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

 

[Timorous]

I agree, except the 7600XT being N32 based, cut down a lot. There is no need to introduce such a product. AMD will likely have excellent yields on N32 dies.

Maybe 7600 as N33 8 GB
7600 XT as N33 16GB

AMD don't really give different VRAM amounts of the same base die config a different brand name.

So my $280 7500XT in the speculation has ended up being a $269 7600 with that kind of performance.

That would probably mean a cut N32 with 40CUs (7600XT, plain 7700 does not really matter how they brand it) with 6750XT tier performance can probably max out at $350.

Sure there will be excellent yields on N32 but AMD are the ones who chose to make just 3 dies to fill a complete stack of products. They are going to need to overcut something somewhere to make it work and since N32 is smaller than N31 it makes more sense to overcut that one for cost reasons.

So the N32 lineup could be

60CU, aggressive V/F binning for high clocks - 6950XT tier performance. (7800XT)
54CU, normal v/f binning - 6800XT tier performance (7700 XT)
48CU, normal v/f binning - 6800 tier performance - maybe with 12GB of VRAM as well to further differentiate from the next SKU up. (7700)
40CU, normal v/f binning - 6750XT tier performance - 12GB of VRAM (7600XT)

Obviously those names can be shifted around a bit and there is room for a 7800XTX to use the 70CU N31 die if AMD want. Depends on N32 clock speeds though IMO.

 

[AtenRa]

$300 for the Custom model ???? no thanks

8GB cards in the meddle of 2023 should be priced at $249 max

AMD announces $269 Radeon RX 7600 RDNA3 graphics card - VideoCardz.com

AMD Radeon RX 7600 now officially The cheapest desktop RDNA3 graphics card thus far. AMD today officially announces its Radeon RX 7600 non-XT graphics card. This is the first desktop GPU that will fall under the $300 price level. The company will launch this GPU tomorrow, just one day after the...

videocardz.com

 

[Timorous]

$300 for the Custom model ???? no thanks

8GB cards in the meddle of 2023 should be priced at $249 max

AMD announces $269 Radeon RX 7600 RDNA3 graphics card - VideoCardz.com

AMD Radeon RX 7600 now officially The cheapest desktop RDNA3 graphics card thus far. AMD today officially announces its Radeon RX 7600 non-XT graphics card. This is the first desktop GPU that will fall under the $300 price level. The company will launch this GPU tomorrow, just one day after the...

videocardz.com

Where does it say $299 for the AIBs. it just says below $300 and it will be $269. If AIBs offer a $299 version it better be a 16GB variant (although I am sure ASUS will try with a STRIX at something stupid).

Also 29% faster than the 6600, if that holds up in actual reviews it will be okay.

It would be around 91 FPS on this chart.

And about 83% on this chart.

Here it would be around 86 FPS so just ahead of the 6700 10GB in performance.

At $269 it is probably just about acceptable IMO. I expect a price cut when the $299 4060 releases and maybe a $299 16GB variant to act as a spoiler for that 4060.

 

[Joe NYC]

AMD don't really give different VRAM amounts of the same base die config a different brand name.

If you are looking for consistency from AMD marketing, I would refer you to the 2023 laptop CPU naming.

Nvidia did it with the 4060 Ti, and that increases the odds.

I think the 7600 with N33 die will end up so cheap that AMD may not want to associate N32 with that price level (offering N32 as 7600 XT). I think the price range of 7600 for next 12 months will be $199 - $299.

N32 will be just fine as:
- 7800 (XT) with 16 GB
- 7700 (XT) with 12 GB slightly cut down die
with price range probably $399 - $599 (depending where the performance lands).

I really have not gone through the math, but I think the cost of N32 will be 2x the cost of N33. The way to look at it is not that N32 will be super expensive to make, but N33 will be super cheap to make. TSMC doesn't know what to do with its N6/N7 capacity.

So my $280 7500XT in the speculation has ended up being a $269 7600 with that kind of performance.

That would probably mean a cut N32 with 40CUs (7600XT, plain 7700 does not really matter how they brand it) with 6750XT tier performance can probably max out at $350.

Sure there will be excellent yields on N32 but AMD are the ones who chose to make just 3 dies to fill a complete stack of products. They are going to need to overcut something somewhere to make it work and since N32 is smaller than N31 it makes more sense to overcut that one for cost reasons.

I don't think AMD market share justifies more than 3 dies, and variable MCDs do add some differentiation. If there is a sliver of a tier left out, that's fine.

So the N32 lineup could be

60CU, aggressive V/F binning for high clocks - 6950XT tier performance. (7800XT)
54CU, normal v/f binning - 6800XT tier performance (7700 XT)
48CU, normal v/f binning - 6800 tier performance - maybe with 12GB of VRAM as well to further differentiate from the next SKU up. (7700)
40CU, normal v/f binning - 6750XT tier performance - 12GB of VRAM (7600XT)

Obviously those names can be shifted around a bit and there is room for a 7800XTX to use the 70CU N31 die if AMD want. Depends on N32 clock speeds though IMO.

I think 7800 XTX would be one tier too many. 7800 XT with full N32 would be just fine, and there is no reason for AMD to reintroduce anything based on N31 (unless there is a re-spin of the die with better performance.)

Cutting down N31 even further is not going to make anyone happy. Offering less performance and lower price at the ~same cost to AMD. What would make people happier is the same N31 products with a bit of a discount and that would not need any further investment on part of AMD and AIBs.

 

[Ranulf]

"Next Gen,1080p and beyond" with 10% more performance than a 6650XT. 29% more than a 6600, that costs around $200. Nevermind the 6700 cards.

Meh. Just not as bad of a meh. Two days of meh.

 

[moinmoin]

If you are looking for consistency from AMD marketing, I would refer you to the 2023 laptop CPU naming.

Those are excellent, I agree, especially wrt consistency. Everything else though...

I really don't understand the hate for the Ryzen Mobile 7000 model numbering scheme. Do you all seriously prefer Processor, Ultra and Z1? And all the model numbering combining both current and rebadged gens with the same scheme without any specific indicator of what gen is actually used?

 

[jpiniero]

I really don't understand the hate for the Ryzen Mobile 7000 model numbering scheme. Do you all seriously prefer Processor, Ultra and Z1? And all the model numbering combining both current and rebadged gens with the same scheme without any specific indicator of what gen is actually used?

Again, that was intentional because OEMs want Barcelo but didn't want it to seem old.

 

[Heartbreaker]

Performance slide from AMD compared to RX 6600 - 29% better

AMD announces $269 Radeon RX 7600 RDNA3 graphics card - VideoCardz.com

AMD Radeon RX 7600 now officially The cheapest desktop RDNA3 graphics card thus far. AMD today officially announces its Radeon RX 7600 non-XT graphics card. This is the first desktop GPU that will fall under the $300 price level. The company will launch this GPU tomorrow, just one day after the...

videocardz.com

If they keep the 30% improvement for the 6600XT, it will need to be Navi 32...

 

[Timorous]

Performance slide from AMD compared to RX 6600 - 29% better

AMD announces $269 Radeon RX 7600 RDNA3 graphics card - VideoCardz.com

AMD Radeon RX 7600 now officially The cheapest desktop RDNA3 graphics card thus far. AMD today officially announces its Radeon RX 7600 non-XT graphics card. This is the first desktop GPU that will fall under the $300 price level. The company will launch this GPU tomorrow, just one day after the...

videocardz.com

If they keep the 30% improvement for the 6600XT, it will need to be Navi 32...

30% ahead of the 6600XT would be around 6700XT performance which can be done with a 40CU 2SE 12GB N32 config I expect, could probably use quite relaxed v/f binning as well.

 

[igor_kavinski]

Let's hope the 7600 XT has at least 10GB, if not 12GB and goes for $329.

 

[PJVol]

new thread

 

[Timorous]

Let's hope the 7600 XT has at least 10GB, if not 12GB and goes for $329.

Judging by how locked down the 7600 is in the BIOS / drivers N33 might have a bit more to give. It seems like you can get 10% more from overclocking so 20gbps ram + 3Ghz boost clocks might get you between the 6700 and 6700XT. Give it 16GB of VRAM as well and you have a potential 7600XT.

Personally I would prefer a 16GB variant of the 7600 we have and make the 7600XT N32 based but AMD may not want to do that if they don't feel they absolutely have to.

 

[moinmoin]

Again, that was intentional because OEMs want Barcelo but didn't want it to seem old.

Again, what's bad about that over how it's been done by all manufacturers up to now? Especially considering consistency this approach is much better than rebagded gens sharing the exact same numbering scheme but without the gen indicator.

 

[Joe NYC]

Let's hope the 7600 XT has at least 10GB, if not 12GB and goes for $329.

If it isn't 16GB, it would have to have reduced bus width (I think), which would defeat the purpose of being 7600 XT a model above 7600, by losing in a number of games that don't need > 8 GB.

 

[Mopetar]

Let's hope the 7600 XT has at least 10GB, if not 12GB and goes for $329.

Seems rather doubtful. It would have to be Navi 32 since there's the 7600 is pretty much the best to be expected out of Navi 33.

The only issue is that Navi 32 is a much more expensive product given it uses the 5nm node for the main die and requires at most 4 memory chipsets, which I'll grant are much less expensive, but still add to the cost.

Even if they do make some kind of salvage tier part, they won't want to sell it for only $330. Expect $400 minimum for it, but probably at least 12 GB of VRAM to compete with the 4060 Ti. The other possibility is they only use 2 memory chiplets with a 128-but bus and 16 GB of VRAM, but that probably cripples the performance since the infinity cache is on the memory chiplet modules now and that'd mean having too little cache for anything above 1080p.

Even if they have a massive price hole between a ~$400 bottom barrel Navi 32 card and the $270 (and probably soon to be much cheaper) 7600, is there much value in filling that hole beyond trying to target people willing to spend $350, but not a penny more? I'm skeptical that there's any ability for AMD to maneuver in that space with RDNA3 unless they make another die with 48 CU or something like that.

Just offer a $400 card with at least 4060 Ti performance but enough memory to alleviate concerns about viability for the next several years and people will probably buy if even if they have to grumble about paying $50 more than they'd like to.

 

[Heartbreaker]

Even if they do make some kind of salvage tier part, they won't want to sell it for only $330. Expect $400 minimum for it, but probably at least 12 GB of VRAM to compete with the 4060 Ti.

Yeah, that's reasonable, and I would expect it to be quite well received in comparison to the 4060 ti.

4060 Ti suffers both from lack of VRAM and lack of BW. 3 Chiplet/12GB N32 can fix both. Plus AMD give it enough CU to beat 4060 ti performance now that it's well known.

 

[igor_kavinski]

N32, the last bastion of hope for cash-strapped gamers

 

[jpiniero]

Just offer a $400 card with at least 4060 Ti performance but enough memory to alleviate concerns about viability for the next several years and people will probably buy if even if they have to grumble about paying $50 more than they'd like to.

May as well just let RDNA2 continue to deplete the supply.

 

[insertcarehere]

N32, the last bastion of hope for cash-strapped gamers

If that's the case might as well plop down some money for a 6800XT while stocks last, with how N33 performs it is highly unlikely N32 will be better by much anyway.

 

[Saylick]

N32, the last bastion of hope for cash-strapped gamers

Me: *takes out my RX 6800 from the case and hugs it*

 

[GodisanAtheist]

Given the 84CU 7900xt and the 32CU 7600 perform essentially identically to their RDNA2 equivalent CU parts (6950XT and 6650xt) I think it's safe to say RDNA3 is a complete flop of an arch.

The idea that a 60/64CU N32 is going to perform like an 80cu 6900xt or even a 72CU 6800xt feels laughable at this point. We're basically looking at 6800 performance +5-10% at best from N32.

Probably why it's totally AWOL at this point, since the packaging costs of a Chiplet design at that tier are just making the economics funky, or N32 needs FSR 3 and framegen to pitch just like Ada.

 

[jpiniero]

Given the 84CU 7900xt and the 32CU 7600 perform essentially identically to their RDNA2 equivalent CU parts (6950XT and 6650xt) I think it's safe to say RDNA3 is a complete flop of an arch.

That's not really true for the 7900 XT. The 6950 XT might be looking worse in more recent games, but Compubase had the average gap at 21% at 1440p.

 

[insertcarehere]

That's not really true for the 7900 XT. The 6950 XT might be looking worse in more recent games, but Compubase had the average gap at 21% at 1440p.

That it does, but given that the 7900XT:
- Clocks at 2.5-2.6ghz stock vs 2.3-ish for 6950XT
- Has marginally more CUs (84 vs 80)
- Has significantly more memory bandwidth (800gb/s vs 576 gb/s)
- Has 50% more ROPs clocking faster (192 vs 128)

What are the 7900XT improvements vs 6950XT that can't be attributed to AMD throwing more CUs, backend, and using the process advantage to clock higher?

To put another way, would RDNA2 on 5nm with over double the transistor budget of N21 (that's what N31 is using) really perform worse than what 7900XTX is doing?

Its now legitimately possible that a very cut down version of AD104 (RTX 4070), which itself is 300mm^2 on N5, would be a rough match for fully enabled N32, which takes a similar area, while being more efficient, better at RT, and dedicating more die area to HW (E. G tensor for DLSS) for other features (irrespective of whether those features are liked). If that's not an indictment on RDNA3 then idk what is.

 

[GodisanAtheist]

-Here is AMD saying the 7900xt is 7% faster than the 6950xt (granted driver improvements might have picked that up a bit, this is from January 23).

RDNA3 seems to really love a handful of games where I guess the dual issue really finds it's footing, but across the vast majority of games it seems to barely outperform RDNA2.

 

[PJVol]

Given the 84CU 7900xt and the 32CU 7600 perform essentially identically to their RDNA2 equivalent CU parts (6950XT and 6650xt)

Not quite.
n31 xt and n22 xtx clocked pretty much the same, 7900 cache-memory speed and ROPs number overprovisioned.
Based on aggregate reviews data 7900xt is ~ 20% faster, which as you might guess, 4 additional CUs (5% more) can't account for. So the rest is basically IPC gain.

Besides, N33 and N31 CU seem to be different. Some reviewers mentioned halfed L0 vector cache compared to n31.
And also take note of the transistors count in N31 and N21 - 57.7 bn vs 26.8 bn resp., while N33 is only 2 bn more compared to N23 (13.3 vs 11.1). Even taking into account the interconnect logic of N31, such big difference suggests the N33 CUs are cut down.