~7 months after GTX480 and it had ~10% higher clocks, lower power consumtion and more shaders enabled at the same time (GTX 580).

 

[desrever]

This is without a doubt N31 and the largest RDNA3 die that will come to market.

Why say that, the die is pretty small compare to what is possible. I doubt RDNA4 will launch any time soon (probably ~2 years from now). Launching 8000 series in 2023 with a bigger RDNA 3 die would be pretty reasonable at this point. Refresh in 2023 makes a lot of sense if there is silicon bugs as well.

 

[jpiniero]

With all that talk going on of potential bug thats limiting clocks and efficiency on RDNA3, I thought when before something similar (may) have happened and I came up with Fermi.
The respin of the silicon was released ~7 months after GTX480 and it had ~10% higher clocks, lower power consumtion and more shaders enabled at the same time (GTX 580).

To me that seems like a reasonable timeframe.

 

[Yosar]

"It (RDNA 3) is also our first gaming GPU architecture that will leverage the enhanced 5nm process and an advanced chip packaging technology. And another innovation includes a architected compute units with enhanced ray-tracing capabilities and an optimized graphics pipeline with even faster clock speeds and improved power efficiency." - David Wang @ FAD 2022

"We’ve driven the frequency up, and that is something unique to AMD. Our GPU frequencies are 2.5 GHz plus now, which is hitting levels not before achieved. It’s not that the process technology is that much faster, but we’ve systematically gone through the design, re-architected the critical paths at a low level, the things that get in the way of high frequency, and done that in a power-efficient way.

Frequency tends to have a reputation of resulting in high power. But in reality, if it’s done right, and we just re-architect the paths to reduce the levels of logic required, without adding a bunch of huge gates and extra pipe stages and such, we can get the work done faster. If you know what drives power consumption in silicon processors, it’s voltage. That’s a quadratic effect on power" - Sam Naffziger @ Venture Beat

So I don't think it's fair to say that clock speed increase was only expected because of leaks.

You're right. I must have missed it. They hinted about much better clocks. Not as loud as with processors (well even Lisa was talking about it). Especially Naffziger claims are now problematic for them. Because Wang was much more reserved ('even faster' ..well they are faster but not that much).
So the ball is in their park. Or maybe Naffzinger went here the way of Raja. If nothing will happen in next 6-7 months we must assume he went the way of Raja (many promises, not many kept).

 

[Joe NYC]

If the rumor of the follow on products being fixed is true, this means that the flaw was found and corrected months ago on N32 & N33. It might be a lot quicker than your estimate.

I am not sure if this means anything but N33 may be delayed, and maybe the first to get the fix. But this is deep in the rumor mill...

https://twitter.com/x/status/1589375356708225024

 

[gdansk]

How can anyone believe rumors about how AMD is fixing the problem when all the leakers were blindsided by AMD even having the problem.

 

[Glo.]

How can anyone believe rumors about how AMD is fixing the problem when all the leakers were blindsided by AMD even having the problem.

Because that is the ONLY next logical step, for AMD. To fix their error.

 

[gdansk]

Because that is the ONLY next logical step, for AMD. To fix their error.

Emphasis on how.
These 'leakers' have no idea how. Even the nature of the problem is nebulous.
Will it be fixed in N32, N33? Fixed in a N31 refresh? Who knows.

 

[leoneazzurro]

The important thing is if AMD found the issue, they are the ones who must fix the problem. And yes, the logical thing to do is to fix it, especially on N32 and N33, because these will be the more mainstream GPUs, and the top GPUS for the mobile.

 

[PJVol]

Because that is the ONLY next logical step, for AMD. To fix their error.

What error? Which of those who can be trusted claimed there's an error or bug or whatever outside of usual suspect's rumbling about missing the perf target?

 

[Kaluan]

With all that talk going on of potential bug thats limiting clocks and efficiency on RDNA3, I thought when before something similar (may) have happened and I came up with Fermi.
The respin of the silicon was released ~7 months after GTX480 and it had ~10% higher clocks, lower power consumtion and more shaders enabled at the same time (GTX 580).

Thought of the same analogy, GTX 480 was impressive in certain ways, but at least equally unimpressive in others. Where as it's refresh (GTX 580) was just so all-around good that it made it into a whole new generation.

The important thing is if AMD found the issue, they are the ones who must fix the problem. And yes, the logical thing to do is to fix it, especially on N32 and N33, because these will be the more mainstream GPUs, and the top GPUS for the mobile.

Yeah I for one am still holding on the hope that Navi 33 may finally the struggling PC gamer's new Polaris 10/20, at some point next year.

 

[beginner99]

Because that is the ONLY next logical step, for AMD. To fix their error.

Not really. Their next step (which surely already happened) is to determine if it is financial viable to fix it at all. Correct me if I'm wrong but a respin will require new masks which are in millions likley double digit millions. Then it will take at least half a year. So it depends on how far RDNA 4 is. the new respinned product will require a certain time on market to make it viable.

 

[Kaluan]

Not really. Their next step (which surely already happened) is to determine if it is financial viable to fix it at all. Correct me if I'm wrong but a respin will require new masks which are in millions likley double digit millions. Then it will take at least half a year. So it depends on how far RDNA 4 is. the new respinned product will require a certain time on market to make it viable.

In that scenario RDNA4 wouldn't be what they'd push ahead first IMO, GFX1150/"RDNA3+" would be, IF it's not just for iGPUs.

"at least half a year"... starting from when tho? We have no idea when the alleged flaws were discovered and even less when the decision to respin was taken (if at all), could be 4 months ago, could be 1 month ago. Either way it has to be after the original Navi3x/31 silicon already started mass production.

 

[desrever]

So it depends on how far RDNA 4 is.

Considering AMD hasn't even mentioned RDNA4 yet, its probably not coming in the next 12 months. Their CPU roadmap has Zen5 in 2024, RDNA4 is probably after that. That gives RDNA3 a lot of time.

 

[dr1337]

I really feel like assuming there is a bug in navi-31 is quite silly when AMD literally hit their exact performance/watt targets. Sticking to 2x8pin on reference cards seems like a very deliberate choice to stay conservative on power and cooling requirements. The architecture might be designed for 3ghz+ but for a flagship with 2x as many transistors as its predecessor with a more novel packaging technique, staying at the same power consumption is probably hard to do without cutting clocks back. Partner cards should shed some light on this 3ghz business but really I want to see n32. A 7800xtx at the same power limit but minus two SEs and two MCDs would have tons more overhead for clocks.

 

[amenx]

Well if a certain report/rumor about an AIB only able to squeeze 3% more performance out of it, then it does sound like an issue. Not that clocks cant go higher, but that performance doesnt scale with it. Can only wait for reviews to know.

 

[Mopetar]

It's probably just cheaper for AMD to keep the current Navi 31 design and shift production to other chips with fixes in place. They don't sell enough high-end parts to justify a new spin. Thankfully the modular design means they're not using as much silicon as they otherwise would be and the memory modules can be used with Navi 32.

Has anyone done any kind of analysis as to the expected performance of Navi 32 assuming that the performance issues are resolved and it is able to reach the 3 GHz design goal?

 

[RnR_au]

Considering AMD hasn't even mentioned RDNA4 yet, its probably not coming in the next 12 months. Their CPU roadmap has Zen5 in 2024, RDNA4 is probably after that. That gives RDNA3 a lot of time.

RDNA4 is in 2024.

 

[jamesdsimone]

Forgive the wiki link but they just posted specs on the RX 7900 XT and XTX. Can't speak to the accuracy.

https://en.wikipedia.org/wiki/Radeon_RX_7000_series

 

[Kaluan]

Has anyone done any kind of analysis as to the expected performance of Navi 32 assuming that the performance issues are resolved and it is able to reach the 3 GHz design goal?

Well that would put it at ~89% of the compute of 7900XT. It all depends how well RDNA3 CUs scale from here. My guess is it would be somewhat close in performance (like 6800XT to 6900XT), but enough difference to tier it down the stack in the consumer's eye.

If it's really 60CUs, not 64, it will offer a great opportunity for RDNA2 v RDNA3 "apples to apples"/PPC comparisons. Including memory subsystem/IC.

 

[Saylick]

Forgive the wiki link but they just posted specs on the RX 7900 XT and XTX. Can't speak to the accuracy.

https://en.wikipedia.org/wiki/Radeon_RX_7000_series

Seems legit... 🤔

 

[zinfamous]

Considering AMD hasn't even mentioned RDNA4 yet, its probably not coming in the next 12 months. Their CPU roadmap has Zen5 in 2024, RDNA4 is probably after that. That gives RDNA3 a lot of time.

If the bug fix is really quite significant, they could release it as some sort of anniversary special edition card (some made-up date they pull out of their butts to make an anniversary, I guess.) as a hold-over between RDNA 4. ...it will be 7900XTX--RT version. or some nonsense.

I was thinking in a hypothetical the other day. What if this bug didn't happen, and their raster was @ maybe 3-5% above 4090. That doesn't seem like a lot of course, and really not that much more than the generous ~10-15% they are behind, but it seems really huge when you think about it. Their Ray Tracing still garbage however...but imagine a 7900XTX with a measured advantage in Raster, bad RT but still a decent improvement from RDNA2. I think that would really make people think a bit:
---significantly smaller die
---significantly less power draw....beating brute monster 4090 in pure raster. I think people would be drooling over that...even if the price were more at 1200 or so. I don't think anyone would be talking about how garbage the RT performance is, even if it is the same as the current reality. Instead, we'd be wondering how much more space is available for the GCU on a bigger chip, and what, maybe, some dedicated RT silicon could do.

I think people would be going nuts even with the same garbage RT, yet a decent boost in raster. I guess we might know if that is indeed something that could have been if the "fixed" N32/33 manage to perform as the rumors and leaks expected, based on paper numbers.