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

 

[insertcarehere]

N32 will be cheaper than N21 and might be a touch more expensive than N22 but I don't think there will be much in it. N33 is cheaper than N23.

With the current prices AMD could offer an N32 based 7800XT with 6950XT performance at $600 with more margin.

They could offer a $500 7700XT with 6800XT performance which would also have more margin than the current 6800XT parts.

I think even a really cut $380 N32 based 7600XT with 6750XT performance would probably have more margin than the current 6750XT does.

A $280 7500XT with 6650XT-6700 performance would be a similar or better margin than the top N23 perts as well.

So at every tier AMD can improve their margin position Vs current RDNA2 products and gain a lot of mindshare.

Also when you consider the 4070 should really be a 4060 at around $400 it is not like the above stack is fantastic, better than NV but nowhere like offering 6800XT perf at $400.

The reason we're seeing $600 6950XTs and so on at the moment is because AMD is trying to get rid of old stock and has presumably amortized/written-off enough of the fixed costs to make the reduced margins work. For reference, said RDNA2 products at launch MSRPs:
- 6950XT: $1099
-6900XT: $999
- 6800XT: $649
- 6800: $579
- 6750XT: $549
- 6600XT: $379

RDNA3 is a different sort of math in that they have a new set of R&D costs that need to be amortized per product, AMD would not spend that money just to earn the margin levels that they can get from clearing out RDNA2. Ideally speaking, their benchmark would be the margins that RDNA2 had at launch.

We'd already seen what that meant with N31 pricing, 7900XT was more cut down vs 7900XTX than 6800XT vs 6900XT, yet the price difference at launch was far narrower at $100 vs $350, which moved the planned "sweet spot" pricing for N31 from $649 to $999. I have no doubts the same will happen with N32/N33 whenever they come out.

 

[PJVol]

~

 

[Timorous]

Can you clarify? N31 has two rows of three SE's. How exactly are the 3 SE's arranged in N32, all in line?

If you draw a line down between the right most 2 MCDS then what is left is the display engine stuff whcih I imagine stays the same, 3 SEs and the links for 4 MCDs + a little extra because N31 SEs have 16CUs rather than the 20 N32 is supposed to have per SE. It seems like the obvious way to layout N32 to me based on what angstronomics leaked and it would also fit with the 200mm^2 die size.

The reason we're seeing $600 6950XTs and so on at the moment is because AMD is trying to get rid of old stock and has presumably amortized/written-off enough of the fixed costs to make the reduced margins work. For reference, said RDNA2 products at launch MSRPs:
- 6950XT: $1099
-6900XT: $999
- 6800XT: $649
- 6800: $579
- 6750XT: $549
- 6600XT: $379

RDNA3 is a different sort of math in that they have a new set of R&D costs that need to be amortized per product, AMD would not spend that money just to earn the margin levels that they can get from clearing out RDNA2. Ideally speaking, their benchmark would be the margins that RDNA2 had at launch.

We'd already seen what that meant with N31 pricing, 7900XT was more cut down vs 7900XTX than 6800XT vs 6900XT, yet the price difference at launch was far narrower at $100 vs $350, which moved the planned "sweet spot" pricing for N31 from $649 to $999. I have no doubts the same will happen with N32/N33 whenever they come out.

It was 100% clear to me that 7900XT was priced to sell 7900XTXs. It was done so AMD did not have to cut perfectly good dies at launch so they dampened demand with poor pricing. Now that the $1,000 tier is saturated (as seen by the fact stock is pretty easy to come by) the 7900XT price can drop down to be positioned more in line with the relative performance if not better . In the UK an MBA 7900XT can be had for £750 and the cheapest 7900XTX is £1,050 making the 7900XT a pretty compelling part if you want to spend that kind of money on a GPU.

Proposed price tiers.

GPU	Proposed Price	Performance	Die Size	Replaces	Old MSRP	Current price	Old Die Size
7800XT	$600	~6950XT	200mm N5 + 4x36mm N6 = 344mm	6800XT	$650	$510	520mm N7
7700XT	$500	~6800XT	200mm N5 + 4x36mm N6 = 344mm	6700XT	$480	$400 (6750XT)	335mm N7
7600XT	$400	~6800	200mm N5 + 3x36 N6 = 308mm	6600XT	$380	$300 (6650XT)	237mm N7
7500XT	$300	~6700	200mm N6	6500XT	$200	$160	107mm N7

Margins on the 7800XT would be better than the 6800XT. I think margins on the 7700XT would be pretty close to those on the 6700XT and the margins on the 7600XT and 7500XT would be worse than the MRSP of the 6600XT and 6500XT but both of those had horrible MSRPs. With the current pricing in mind the $300 7500XT would be slightly better perf/$ than the currently available 6650XT. The $400 7600XT would be slightly better perf/% than the currently available $400 6750XT and the $500 7700XT would be about the same as the currently available 6800XT.

This is not really moving the needle much for current buyers and while AMD can wait for RDNA2 stock to dwindle and then come in at higher prices (maye $50 across the board) if they did that the 7700XT is not looking that great vs a 4070. More vram and similar performance sure but people will gladly pay $50 for DLSS3 and the NV mindshare.

AMD have an opportunity to really steal some of that mindshare if they spec and price their parts correctly and I think the above does that. That 7600XT with 12GB of VRAM vs the $400 4060(Ti) with 8GB would be a great selling point. The 16GB 7700XT at $500 vs the similar performing $600 4070 would also get people to move to AMD and then the same priced 7800XT with performance that knocks on the door of the $800 4070Ti would also be something worth considering for a lot of people.

Do that and announce the new 7900XT MSRP to be $750 and AMD solid stack that would sell well IMO.

 

[Heartbreaker]

AMD have an opportunity to really steal some of that mindshare if ...

The excitement before every release, about AMD making a move this time, reminds me of when nerds used to get excited about this year, finally being the Year of the Linux Desktop. 😉

 

[PJVol]

If you draw a line down between the right most 2 MCDS then what is left is the display engine stuff whcih I imagine stays the same, 3 SEs and the links for 4 MCDs + a little extra because N31 SEs have 16CUs rather than the 20 N32 is supposed to have per SE. It seems like the obvious way to layout N32 to me based on what angstronomics leaked and it would also fit with the 200mm^2 die size.

This is too oversimplified. You've got 48 СU's left this way on the left side, but 60 takes 25% more space and the N32 SE layout is 2 SA * 5 WGP, not 2x8.
Here I added 2 WGPs missing to each leftside SE and cut three rightmost SEs from the n31 die shot :

 

[jpiniero]

6950 XT is already about $600 right now.

 

[Timorous]

This is too oversimplified. You've got 48 СU's left this way on the left side, but 60 takes 25% more space and the N32 SE layout is 2 SA * 5 WGP, not 2x8

N31 is 8 WGPs per SE in 2 banks of 4.
N32 would change this to 10WGPs per SE in 2 banks of 5.

I also expect L2 cache to be halved for N32 because it has half the shader engines.

Here is a clearer image.

If you can't figure out how N32 will be arranged from that without me actually sitting at my PC and doing the edit I can't really help you. You can very clearly see there is more than 6 WGPs worth of space available after you slice off the right hand side of the die to leave only 4 MCDs

Edit. Just seen your edit. N32 won't have the same L2 as N31. It has half the SEs so L2 is probably half as well and the command front end can be shrunk as well.

 

[insertcarehere]

It was 100% clear to me that 7900XT was priced to sell 7900XTXs. It was done so AMD did not have to cut perfectly good dies at launch so they dampened demand with poor pricing. Now that the $1,000 tier is saturated (as seen by the fact stock is pretty easy to come by) the 7900XT price can drop down to be positioned more in line with the relative performance if not better . In the UK an MBA 7900XT can be had for £750 and the cheapest 7900XTX is £1,050 making the 7900XT a pretty compelling part if you want to spend that kind of money on a GPU.

Proposed price tiers.

Again, comparing to RDNA2 where AMD were positioning the 6800XT at $649 as the sweetspot, they were likely plenty willing to cut down good dies for that pricing. Even a new "street" price of $800 for a 7900XT would still be a significant pricing increase gen-on-gen.

GPU	Proposed Price	Performance	Die Size	Replaces	Old MSRP	Current price	Old Die Size
7800XT	$600	~6950XT	200mm N5 + 4x36mm N6 = 344mm	6800XT	$650	$510	520mm N7
7700XT	$500	~6800XT	200mm N5 + 4x36mm N6 = 344mm	6700XT	$480	$400 (6750XT)	335mm N7
7600XT	$400	~6800	200mm N5 + 3x36 N6 = 308mm	6600XT	$380	$300 (6650XT)	237mm N7
7500XT	$300	~6700	200mm N6	6500XT	$200	$160	107mm N7

Margins on the 7800XT would be better than the 6800XT. I think margins on the 7700XT would be pretty close to those on the 6700XT and the margins on the 7600XT and 7500XT would be worse than the MRSP of the 6600XT and 6500XT but both of those had horrible MSRPs. With the current pricing in mind the $300 7500XT would be slightly better perf/$ than the currently available 6650XT. The $400 7600XT would be slightly better perf/% than the currently available $400 6750XT and the $500 7700XT would be about the same as the currently available 6800XT.

Proposals:
- 7800XT vs 6800XT at launch: Possibly, although that would heavily depend on the tradeoff between chiplet packaging + N5 vs a larger monolithic N7
- 7700XT vs 6700XT at launch: Absolutely not, N5+N6 with roughly equivalent die size in 2023 is more expensive than N7 in 2020, not to mention chiplet packaging costs, more GDDR6 VRAM,
- 7600XT vs 6600XT at launch: Setting aside that it is unlikely for N32 to have 3 cuts when the bigger N31 only has 2 for consumer at the moment, unless there are big problems with yields at N5 (there shouldn;t be), even a heavily binned N32 will have significantly thinner margins than N23 at similar pricing.
- 7500XT vs 6500XT: Same here, double the die space, vram...does not make for good margins, even if we're assuming a $100 increase in MSRP.

Keep in mind that R&D costs in the development period for RDNA3 (~2019-22) is way more expensive than RDNA 2 (~2016-19), on account of the massively increased personnel costs that every single U.S tech company faced during the COVID years as they outbid each other for talent. AMD's SG&A costs jumped from $995m in 2020 to $2.3B in 2022 for example, all these additional costs need to get amortized.

AMD have an opportunity to really steal some of that mindshare if they spec and price their parts correctly and I think the above does that. That 7600XT with 12GB of VRAM vs the $400 4060(Ti) with 8GB would be a great selling point. The 16GB 7700XT at $500 vs the similar performing $600 4070 would also get people to move to AMD and then the same priced 7800XT with performance that knocks on the door of the $800 4070Ti would also be something worth considering for a lot of people.

Do that and announce the new 7900XT MSRP to be $750 and AMD solid stack that would sell well IMO.

It's been nearly six months since RDNA3 has launched and so far we still don't have much in the way of official information, let alone products for RDNA3 GPUs other than N31, this doesn't strike me as a company that wants to get attention and marketshare. Frankly speaking, I don't think they care about mindshare and would be plenty happy taking those wafers and making datacenter GPUs/CPUs instead..

 

[Ajay]

I'm sure the sales distribution between them is better now that the 7900 XT can be purchased for as low as $760. The MSRP was clearly a marketing failure since it's only been 4.5 months since these cards released and it's already received a significant price cut. I think the naming scheme is bad marketing since the 6950 XT, 6900 XT, 6800 xt, and 6800 were actually competitive in raster performance with their 3090 TI, 3090, 3080 ti, and 3080 counterparts. The 7900 XTX gets curb stomped by the 4090. Based on the precedent that was set by the 6000 series cards, the current 7000 cards have no business being marketed as 90 tier cards when they actually compete with the 4080 and 4070 ti (AKA 4080 12 GB). The 4090 was on the market for a month before the 7900 cards were officially announced. They had time to call an audible and they didn't change course. They knew the 4090 was going to destroy the 7900 XTX and they didn't adjust the marketing to compensate. These cards should have been branded as 7800 XT and 7800 and priced a bit more aggressively IMO.

Yes, initial pricing was unjustified for the 7900XT. Whether or not it was worthy of the title, there needed to be at least one 7090 card in the stack - simply to imply that they had a 'halo' card. As we've often seen in the video card segment, marketing trumps logic.

Assuming that it would be financially feasible, there seems to be silicon level issues that might be helped with a new stepping. All they did for 6950 XT was give it more clock speed via power budget and some slightly faster memory.

There has been a lot of speculation about what is 'wrong' with RDNA 3 GPUs. Drivers, silicon implementation, flaw in architecture. I used to track this sort of thing on the Beyond 3D forums. But, the Architecture and Products sub-forum was locked as read only in January of this year - apparently due to a decline in civility. There are those on twitter who know the score better, but they are invite only accounts. That said, why do a re-spin if RDNA 4 products are only about a year away. Hopefully, RDNA 4 is on TSMC N3E, so that there will not be a half node difference between Nvidia and AMD has there is now.

 

[jpiniero]

Hopefully, RDNA 4 is on TSMC N3E

I'd be surprised if it is. N4 is more likely.

 

[Mopetar]

We have seen this for multiple generations. Why would you suddenly expect change?

Because AMD has hopefully also realized that they're not going to win at hearts and minds by following NVidia's pricing and only offering a slight performance edge in rasterization.

It's a down market so they can just ignore the bottom and try not to produce much, effectively writing off RDNA3 or they can price aggressively to tap in to consumers that want a new card, but not at $600 for something that might not last more than a few years due to low memory or other reasons.

If I were AMD I'd find the price point where they can come in a lot lower than NVidia even if they have to sell three times as many cards to make the same profit. They need to do something to get most of the market that won't even look at them to buy an AMD card. Anything else is just kicking the same can down the road for another generation.

 

[Timorous]

This is what Locuza came up with for N7 and N5 onwards

Again, comparing to RDNA2 where AMD were positioning the 6800XT at $649 as the sweetspot, they were likely plenty willing to cut down good dies for that pricing. Even a new "street" price of $800 for a 7900XT would still be a significant pricing increase gen-on-gen.


Proposals:
- 7800XT vs 6800XT at launch: Possibly, although that would heavily depend on the tradeoff between chiplet packaging + N5 vs a larger monolithic N7
- 7700XT vs 6700XT at launch: Absolutely not, N5+N6 with roughly equivalent die size in 2023 is more expensive than N7 in 2020, not to mention chiplet packaging costs, more GDDR6 VRAM,
- 7600XT vs 6600XT at launch: Setting aside that it is unlikely for N32 to have 3 cuts when the bigger N31 only has 2 for consumer at the moment, unless there are big problems with yields at N5 (there shouldn;t be), even a heavily binned N32 will have significantly thinner margins than N23 at similar pricing.
- 7500XT vs 6500XT: Same here, double the die space, vram...does not make for good margins, even if we're assuming a $100 increase in MSRP.

Keep in mind that R&D costs in the development period for RDNA3 (~2019-22) is way more expensive than RDNA 2 (~2016-19), on account of the massively increased personnel costs that every single U.S tech company faced during the COVID years as they outbid each other for talent. AMD's SG&A costs jumped from $995m in 2020 to $2.3B in 2022 for example, all these additional costs need to get amortized.



It's been nearly six months since RDNA3 has launched and so far we still don't have much in the way of official information, let alone products for RDNA3 GPUs other than N31, this doesn't strike me as a company that wants to get attention and marketshare. Frankly speaking, I don't think they care about mindshare and would be plenty happy taking those wafers and making datacenter GPUs/CPUs instead..

Seems like N5 is about 30% more expensive than N7 and N6 is supposed to be cheaper than N7.

N32 Vs N21 the cost is massively in favour of N32. N32 Vs N22 and it looks close to me if AMD has similar relative costs to those above.
N32 Vs N23 is in favour of N23 obviously but I don't think the 380 MSRP of the 6600XT is realistic outside of a mining/pandemic induced shortage so AMD expecting that level of margin in that segment long term is unrealistic.

Chances are 7600 will be N33 anyway so that will also be cheaper but AMD could surprise us.

 

[PJVol]

It has half the SEs so L2 is probably half as well and the command front end can be shrunk as well.

It's WGP area 37.5% less and it has 4mb of L2, and btw please keep that preachy tone away, 'cause I didn't ask for help and just shared my thoughts.

 

[Heartbreaker]

If I were AMD I'd find the price point where they can come in a lot lower than NVidia even if they have to sell three times as many cards to make the same profit.

That reminds me of the old joke:
"We lose money on every sale, but make it up on volume"

How much lower do you think they would have to go?

Look at 2070 Super ($500) vs 5700 XT ($400). Similar performance, similar release date, and 2070S outsold 5700XT about 4:1 (according to Steam) with AMD having a 20% lower price...

So how much lower? 30%? 40%? How much margin will they have left?

Also do you think an economic downturn where people are worried about fuel/food/shelter costs are the best time for a "make it up on volume" strategy? A lot of these people are just out of the market until their future looks more solid. They aren't looking for better GPU prices to get back in, they are looking for an economic upturn.

Finally, do you think NVidia won't respond? What happens to the screw the margins strategy when your competitor makes a cut to get back to the same spread as before??

All of the above, is likely why AMD hasn't engaged in this strategy in recent years, and won't this time either.

 

[Kepler_L2]

N31 is 8 WGPs per SE in 2 banks of 4.
N32 would change this to 10WGPs per SE in 2 banks of 5.

I also expect L2 cache to be halved for N32 because it has half the shader engines.

Here is a clearer image.

If you can't figure out how N32 will be arranged from that without me actually sitting at my PC and doing the edit I can't really help you. You can very clearly see there is more than 6 WGPs worth of space available after you slice off the right hand side of the die to leave only 4 MCDs

Edit. Just seen your edit. N32 won't have the same L2 as N31. It has half the SEs so L2 is probably half as well and the command front end can be shrunk as well.

N32 has 4MB L2.

 

[KompuKare]

Keep in mind that R&D costs in the development period for RDNA3 (~2019-22) is way more expensive than RDNA 2 (~2016-19), on account of the massively increased personnel costs that every single U.S tech company faced during the COVID years as they outbid each other for talent. AMD's SG&A costs jumped from $995m in 2020 to $2.3B in 2022 for example, all these additional costs need to get amortized.



It's been nearly six months since RDNA3 has launched and so far we still don't have much in the way of official information, let alone products for RDNA3 GPUs other than N31, this doesn't strike me as a company that wants to get attention and marketshare. Frankly speaking, I don't think they care about mindshare and would be plenty happy taking those wafers and making datacenter GPUs/CPUs instead..

Is that not contractionary?

If the have increased R&D costs which they need to amortise but would prefer the margins of data centre only (i.e. low volume) then the fixed R&D costs would be much higher on each part.

This has been my critique of their "we want high margins at any cost" strategy: high R&D fixed costs spread over far too little volume.

 

[Rigg]

Whether or not it was worthy of the title, there needed to be at least one 7090 card in the stack - simply to imply that they had a 'halo' card. As we've often seen in the video card segment, marketing trumps logic.

Fair point. Naming them 7900 XT & 7800 XT would have at least been consistent with previous gen naming even though they fall short on performance. I think I've probably posted a few variations on what I think they should have been called in different threads so what the hell do I know. When I checked back on this thread last night and reread my post the same thought crossed my mind about not having a 7900 card to market. $1000 7900 XT and $750 7800 XT would have been much better received at launch than $1000 7900 XTX and $900 7900 XT. Honestly, I would have been okay with it.

 

[Ajay]

I'd be surprised if it is. N4 is more likely.

Why do you think that? I know AMD prefers to be a bit less aggressive with Zen CCD chiplets. But, it would seem to be a bad move to do so with their RDNA lineup. After it's misstep with Samsung Logic, Nvidia is back to going full balls on using the best process node available.

 

[maddie]

Great new interview on MLID with semiwiki.com founder. There have been some great interviews lately.

1) He basically scoffs at all the silicon cost estimates floating around and insinuates that they're way too high.

2) Chiplets and the "glue" as sarcastically referred to by some, are extremely difficult + AMD is far ahead of all others.

3) Intel nodes are good ( but TSMC cost + ), Samsung's are not.

I hope some posters here watch it.

 

[jpiniero]

Why do you thing that? I know AMD prefers to be a bit less aggressive with Zen CCD chiplets. But, it would seem to be a bad move to do so with their RDNA lineup. After it's misstep with Samsung Logic, Nvidia is back to going full balls on using the best process node available.

Too early. Best I am expecting from nVidia next year is a Ada Refresh with GDDR7... which could be a solid improvement albeit power hungry.

 

[Saylick]

Great new interview on MLID with semiwiki.com founder. There have been some great interviews lately. He basically scoffs at all the silicon cost estimates floating around and insinuates that they're way too high.

But but but Daddy Jensen told me that they're raising prices because the cost of wafers is skyrocketing...

 

[Ajay]

Too early. Best I am expecting from nVidia next year is a Ada Refresh with GDDR7... which could be a solid improvement albeit power hungry.

I was talking about RDNA 4 being on N3E. Not NV, which is currently on N4 for Ada - IIRC.

 

[jpiniero]

I was talking about RDNA 4 being on N3E.

Yes I'm aware of that. I guess we don't know a realistic timeframe for RDNA 4 but I was guessing it would be sooner rather than later.

 

[Rigg]

Look at 2070 Super ($500) vs 5700 XT ($400). Similar performance, similar release date, and 2070S outsold 5700XT about 4:1 (according to Steam) with AMD having a 20% lower price...

That data encompasses the Pre-built/OEM PC's which represent a huge chunk of this market. This massively skews the data because AMD does not (and probably can not) serve that market in anywhere near the volume that Nvidia does. There is a reason the top of the survey is flooded by the same GPU's that Dell and HP crap out in their high volume sub $1000 gaming desktops. Using the steam survey as a data point to argue anything about DIY GPU sales is inherently flawed.

 

[Heartbreaker]

That data encompasses the Pre-built/OEM PC's which represent a huge chunk of this market. This massively skews the data because AMD does not (and probably can not) serve that market in anywhere near the volume that Nvidia does. There is a reason the top of the survey is flooded by the same GPU's that Dell and HP crap out in their high volume sub $1000 gaming desktops. Using the steam survey as a data point to argue anything about DIY GPU sales is inherently flawed.

What would you suggest using instead of Steam Survey?