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

 

[Heartbreaker]

I personally think it's more the number of chips than the capacity. Perhaps nVidia's decisions with the bus size makes more sense if they were expecting the Ada Refresh to get 3 or 4 GB/chip.

IIRC, it was you that pointed out 16GB of GDDR is ~$200. That's a big hit to the parts list cost for a low to mid range card.

 

[Rigg]

Maybe low and mid don't need 16GB, which is going to add significantly to the BOM.

I guess this depends on how you delineate the tiers. I consider 60 class to be mid range although some might consider them low end. Raster price to performance and graceful aging due to more VRAM are AMD's primary selling points in the mid range. Given all the attention in the media surrounding VRAM it's good marketing to use it as a major selling point. I believe they need all the selling points they can get vs the Nvidia cards to make up for their inferior RT, upscaling, and frame gen features.

I think the full N32 SKU is DOA without 16 GB. Cut down N32 could probably get away with 12 GB if they need to get real aggressive with pricing. They could probably do fine with 8 GB for N33 if the cards are cheap enough. 8 GB is quickly becoming the entry point for DGPU and there should be an entire tier below the 7600 cards. Since 12 GB isn't an option for N33, offering more expensive 16 GB options might be a more effective way to compete against the Nvidia 60 cards. It's hard to say until we know how good the price to performance is on the 60 cards.

 

[MrTeal]

Is there a source on 16GB being $200, and is that retail or cost?

Seems a little out when the 12GB launched at and is currently available for $330 at retail, and the possibly subsidized A770 16GB is $350. DRAM exchange said a 1GB GDDR6 chip is average $3.386, or $54 for 16GB.

 

[leoneazzurro]

200$ for 16 Gbytes of GDDR6 is too much. It may have been as high as that during the covid and mining crysis, but prices went down consistently and you can bet now it's -at best- half that value (and that is still too much in my opinon). Otherwise 8Gbyte cards could not be sold at the 230€ mark (in average, we are seeing 8Gbyte 6600 being as low as 200$-220$ at Newegg) we are seeing on the 6600 nowadays.

 

[jpiniero]

Is there a source on 16GB being $200, and is that retail or cost?

Just a guess. I do imagine the speed affects the price a bit.

 

[Heartbreaker]

Is there a source on 16GB being $200, and is that retail or cost?

Seems a little out when the 12GB launched at and is currently available for $330 at retail, and the possibly subsidized A770 16GB is $350. DRAM exchange said a 1GB GDDR6 chip is average $3.386, or $54 for 16GB.

From when this came up before, you can check memory wholesalers for GDDR6:

Avnet: Quality Electronic Components & Services

Avnet is a global leader of electronic components and services, guiding makers and manufacturers from design to delivery. Let Avnet help you reach further.

www.avnet.com

In volumes of 100,000+ it's about $10 to $15/GB, depending on speed. Bleeding edge speed $15, lower speed $10.

AMD/NVidia will cut better deals than that though.

In volume electronics productions, saving $1/product shaving off a few passive components is a big deal, so regardless of fluctuations, VRAM is going to be a big expense that they will want to minimize in order to maximize profits.

 

[MrTeal]

You'd be much better looking at dramexchange.com for spot prices rather than looking at tape reel prices at Avnet or findchips.com.
They also list 8Gbit DDR4 ICs all over $5 in quantity (https://www.avnet.com/shop/apac/c/m...umber=1#facet=ads_f13671_ntk_cs:"DDR4+SDRAM"&) which would be $80 for 16GB, whereas you can go on Amazon today and buy a 16GB kit of DDR4-3200 for $30.

 

[maddie]

There goes that whole fantasy of the HUGE cost of adding ram to cards. About time.

 

[maddie]

200$ for 16 Gbytes of GDDR6 is too much. It may have been as high as that during the covid and mining crysis, but prices went down consistently and you can bet now it's -at best- half that value (and that is still too much in my opinon). Otherwise 8Gbyte cards could not be sold at the 230€ mark (in average, we are seeing 8Gbyte 6600 being as low as 200$-220$ at Newegg) we are seeing on the 6600 nowadays.

The argument used to rationalize the high memory cost was that these cards are losing money for the AIB vendors, or at best, selling at cost. Bizarre,

 

[TESKATLIPOKA]

The cut down N31 was universally panned for that $900 price by reviewers. It was also widely rejected by consumers which is why it can be had for $770 currently. The MSRP on that card should have been $700 IMO. $750 at most. The 7900 xt is a terrible example to use in a pricing argument. It's MSRP was a complete and total sales & marketing blunder.

The consumer GPU market seems to dictate that AMD cards need to cost something like 15-20% less than Nvidia at the same tier of performance to sell well. The market appears to be rejecting the 4070 at it's $600 MSRP despite it's software feature and mind share advantage over AMD. A major factor in this is obviously price/VRAM. I'd argue AMD doesn't have any choice but to offer 16 GB in the mid price tiers. How much do you think an AMD 16 GB 4070 competitor should cost to sell in high volume? I think it's pretty clear that it can't be more than $500 to succeed. While I'd like to see it at $450, it's probably more realistic to hope for $500.

N32 performance should be ~75-80% of 7900 XT depending on final clocks.
0.75(0.8) * $899 -> $674(719)
0.75(0.8) * $770 -> $578(616)
0.75(0.8) * $750 -> $563(600)
0.75(0.8) * $700 -> $525(560)

If we go according to current RX 7900XT price, then It should be $579(619).
Weaker cards have better perf/price so let's say It will cost $549(579).

$449-499 is very optimistic considering the direct competitor RTX 4070 still costs $599, which is 20-33% more.
Why should N32 cost only $449-499?
It would be cheaper than RX 6800($469.99) or RX 6800XT($529.99).

 

[coercitiv]

N32 performance should be ~75-80% of 7900 XT depending on final clocks.
....
It would be cheaper than RX 6800($469.99) or RX 6800XT($529.99).

Well, to be fair... the 6800XT is 75-80% of 7900XT.

I think @Rigg 's point was that the 4070 is due for a price cut to invigorate sales. Something like $550+. In this case AMD would need to price themselves 10% lower, which is $500. The price of 6800 and 6800XT is irrelevant in the longer run (supply will dry out), except for the fact that it's a good indicator for the price/performance ratio the market is expecting today.

If the 4070 starts selling well at $600 then N32 will also make sense above $500.

 

[Rigg]

N32 performance should be ~75-80% of 7900 XT depending on final clocks.
0.75(0.8) * $899 -> $674(719)
0.75(0.8) * $770 -> $578(616)
0.75(0.8) * $750 -> $563(600)
0.75(0.8) * $700 -> $525(560)

If we go according to current RX 7900XT price, then It should be $579(619).
Weaker cards have better perf/price so let's say It will cost $549(579).

4070 performance is ~80% of 4070 ti

0.8 * $800 -> $640

If we go according to current 4070 ti price, then It should be $640.
Weaker cards have better perf/price so let's say It should cost $600.

How is that working out for sales?

Edit: Used wrong MSRP by mistake

 

[Rigg]

Well, to be fair... the 6800XT is 75-80% of 7900XT.

I think @Rigg 's point was that the 4070 is due for a price cut to invigorate sales. Something like $550+. In this case AMD would need to price themselves 10% lower, which is $500. The price of 6800 and 6800XT is irrelevant in the longer run (supply will dry out), except for the fact that it's a good indicator for the price/performance ratio the market is expecting today.

If the 4070 starts selling well at $600 then N32 will also make sense above $500.

Yes this exactly my point. If the 7900 xtx needs to be 20% cheaper to sell well against the 4080, and the 4070 isn't selling well at $600, it is reasonable to assume that a direct competitor to the 4070 isn't going to sell well at more than $500.

 

[Timorous]

How about this option for AMD?

GPU	Die	Approx Config	Ram	Speed	Performance	Cost
7800XT	N32	60CU + 4MCDs	16GB	20gbps	~6950XT	$600
7800	N32	54CU + 4MCDs	16GB	20gbps	~6900XT	$550
7700XT	N32	48CU + 4MCDs	16GB	18gbps	~6800XT	$500
7600XT	N32	40CU (2SE) + 3MCDs	12GB	18gbps	~6800	$400
7600	N32	36CU (2SE) + 3MCDs	12GB	18gbps	~6700XT	$350
7500XT	N33	32CU	8GB	18gbps	~6700 non XT	$280
7500	N33	28CU	8GB	16gbps	~6650 XT	$220

Not sure they would like the margins on this stack tbh but they might be forced down it given the lacklustre sales of the 4070 and the already negative perception of the 8GB 4060 and 4060Ti.

EDIT: If you look at the 7600XT there it would perform close to the current 4070 (a little behind perhaps) but given most people think the 4070 should actually be a 4060 and priced around the $400 mark anyway it would fit into that perf/$ realm.

 

[leoneazzurro]

The argument used to rationalize the high memory cost was that these cards are losing money for the AIB vendors, or at best, selling at cost. Bizarre,

Well, it IS a cost, also because there is not only material cost there, you have mounting, solderign, testing, and more. But simply nowadays the cost went down and the trend is going further on that route.

 

[Aapje]

Well, it IS a cost, also because there is not only material cost there, you have mounting, solderign, testing, and more. But simply nowadays the cost went down and the trend is going further on that route.

Most of those costs are minimal. The most expensive part is when they need a bigger PCB with more wiring, but that is still not that much more costly.

 

[Rigg]

How about this option for AMD?

GPU	Die	Approx Config	Ram	Speed	Performance	Cost
7800XT	N32	60CU + 4MCDs	16GB	20gbps	~6950XT	$600
7800	N32	54CU + 4MCDs	16GB	20gbps	~6900XT	$550
7700XT	N32	48CU + 4MCDs	16GB	18gbps	~6800XT	$500
7600XT	N32	40CU (2SE) + 3MCDs	12GB	18gbps	~6800	$400
7600	N32	36CU (2SE) + 3MCDs	12GB	18gbps	~6700XT	$350
7500XT	N33	32CU	8GB	18gbps	~6700 non XT	$280
7500	N33	28CU	8GB	16gbps	~6650 XT	$220

Not sure they would like the margins on this stack tbh but they might be forced down it given the lacklustre sales of the 4070 and the already negative perception of the 8GB 4060 and 4060Ti.

EDIT: If you look at the 7600XT there it would perform close to the current 4070 (a little behind perhaps) but given most people think the 4070 should actually be a 4060 and priced around the $400 mark anyway it would fit into that perf/$ realm.

According to TPU full N32 is 64 CU. 5 cuts seems like a lot for a 200mm² GCD. If they cut down the same percentage as N31 you'd get 64, 56, and 46 CU SKU's although that's probably not the right way to look at it. 6950XT performance seems a bit optimistic to me for full N32. The 7900XT is only 10-15% faster than the 6950XT in raster.

 

[Mopetar]

I think the bottom bin for N32 would be 40 CU. N33 has 32 CU at most so there's no reason to venture much closer to that. N32 also has the same chiplet based approach as N31 with separate MCDs. The GCD for N32 is going to be close in size to the total due size for N33, in part because of the memory controllers being off die as well as using the 5nm process.

I think prices will need to come down as well. The market right now isn't great and the 4070 struggling at $600 suggests that while consumers will pay top dollar for a halo product, the mid-range has become far more price sensitive. I don't think AMD can expect to see a lot of sales if they launch above $500. It would be far better to cut back on the last 10% of clock speeds so they can sell a less power hungry card with more moderate cooling at a lower price. Let the partners have that leftover space to try to sell OC cards and they'll probably be quite happy.

AMD might have a bit of an advantage due to the higher VRAM as there are a lot of newer titles where anything under 12 GB isn't going to run without settings turned down. The cut down 12 GB cards will probably stack up quite well against everything below the 4070 for that reason alone.

 

[Timorous]

According to TPU full N32 is 64 CU. 5 cuts seems like a lot for a 200mm² GCD. If they cut down the same percentage as N31 you'd get 64, 56, and 46 CU SKU's although that's probably not the right way to look at it. 6950XT performance seems a bit optimistic to me for full N32. The 7900XT is only 10-15% faster than the 6950XT in raster.

TPU have it wrong unless the angstronomics leak which has been 100% so far is incorrect. Looking at N31 though 3SEs with 20 CUs per SE seems like the config AMD will go with for N32 and if you look at an N31 die shot you can see exactly how it will be laid out.

5 cuts is a lot and they could easily scrap the 7800 and 7600 and just go with the 3 XT models instead.

 

[Heartbreaker]

AMD might have a bit of an advantage due to the higher VRAM as there are a lot of newer titles where anything under 12 GB isn't going to run without settings turned down. The cut down 12 GB cards will probably stack up quite well against everything below the 4070 for that reason alone.

Below? I think it will mainly be aimed directly at 4070 perhaps slightly above. Maybe extending as low as 4060 Ti Competitor.

7800 16 GB - Full fat everything. Out perform the 4070 by a bit, for about the same price (~$600).

7700 XT 12 GB - Similar to 4070 for $100 less (~$500)

7700 12 GB - Aimed slightly above 4060 Ti for better price. (~$400) - The one I find most interesting.

7600/XT are Navi 33 aimed at 4060 and below.

 

[insertcarehere]

Radeon 780m review here:

With both systems running LPDDR5-6400, 780m/7840HS is around 12-13% faster than 680m/7735HS:


LPDDR5-6400 looks to be adequate enough for 780m, not a ton of difference moving to LPDDR5-7500


Quoting the reviewer: "Instead of calling the iGPU RDNA3, might as well call it RDNA2.5".

 

[Saylick]

Radeon 780m review here:

With both systems running LPDDR5-6400, 780m/7840HS is around 12-13% faster than 680m/7735HS:
View attachment 80142
View attachment 80143
LPDDR5-6400 looks to be adequate enough for 780m, not a ton of difference moving to LPDDR5-7500
View attachment 80144

Quoting the reviewer: "Instead of calling the iGPU RDNA3, might as well call it RDNA2.5".

That's rather disappointing for a node shrink...

 

[IEC]

This is probably going to be an unpopular opinion, but IMO AMD cares little about the dGPU consumer market beyond doing what work is necessary to keep their chips reasonably competitive and in consoles. It just doesn't make them the money that CDNA and data center chips do. So they'd rather price their chips high (e.g. RDNA3) and get good margins even if it means they're 1) not making that many of them and 2) not gaining market share.

I'd be happy to be proven wrong with RDNA4.

 

[igor_kavinski]

Quoting the reviewer: "Instead of calling the iGPU RDNA3, might as well call it RDNA2.5".

Badly needs V-cache.

 

[gdansk]

That's rather disappointing for a node shrink...

They're packing the same number of CU. RDNA3 has already shown it doesn't clock high on N5. It has also shown the dual issue capability amounts to about 0% IPC increase. Another case of RDNA3 being a bit of a dud as usual.

And finally despite 780M having ~25% higher boost clock it doesn't seem to have a 25% higher typical clock rate. I guess that's the disappointing part.