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

 

[jpiniero]

There, that's the pretty good part. How exactly is that a bad thing and not a win for us?
If anything, have people not learned anything from the success of Polaris?

Presumably it would greatly impact whether AMD bothers with the N33 desktop release.

The 6600 is under $200 now.

 

[Timorous]

- The 7600S that was tested was clocking at ~2.4 ghz, we don't know what the 6800S clocked at in the comparison, but from other tests on the laptop it tends to settle at around ~2.1ghz in an 80w envelope, so 7600S most likely enjoyed a sizable clock speed advantage...and still lost out.
- The 9% IPC increase was based on comparing a 6900XT vs 7900XT at 4K, but there are other factors at play there which I have already explained before:



A 7600XT wouldn't have 50% more bandwidth nor 50% more ROPs than a 6650XT...

Cross comparing laptop results is always a problem. The TUF they had was the FA617NS judging by the fact it had the 7735HS CPU. Also means it was a Zen 3+ CPU and given there are two configs with just single channel ram and they do not specify what ram config their model had we don't know if it was running dual channel or single channel.

Given the number of potential variables between this system and the G14 that had the 6800S in it drawing any conclusions from such data has rather wide error bars that only grow when trying to extrapolate such results from the laptop arena to the desktop arena. We also see frequently that two different laptops with the same GPU, CPU and memory config can offer very different gaming performance just based on the cooling solution so even when the components are the same it is still not apples to apples unless the chassis is a match as well.

It might be flawed data but it is the best independent test we had. If we went with AMDs claimed IPC uplift number the target would be closer to 6750XT / 6800 which from what we know so far seems a bit far fetched. It is always going to be a ballpark figure anyway due to the nature of guessing based on specs. If it ends up being closer to an OC'd 6650XT I would not be surprised. If it ends up just about matching the 6700XT at 1080p and maybe 1440p also would not be surprised. Slower than the former and quicker than the latter would be surprising though.

 

[Mopetar]

AMD might not be in a rush to release Navi 33 parts to the desktop market, particularly if there's still a lot of Navi 23 stock left in the channel.

Navi 33 should be less expensive for them considering that the die area is smaller and that TSMC was alleged to be offering a discount for moving wafers from 7nm to 6nm.

It's hard to say what AMD uses for a launch price as the 6600/XT and other Navi 23 parts were launched during the mining boom and were widely regarded as being too high. At $250 they probably would do quite well, just because there's nothing new at that price point.

 

[jpiniero]

It's hard to say what AMD uses for a launch price as the 6600/XT and other Navi 23 parts were launched during the mining boom and were widely regarded as being too high. At $250 they probably would do quite well, just because there's nothing new at that price point.

But that was before the price hikes. Maybe TSMC will cut N7/N6 prices back if demand for it continues to wane but so far they haven't admitted to doing so. So unless that changes AMD is likely paying more for N33 than N23 when accounting for the hikes.

I don't know what the magic number is, but $200/$250 definitely aint it.

 

[Timorous]

AMD might not be in a rush to release Navi 33 parts to the desktop market, particularly if there's still a lot of Navi 23 stock left in the channel.

Navi 33 should be less expensive for them considering that the die area is smaller and that TSMC was alleged to be offering a discount for moving wafers from 7nm to 6nm.

It's hard to say what AMD uses for a launch price as the 6600/XT and other Navi 23 parts were launched during the mining boom and were widely regarded as being too high. At $250 they probably would do quite well, just because there's nothing new at that price point.

I think AMD might try and stick to the current 6600 and 6650XT pricing given N33 can drop into the same PCBs as those parts. Just do a box update and you have a 7600 in the exact price point of the 6600 but offering better performance and the same for the 7600XT replacing the 6650XT with more performance. AMD could also offer a 16GB 7600XT variant if they so wished and charge a $50 or so premium for it. Would be the one to go for if you wanted to hold onto it for a few years where as the 8GB version would be great if you are using it as a stop gap card until the next gen parts are released.

But that was before the price hikes. Maybe TSMC will cut N7/N6 prices back if demand for it continues to wane but so far they haven't admitted to doing so. So unless that changes AMD is likely paying more for N33 than N23 when accounting for the hikes.

I don't know what the magic number is, but $200/$250 definitely aint it.

The price hikes will impact both N23 and N33 going forward and if the price of a 6600 is $200 because that is what that level of performance from AMD can command selling a 7600 in that price point using the cheaper to make N33 will net AMD more margin so that is a win for AMD.

8GB cards have a price ceiling now so unless AMD decide to make all N33 designs 16GB sub $300 is where it needs to land.

 

[Aapje]

16 GB is not going to happen as no 32 Gbit memory chips exist.

I expect AMD to initially have higher prices for the 7600 (XT), so they can gradually lower them & still get rid of the remaining 6000-cards.

8GB cards have a price ceiling now so unless AMD decide to make all N33 designs 16GB sub $300 is where it needs to land.

AMD has a habit of pricing their cards too high at first, so once you determine what price it should be, you need to increase that price to where AMD will price it at first, to ensure poor reviews and low sales.

 

[Kaluan]

Presumably it would greatly impact whether AMD bothers with the N33 desktop release.

The 6600 is under $200 now.

Speak of the Devil...

NVIDIA stops supply of chips, the RTX 4060 Ti comes at the end of May/beginning of June and AMD shows the RX 7600 at Computex (Update) | igor´sLAB

Let’s start the graphics card week with today’s review of an RTX 4070 from Palit and an information that I also got confirmed in the meantime: NVIDIA will stop supplying its board partners for a few…

www.igorslab.de

What I can confirm, however, is that individual (!) AMD board partners will show a Radeon RX 7600 as a finished product at Computex, while other partners, who use both AMD and NVIDIA chips, are first exercising wait-and-see restraint. To paraphrase politely. The statement that you don’t want to produce anything just to satisfy production and where you don’t see any chances of profits then hurts a bit. There is no price basis for a Radeon RX 7700XT at the moment, because you could even slip into the loss zone because the target group shifts and the price shouldn’t fit anymore.


Of course, it’s a lot of speculation, but the current price distortions due to a strong increase in inflation are real, but obviously haven’t reached the consciousness of many consumers yet. Most of them only see butter, bread, vegetables and fruit. And there we have already reached 20 to 30 percent often enough. And this is not only due to the evil retail trade (but probably also with). And so, in the foreseeable future, the graphics card classes will probably also be oriented more towards the income classes and no longer only towards the actual performance.

 

[TESKATLIPOKA]

16 GB is not going to happen as no 32 Gbit memory chips exist.

You don't need 32 Gbit chips to have 16 GB with 128-bit. Clamshell could be used for doubling Vram. The question is If It's worth It for N33 and some AIB or AMD are willing to do It.

 

[jpiniero]

Speak of the Devil...

Like if they try to sell only the cut N33 model, and at the original 6600 MSRP ($329)... that probably wouldn't go over well given that you should be able to get N22 for only slightly more.

 

[Aapje]

You don't need 32 Gbit chips to have 16 GB with 128-bit. Clamshell could be used for doubling Vram. The question is If It's worth It for N33 and some AIB or AMD are willing to do It.

It's indeed not needed, but the consequence would be that the 16 GB cards would have half the bandwidth per GB on the card, compared to a 256 bit card with 16 GB. It's very doubtful that they would do that because of all the headaches it would create for them (both in the marketing and because it could severely cannibalize the 7800 and 7700 sales).

I could only see them do this if they somehow cancel 7800 and 7700.

 

[Heartbreaker]

It's indeed not needed, but the consequence would be that the 16 GB cards would have half the bandwidth per GB on the card, compared to a 256 bit card with 16 GB. It's very doubtful that they would do that because of all the headaches it would create for them (both in the marketing and because it could severely cannibalize the 7800 and 7700 sales).

Navi 33, is a 128bit bus GPU, so 256 bit cards are not possible, so that's an irrelevant point.

It's simply a case of 8GB vs 16GB on 128 bit bus.

The downsides that make it unlikely are, that you need a new, more expensive different design board, for the 16 GB cards, to double up chips on each channel, plus the added expense of extra VRAM, so probably at least $100 added to the retail price of the card, and considering this is kind of low end product aimed mostly at 1080P, that is probably too much.

Though it would be nice if AMD just gave their AIBs permission to do that if they feel like it.

 

[TESKATLIPOKA]

It's indeed not needed, but the consequence would be that the 16 GB cards would have half the bandwidth per GB on the card, compared to a 256 bit card with 16 GB. It's very doubtful that they would do that because of all the headaches it would create for them (both in the marketing and because it could severely cannibalize the 7800 and 7700 sales).

I could only see them do this if they somehow cancel 7800 and 7700.

As @guidryp already mentioned, N33 has 128-bit bus, you can't make 258-bit out of It. Because It has only 128-bit bus, you can have 8GB Vram or 16GB Vram in clamshell, BW won't change only Vram can double.
Even If N33 had 256-bit bus, It wouldn't cannibalize 7700-7800, because N32 has 60CU instead of 32CU in N33 and even the cut-down version should have at least 48CU enabled, most likely more.

 

[Aapje]

@guidryp @TESKATLIPOKA

I didn't compare it to a 256 bit card to argue that AMD could change the IO of N33, but to argue that it would mess up their marketing. They somehow have to convince normies that a 16 GB card with less bandwith is considerable better than 8 GB, so they are willing to pay more, but not as good as a proper 16 GB card with a 256 bit bus. And that it's also worse than a 12 GB 7700 XT, assuming that they want to sell that one as well.

Even If N33 had 256-bit bus, It wouldn't cannibalize 7700-7800, because N32 has 60CU instead of 32CU in N33 and even the cut-down version should have at least 48CU enabled, most likely more.

Except that a very large percentage if not a majority buys based on 'bigger numbers better' and the CU count is not on the box. If you look on front of AMD Radeon boxes, the three information boxes that are on every box and thus almost certainly are mandatory are:
- Targeted resolution
- VRAM quantity
- PCIe standard

So the VRAM quantity is actually one of the things they market most prominently with and that they've conditioned casual buyers into comparing between cards. One of the three other boxes, the PCIe version, will be equal between all cards. If AMD chooses to market both the 7600 XT 16GB and 7700 XT 12 GB as a 1440p card, the resolution box is also going to be the same. So then the 7700 XT 12 GB looks worse than the 7600 XT 16GB when comparing the only specs visible on the front of every box. If people are picking a card for 1440p, then even if the 7800 XT 16GB shows that it is good for 4K, buyers may think: "I'm gaming on 1440p, so the 7600 XT provides identical performance to the 7800 XT, since the only difference is 4K support, which I don't need."

Of course that isn't true, but a noob picking out a card in the store often isn't going to realize that.

 

[linkgoron]

Speak of the Devil...

NVIDIA stops supply of chips, the RTX 4060 Ti comes at the end of May/beginning of June and AMD shows the RX 7600 at Computex (Update) | igor´sLAB

Let’s start the graphics card week with today’s review of an RTX 4070 from Palit and an information that I also got confirmed in the meantime: NVIDIA will stop supplying its board partners for a few…

www.igorslab.de

Smaller RDNA3 arriving in June is just sad. AMD still using RDNA2 kind of reminds me of AMD reusing Polaris for a few gens, while not having something very competitive in the high end because Vega was just barely competitive with the 1080. Sure, N31 is much more competitive with the 4080 than what Vega64 was with the 1080, but it's just bad. I hope that AMD can fix the supposed bugs that exist with N33 and N32 or a N31v2 (if it's not just that the architecture is fundamentally bad), but I'm doubtful. Hopefully RDNA4 will be better, but it's surely far away in the future.

 

[GodisanAtheist]

Wondering if it would have made more sense for AMD to just port their RDNA 2 designs to TSMC 6N and just rebrand them.

There is no way AMD started designing N32 expecting to sell it for $800 or whatever N22 was going for during the crypto boom, it was built to be sold taking TSMC's 5nm waffer costs into account as well as market segmentation, so the more I think about it the less I think N32 is waiting in the wings for pricing and margin reasons.

 

[gdansk]

Navi32 would have had a similar target MSRP as Navi22. $600 maximum.

If Navi32 doesn't ever ship it'll be because they forecast low sales. With low enough sales the relatively cheaper manufacturing cost would never pay off the GCD masks etc. If that is the case then the low sales also deters a shrink of Navi21.

 

[jpiniero]

Wondering if it would have made more sense for AMD to just port their RDNA 2 designs to TSMC 6N and just rebrand them.

You'd still have the same problem you have now in that there's still aparently plenty of unsold RDNA 2 floating around.

 

[Aapje]

With low enough sales the relatively cheaper manufacturing cost would never pay off the GCD masks etc. If that is the case then the low sales also deters a shrink of Navi21.

The real cost is in the initial design and the validation of the initial mask set, which should already have been done. Replicating masks is not that expensive, although if they need a respin, that is more costly. But it still seems unlikely that they would scrap an entire design until it would actually sell worse than to keep selling Navi 2x.

But I agree that a node shrink of Navi 2x doesn't make much sense.

 

[gdansk]

The real cost is in the initial design and the validation of the initial mask set, which should already have been done. Replicating masks is not that expensive, although if they need a respin, that is more costly. But it still seems unlikely that they would scrap an entire design until it would actually sell worse than to keep selling Navi 2x.

Let me put it the other way: If Navi 32 does appear it will be because it is substantially cheaper than Navi 21 to make. If it doesn't see mass production, then reasonably it is not cheaper to make.

Or perhaps tape out was delayed and the part is now incorporating RDNA3 bug fixes. But I find it very unlikely. Why invest so much - in this economy - to 'fix' the design in a respin when more Navi 21 would sell nearly as well.

 

[Kepler_L2]

Let me put it the other way: If Navi 32 does appear it will be because it is substantially cheaper than Navi 21 to make. If it doesn't see mass production, then reasonably it is not cheaper to make.

Or perhaps tape out was delayed and the part is now incorporating RDNA3 bug fixes. But I find it very unlikely. Why invest so much - in this economy - to 'fix' the design in a respin when more Navi 21 would sell nearly as well.

Why would anyone believe N21 is cheaper to produce? We know the die sizes already.

 

[Timorous]

Let me put it the other way: If Navi 32 does appear it will be because it is substantially cheaper than Navi 21 to make. If it doesn't see mass production, then reasonably it is not cheaper to make.

Or perhaps tape out was delayed and the part is now incorporating RDNA3 bug fixes. But I find it very unlikely. Why invest so much - in this economy - to 'fix' the design in a respin when more Navi 21 would sell nearly as well.

N32 will appear because it will be used in higher end laptops. N33 only covers 7600M XT and 7700S. The 7700M, 7800M and 7800S are going to need bigger dies which is where N32 comes in.

 

[jpiniero]

Why would anyone believe N21 is cheaper to produce? We know the die sizes already.

The price gap between N7/N6 and N5/N4 would have to get rather wide for that to happen. That being said, N32 would be a lot more expensive than N22.

 

[leoneazzurro]

The price gap between N7/N6 and N5/N4 would have to get rather wide for that to happen. That being said, N32 would be a lot more expensive than N22.

Total area for N32 is WAY lower than N21 (approx. 350mm^2 vs 521mm^2), yields for smaller chips are higher and wafer area is better utilized. Yes, it is more expensive than N22 but N22 is clearly significantly less powerful. And prices went high for everyone. Only problem in the confrontation with Nvidia is that they used too little N5 silicon OR they overestimated the target clocks. Something is definitely off because a 384bit bus+IC is defiitely overkill for the 7900XTX and probably it was meant to feed something more than what we got.

 

[jpiniero]

because a 384bit bus+IC is defiitely overkill for the 7900XTX and probably it was meant to feed something more than what we got.

Clearly N31 was intended to be a 4090+ competitor.

 

[Aapje]

Clearly N31 was intended to be a 4090+ competitor.

It's quite peculiar that the dual issue shaders seem to do nearly nothing, as the performance difference between the 6950 XT and 7900 XT can be explained by the higher clock speed, wider bus, and the extra shaders. The choice by AMD to pretend that the extra execution units don't exist also suggests that they do nearly nothing.

'Feeding the beast' is always one of the bigger challenges with processors. It's hard to make optimal use of the hardware that exists, when the workload is often not optimal for the hardware, so all kinds of tricks need to be done to alter the workload so it can be run efficiently. Perhaps AMD ran into big issues when trying to build a new scheduler that assigns work to their shaders and they stuck to the old one, that treats the shaders as single issue?

Delaying N32 could also be because they want to release it with a new scheduler.