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

 

[Mahboi]

Why so sure? AMD has a downturn in console demand and a vertical with falling revenue to prop up

You mean RDNA 4 driving prices down? It stops at roughly between a 5060 Ti and 5070's perf.
There's just no way Nvidia is going to care to price compete for any card above a 5070. And the lack of Halo product means they can easily carry on the narrative that "only Nvidia can do the best hardware".
It's just Lovelace vs heavily Gimped RDNA again, except this time it's lack of will from AMD, not an electrical problem.
Also AI is still in high demand for now.

 

[inquiss]

You mean RDNA 4 driving prices down? It stops at roughly between a 5060 Ti and 5070's perf.
There's just no way Nvidia is going to care to price compete for any card above a 5070. And the lack of Halo product means they can easily carry on the narrative that "only Nvidia can do the best hardware".
It's just Lovelace vs heavily Gimped RDNA again, except this time it's lack of will from AMD, not an electrical problem.
Also AI is still in high demand for now.

It will provide good value in the segments it's in yeah. They need to sell a lot of cards. Yep, Nvidia will continue with that narrative of having the best, so what? This is about whether AMD will price them for value, what's it got to do with Nvidia having the high end. Even if AMD had the best card at the top there would be some other reason why Nvidia is the best anyway.

Think adroc had been mentioning. It's not lack of will, it's to keep the roadmaps going. RDNA3 underperformed, can't have that again and can't waste all this time fixing. Considering the pressure on advanced packaging to sell AI chips, looks a good decision. RDNA 5 will probably have a very big version and supply for packaging should have improved by then, maybe.

This new version of RDNA shouldn't be gimped. It covers the biggest volume segments. Should compete pretty well.

 

[Mahboi]

It will provide good value in the segments it's in yeah. They need to sell a lot of cards. Yep, Nvidia will continue with that narrative of having the best, so what?

So it's hard to break when your marketing is braindead, and you have nothing to compete back.

This is about whether AMD will price them for value, what's it got to do with Nvidia having the high end.

Monstrous mistake there.
AMD's main problem is that they can't market the cards for their value because Nvidia has total narrative control due to the aforementioned braindead marketing.
If at least they have a strong flagship, it is a different story. But they won't even have a high end card.

So it's like going in a brawl with your mouth taped and both your legs tied. Nvidia will have no difficulty in setting the prices they want across the board, including low-end.

Even if AMD had the best card at the top there would be some other reason why Nvidia is the best anyway.

Think adroc had been mentioning. It's not lack of will, it's to keep the roadmaps going.

I don't buy that story for a second.
It's the same cope as the "AMD HAS to do chiplets because X Y Z".
Meanwhile Nvidia throws a 609mm² die with 144SMs out of which only 120 are active and they literally laugh at the wasted area and money.

It's not a "technical limitation", it's just that there's a corpo that doesn't mind wasting money and one that penny pinches every speck of copper on the road.

RDNA3 underperformed, can't have that again and can't waste all this time fixing. Considering the pressure on advanced packaging to sell AI chips, looks a good decision.

That is a completely backwards take?
RDNA 3 was bad, so let's make RDNA 4 tiny? What?

Shouldn't it be "RDNA 3 was bad, so let's make RDNA 4 great to compensate"? Yes it should.

RDNA 5 will probably have a very big version and supply for packaging should have improved by then, maybe.

This new version of RDNA shouldn't be gimped. It covers the biggest volume segments. Should compete pretty well.

Shoulda woulda coulda...
RDNA 3 shoulda been great
RDNA 4 woulda been great

Just don't expect me to be surprised or shocked if in a year we hear rumors of cancelled half the stack but "RDNA 5 coulda been great".
Actually considering the number of people who are, despite the recent track record, 100% certain of "RDNA 5 OMEGA HYPE", expect me to post every gif I can find on the internet about "I TRUSTED YOU" and "BETRAYAL".

 

[inquiss]

So it's hard to break when your marketing is braindead, and you have nothing to compete back.

Monstrous mistake there.
AMD's main problem is that they can't market the cards for their value because Nvidia has total narrative control due to the aforementioned braindead marketing.
If at least they have a strong flagship, it is a different story. But they won't even have a high end card.

So it's like going in a brawl with your mouth taped and both your legs tied. Nvidia will have no difficulty in setting the prices they want across the board, including low-end.

I don't buy that story for a second.
It's the same cope as the "AMD HAS to do chiplets because X Y Z".
Meanwhile Nvidia throws a 609mm² die with 144SMs out of which only 120 are active and they literally laugh at the wasted area and money.

It's not a "technical limitation", it's just that there's a corpo that doesn't mind wasting money and one that penny pinches every speck of copper on the road.

That is a completely backwards take?
RDNA 3 was bad, so let's make RDNA 4 tiny? What?

Shouldn't it be "RDNA 3 was bad, so let's make RDNA 4 great to compensate"? Yes it should.

Shoulda woulda coulda...
RDNA 3 shoulda been great
RDNA 4 woulda been great

Just don't expect me to be surprised or shocked if in a year we hear rumors of cancelled half the stack but "RDNA 5 coulda been great".
Actually considering the number of people who are, despite the recent track record, 100% certain of "RDNA 5 OMEGA HYPE", expect me to post every gif I can find on the internet about "I TRUSTED YOU" and "BETRAYAL".

Whole lotta text. The original reply to your point was that value was gonna be bad again this gen. I don't think it will be in the segments AMD will compete in with RDNA4. That's it.

I agree you need a big win if you want to change mind share. This gen isn't it.

On you point about RDNA, if it tanks then you make a bigger RDNA 4? Like, what? It was already being designed and made ready. In a few weeks you'll see what RDNA3 should have been. If you have issues in an org and execution isn't working, what you gonna cut to get the roadmap back on track. You need to launch things, you have to invest extra man hours fixing the damn thing that was broken, for future iterations. Something has to give. I don't need to convince you that RDNA3 failed in post SI, it's been written about in this thread and it's not long till you see what RDNA3 should have been. Think it's even been say who they brought in the fix things.

Anyway, the damned point was that in the segments that AMD is in I think they will offer good value because they need to fill graphics revenue while consoles suck. They need the share.

 

[Mahboi]

Whole lotta text. The original reply to your point was that value was gonna be bad again this gen. I don't think it will be in the segments AMD will compete in with RDNA4. That's it.

And if you had bothered to read the text, you'd have noted that no, it won't. And why. But you didn't so I'm not repeating myself.

I agree you need a big win if you want to change mind share. This gen isn't it.

On you point about RDNA, if it tanks then you make a bigger RDNA 4? Like, what? It was already being designed and made ready. In a few weeks you'll see what RDNA3 should have been. If you have issues in an org and execution isn't working, what you gonna cut to get the roadmap back on track.

All this cope text just to say "we have internal problems".
You could have just said "we have problems and we gave up". Which is pretty much what I'm saying.

Also, there's a world of difference between a "we had problems" and "we had problems so bad that our biggest card next gen is a 240mm² die". That's like "I had an oopsie with the car dad" and "I had an oopsie with the car and now there's only half a car left dad" difference. RDNA 4 is gimped down to so small you almost want to laugh.

Anyway, the damned point was that in the segments that AMD is in I think they will offer good value because they need to fill graphics revenue while consoles suck. They need the share.

We'll see. I expect a perfectly Dyatlov pricing.

Not great, not terrible.

 

[inquiss]

Ah, it makes a bunch of sense now.

I think value will be good this gen in the segments in which AMD competes, because I think AMD will offer reasonable value.

You think this will be a poor value gen because AMD won't be driving Nvidia pricing down. Have I got that right? So there will be a cheaper set (AMD) and a more expensive set (Nvidia), in the segments AMD competes in this gen) and that means this gen has no value. Only if you only wanna buy an Nvidia GPU. Get fleeced is how I feel about those customers. Who cares.

 

[Mahboi]

Have I got that right?

Not a single thing since earlier, not here nor in the Z5 thread, and it's time for bed. Good night.

 

[inquiss]

Not a single thing since earlier, not here nor in the Z5 thread, and it's time for bed. Good night.

Fair enough. At some point I'll work out what you're getting at. Sleep well, dream of NUCs with big market shares

 

[PJVol]

RDNA 3's entire problem is the power efficiency bust

I believe there's no power efficiency issues with the GCD, but there are with the rest of the GPU, i.e. fanout links and MCDs.
You can see this by looking at TimeSpy scores for the 6950xt and 7900xtx with the HWInfo telemetry attached, having the ~same power supplied by the GFX rail. I found two such results on forums where the reported gfx power was ~265W.

Graphics scores are 22,750 and 35,400 respectively, so for the GFX IP we're actually getting ~50% increase in perf/W. The total power reported was
~315W (370W*) and ~385W (460W) - tgp (tbp)

VRM losses aside, you can see that "uncore" accounts for
~120W and ~50W of their respective TGPs.

In other words, the "uncore" in rdna3 eats up a third! of the TGP, whereas it's only 17-18% for the rdna2.

 

[adroc_thurston]

I believe there's no power efficiency issues with the GCD, but there are with the rest of the GPU, i.e. fanout links and MCDs.

the opposite

 

[gdansk]

Phoenix IGP is a dud too (10% improvement going from 6nm to 4nm? lol) and there's no MCD to blame.

 

[linkgoron]

Phoenix IGP is a dud too (10% improvement going from 6nm to 4nm? lol) and there's no MCD to blame.

Navi 33 (7600/7600XT) has also already shown that RDNA3 is barely any better than RDNA2 even without MCDs.

 

[PJVol]

the opposite

Not sure about the "uncore" - maybe it really cannot eat less, but as for GCD, I don't see a noticeable problem, at least up to 3000 MHz gfx, or it's not as pronounced as if we move further along the VF curve towards Fmax, but what exactly was it at the STA?
And by saying that there's an issue anyway, what could it be then? Too high IR losses?

 

[igor_kavinski]

In other words, the "uncore" in rdna3 eats up a third! of the TGP, whereas it's only 17-18% for the rdna2.

That's awful! They could've fixed that with a new uncore. But I guess they decided to take the easy way out. People rightly got fired for such an obvious overlooked mistake.

Question is, was the uncore consuming more power on its own or was it doing that to satisfy the demands of the other chips? In which case, they had no concrete idea about what the chips' bandwidth/data consuming requirements would be.

 

[igor_kavinski]

the opposite

Why is the RDNA3 uncore power consumption higher then?

 

[Ajay]

Hmm, when do we start the RDNA5 and CDNA3 threads? Actually, maybe they should be split. Joking/not joking.
RDNA4 isn't blowing up my dress, and I'm eager to hear more about the MI400.

[Also trying to gauge whether or not AMD is a good buy - it's pretty high relative to forward revenue right now, waited too long]

 

[Ajay]

Why is the RDNA3 uncore power consumption higher then?

The fact that mobile required fixes (RDNA 3.5) to get power down tells you all you need to know.

 

[igor_kavinski]

The fact that mobile required fixes (RDNA 3.5) to get power down tells you all you need to know.

What were those GPU people doing??? Hiding their progress reports from Lisa??? How did she end up giving them approval to go ahead producing tens of thousands or even hundreds of thousands of GPUs with such a glaring flaw? She was most likely tricked.

 

[Aapje]

[Also trying to gauge whether or not AMD is a good buy - it's pretty high relative to forward revenue right now, waited too long]

It is probably too late to buy into both Nvidia and AMD at this point. The mindless investors who think that AI hardware will boom forever have overvalued these companies too much.

 

[PJVol]

Why is the RDNA3 uncore power consumption higher then?

It's obviously the cost of moving data - the fundamental challenge in the modern design. Going through the chiplet interconnects is quite expensive in terms of energy, plus MCDs being manufactured using a less advanced process node.

Phoenix IGP is a dud too (10% improvement going from 6nm to 4nm? lol) and there's no MCD to blame.

Actually, I didn't claim that the rdna 3 issues are solely attributed to the chiplet design. Obviously there are other factors.
And btw, IIRC the N31/N32 and N33/APU use different gfx IPs, so I'm not sure such comparisons are relevant.

 

[Ajay]

What were those GPU people doing??? Hiding their progress reports from Lisa??? How did she end up giving them approval to go ahead producing tens of thousands or even hundreds of thousands of GPUs with such a glaring flaw? She was most likely tricked.

Well, nobody's perfect. The engineering team failed a bit on execution, but then, AMD had to ship what they had to hit the market window. I'm sure she was appraised of the situation in advance.

 

[Ajay]

It is probably too late to buy into both Nvidia and AMD at this point. The mindless investors who think that AI hardware will boom forever have overvalued these companies too much.

Oh, it is right now. I should have said I'm looking for more positive news on MI400, and a lower stock price - which of course, is an oxymoron.

 

[igor_kavinski]

Oh, it is right now. I should have said I'm looking for more positive news on MI400, and a lower stock price - which of course, is an oxymoron.

Well buy now and then you won't regret it later

The June reveal of Zen 5 could cause the stock price to go up up up!

 

[jpiniero]

Navi 33 (7600/7600XT) has also already shown that RDNA3 is barely any better than RDNA2 even without MCDs.

Just based upon the 3dmark data, even if N33 had higher clock speeds, you're maybe talking an additional 10%. The 7600 only has a tad more memory bandwidth than the 6650 XT does.

 

[Mopetar]

Irrelevant. This isn't inflation, that's the story Nvidia fed people.

I wasn't claiming that it was, I was merely stating that perhaps both companies would continue selling cards at these existing price points for the next decade, over which the value of those price points would be eroded by inflation.

In other words once you adjust for everything and look at it from the perspective of PPP, the cards aren't more expensive than what would have been a $150 card a few decades ago.

That's not guaranteed to happen, but I'd like for that to be the case.