Question Speculation: RDNA3 + CDNA2 Architectures Thread

Page 78 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

uzzi38

Platinum Member
Oct 16, 2019
2,565
5,574
146

Leeea

Diamond Member
Apr 3, 2020
3,599
5,340
106
I mean that is very believable so regardless of the source it makes sense. for gamers, a 5800x3d is the better choice overall and with the upcoming if not already happening recession wallets will be tighter and x670 boards and ddr5 aren't exactly cheap.
Their are 100% unconfirmed rumors of unlocked rx5800x3ds floating around:


I am going to speculate that the only way a 5800x3d gets to zen4 clock speeds is if it has some sort of refresh involved.


A new round of cpus for AM4 might be a way for AMD to sell a lot of inventory.
 
Last edited:

H T C

Senior member
Nov 7, 2018
549
395
136
That is not going to happen.

What we saw with Ryzen was as soon as AMD was able to, they charged Intel prices. The GPUs will be no different.

AMD will figure out the point on the supply/demand chart where they are likely to make the most money now, and shoot for that.

Imagine this scenario:

- AMD launch just 2 GPUs, for now
- AMD sells their flagship GPU that trades blows @ least in raster with 4090 (if not higher performance) for ... say ... $1000
- AMD lowers the prices of ALL their current line accordingly, so that every tier has their own pricing

How do you think will nVidia respond to that?

It's not like there hasn't been a precedent for this: remember when Zen launched? 8c / 16t @ CPU around half the price of what Intel priced their premium CPU.
 

Kaluan

Senior member
Jan 4, 2022
500
1,071
96
So, any takers on some last minute RDNA3 SKUs and their specifications speculation?

I wrote down quite a bit in my notepad. Will share some when I get on my PC.
 

GodisanAtheist

Diamond Member
Nov 16, 2006
6,717
7,014
136
Imagine this scenario:

- AMD launch just 2 GPUs, for now
- AMD sells their flagship GPU that trades blows @ least in raster with 4090 (if not higher performance) for ... say ... $1000
- AMD lowers the prices of ALL their current line accordingly, so that every tier has their own pricing

How do you think will nVidia respond to that?

It's not like there hasn't been a precedent for this: remember when Zen launched? 8c / 16t @ CPU around half the price of what Intel priced their premium CPU.

Re: Zen1 AMD was still looking to redeem itself after the Bulldozer arch debacle in a big big way. As AMD's Zen rep increased, so did the prices, until they were at parity with Intel.

AMD is going to price lower than NV, it's just not going to price that much lower.

Let's take a walk down memory lane: 2080ti was $1000 (AIB $1200), Radeon VII was $700 (5700xt was $400). 3090 was $1500, 6900xt was $1000. 4090 is $1600 ($2000+ AIB), 7900xt will be...

More than $1000 and the more competitive it is the more it will cost.
 
  • Like
Reactions: Tlh97 and scineram

H T C

Senior member
Nov 7, 2018
549
395
136
@GodisanAtheist

You're forgetting one small but very important detail: we're in a Worldwide economic downturn. This means price will matter MUCH MORE than before.

If AMD undercuts nVidia in price VS performance in EVERY SINGLE SEGMENT, not by a little but BY A LOT, it's likely that people will buy the cheaper product, specially if it's MUCH CHEAPER.

Like i said, it's a gamble: one that would deal a HUGE blow to AMD if it were to backfire ... but if it succeeds ...


Will AMD "have the stones" to try it, though ...
 
  • Like
Reactions: Kaluan

Timorous

Golden Member
Oct 27, 2008
1,532
2,535
136
Re: Zen1 AMD was still looking to redeem itself after the Bulldozer arch debacle in a big big way. As AMD's Zen rep increased, so did the prices, until they were at parity with Intel.

AMD is going to price lower than NV, it's just not going to price that much lower.

Let's take a walk down memory lane: 2080ti was $1000 (AIB $1200), Radeon VII was $700 (5700xt was $400). 3090 was $1500, 6900xt was $1000. 4090 is $1600 ($2000+ AIB), 7900xt will be...

More than $1000 and the more competitive it is the more it will cost.

I think AMD will price based on their desired margins and their supply capacity.

If they have good supply then they should be able to undercut NV while having better margins than they did with RDNA2 GPUs across the entire stack. As such AMD can win with higher margins, mindshare and marketshare and consumers can win with better perf/$ than existing parts.

I think this is the path AMD will take but lets wait and see.
 
  • Like
Reactions: Joe NYC and Tlh97

Gideon

Golden Member
Nov 27, 2007
1,608
3,573
136
AMD is certainly pricing the last gen quite well already:


Let's wait and see.

I don't see the top SKU anywhere near $1000. It's final price is anyones guess (and depends heavily on performance), but I think it will be the halo product to bring up the margins.
What I really do hope, is that the lower-end NAVI 31 SKUs (the equivalent of current 6800 and 6800XT) will be sold around the 1000€ mark and slightly below.

History doesn't look kindly to such fools-hopes though.
 
Last edited:
  • Like
Reactions: Tlh97 and scineram

Timorous

Golden Member
Oct 27, 2008
1,532
2,535
136
AMD is certainly pricing the last gen quite well already:


Let's wait and see.

I don't see the top SKU anywhere near $1000. It's final price is anyones guess (and depends heavily on performance), but I think it will be the halo product to bring up the margins.
What I really do hope, is that the lower-end NAVI 31 SKUs (the equivalent of current 6800 and 6800XT) will be sold around the 1000€ mark and slightly below.

History doesn't look kindly to such fools-hopes though.

I don't think N31 will be x8 tier, that is where N32 comes in.

N31 can easily have 3 SKUs, The 3d cache one, the non 3d cache one and the 5MCD one.

I can see the 5MCD one being in the $1,000 region give or take. I could see the non 3d cache version being $1,200 give or take and the 3d cache version they could just go big with it knowing it will be low volume but some people will have to have it because it is the best or because it can do 8k better or something.
 

Kaluan

Senior member
Jan 4, 2022
500
1,071
96
I don't think N31 will be x8 tier, that is where N32 comes in.

N31 can easily have 3 SKUs, The 3d cache one, the non 3d cache one and the 5MCD one.

I can see the 5MCD one being in the $1,000 region give or take. I could see the non 3d cache version being $1,200 give or take and the 3d cache version they could just go big with it knowing it will be low volume but some people will have to have it because it is the best or because it can do 8k better or something.
If nVidia can have a die/SKU that is "only" 2/3rds of it's bigger sibling and call it xx8 class SKU (it was even worse before the 12GB cancellation), then so can AMD. I have almost no doubt that at least some 7800 SKUs will be based on N32, not N31. Their die tiering are much more balanced than ever this gen, I see them cashing in on that hard.

AMD will likely market N31+V-Cache in 4K and 8K scenarios vs the rest of the stack and the competition hard. 😅
If it outclasses the 4090, I could see them ask $1400 for it. While the regular N31 could stay at $1000 for maximum pressure on nV.

Anyone have a guesstimate of how many Texture Mapping Units and Raster Output Units the N3x chips will have? Something tells me that on full N31 just 386 TMUs and 160 ROPs is unlikely, but so is 768 and 320 IDK. AD102 has 576 and 192 I believe. Considerably up from 320/336 and 128/112 on N21/GA102.
Or if the 2nd gen Ray Accelerators will stay at 2 per WGP or be 4 per WGP?
 
Last edited:

TESKATLIPOKA

Platinum Member
May 1, 2020
2,329
2,811
106
Anyone have a guesstimate of how many Texture Mapping Units and Raster Output Units the N3x chips will have? Something tells me that on full N31 just 386 TMUs and 160 ROPs is unlikely, but so is 768 and 320 IDK. AD102 has 576 and 192 I believe. Considerably up from 320/336 and 128/112 on N21/GA102.
Or if the 2nd gen Ray Accelerators will stay at 2 per WGP or be 4 per WGP?
Hard to tell. Angstronomics didn't mention this info, only WGPs and shader engines.
I think 192 ROPs is realistic.
Ray accelerator I think will be doubled per WGP considering the massive shader increase. TMUs will be 12-16(+50-100%) per WGP in my opinion.
 
Last edited:
  • Like
Reactions: Tlh97 and Kaluan

Mopetar

Diamond Member
Jan 31, 2011
7,797
5,899
136
If nVidia can have a die/SKU that is "only" 2/3rds of it's bigger sibling and call it xx8 class SKU (it was even worse before the 12GB cancellation), then so can AMD.

The modular nature of RDNA3 makes it easier for AMD to reuse a chip across multiple product lines and differentiate those without wasting silicon to the same degree as a monolithic chip being stretched across multiple products would be. I don't think we'll see N32 getting branded as a 7800. If AMD were considering it, you'd think they would have learned from NVidia's mistake and quietly scrapped those plans.

Another interesting thought is that with creating separate memory chiplets, in addition to having the option of a high-end product with v-cache, AMD also has the option of swapping to a chiplet that interfaces with HBM. In the past it was never worth the die space to support both memory types, but you might already be producing such chiplets for professional or data center products and there's no reason those couldn't be used for a limited edition consumer product that just acts as a halo product even if it's far more expensive.

Chiplets make for a far more flexible product lineup and while AMD may not yet be in a position to execute on all of those possibilities, it's certainly something that they can move towards. Even if they're not creating a mass market product, a top end card with HBM and v-cache will generate a lot of hype and publicity just by existing. AMD could even position it as a prosumer product similar to what NVidia did with their Titan cards for several years. It doesn't matter if they have to charge $2,500 for it when it occupies the top spot on any chart that someone looks at when a new GPU is being reviewed. That likely does more to change AMD's perception in the market than trying to offer similar performance at a much lower price.
 

jpiniero

Lifer
Oct 1, 2010
14,510
5,159
136
The modular nature of RDNA3 makes it easier for AMD to reuse a chip across multiple product lines and differentiate those without wasting silicon to the same degree as a monolithic chip being stretched across multiple products would be. I don't think we'll see N32 getting branded as a 7800. If AMD were considering it, you'd think they would have learned from NVidia's mistake and quietly scrapped those plans.

You're not going to get 100% yield even at TSMC. So you need a place to dump partially busted dies.You should only need one cut product per die but you do need one.

Making N31 be 79xx , N32 78xx and N33 77xx makes sense.
 

Timorous

Golden Member
Oct 27, 2008
1,532
2,535
136
You're not going to get 100% yield even at TSMC. So you need a place to dump partially busted dies.You should only need one cut product per die but you do need one.

Making N31 be 79xx , N32 78xx and N33 77xx makes sense.

Except N33 is 128bit with 8GB ram so it won't be a 7700.

N32 with 4 MCDs will be 7800XT and cut N32 with 3 MCDs will be 7700XT. There may not be vanilla variants, atleast at N32 launch.

I don't see the issue of a 256bit 16GB card with ~50% more shaders than 6800XT being the 7800XT.
 

maddie

Diamond Member
Jul 18, 2010
4,723
4,628
136
You're not going to get 100% yield even at TSMC. So you need a place to dump partially busted dies.You should only need one cut product per die but you do need one.

Making N31 be 79xx , N32 78xx and N33 77xx makes sense.
These are not monolithic. You have defects in the core and defects in the MCD bonding, and the two cases will not always align.. So no, there are at least 2 cuts needed, to maximize die utilization.
 
  • Like
Reactions: Leeea

eek2121

Platinum Member
Aug 2, 2005
2,904
3,906
136
AMD is certainly pricing the last gen quite well already:


Let's wait and see.

I don't see the top SKU anywhere near $1000. It's final price is anyones guess (and depends heavily on performance), but I think it will be the halo product to bring up the margins.
What I really do hope, is that the lower-end NAVI 31 SKUs (the equivalent of current 6800 and 6800XT) will be sold around the 1000€ mark and slightly below.

History doesn't look kindly to such fools-hopes though.

Top end will likely be $1,200-1,300 unless it doesn’t compete.
 

maddie

Diamond Member
Jul 18, 2010
4,723
4,628
136
Top end will likely be $1,200-1,300 unless it doesn’t compete.
For all of 1 month maybe.This should add some context.

Apple has told at least one manufacturer in China to immediately halt production of iPhone 14 Plus components while its procurement team reevaluates demand for the product, which Apple has positioned as a cheaper alternative to its more expensive iPhone Pro models but equipped with a large screen, according to one of the people.

Two downstream Apple suppliers in China that rely on the parts and assemble them into larger modules are also cutting their production 70% and 90% respectively, the person said.

 

linkgoron

Platinum Member
Mar 9, 2005
2,286
810
136
For all of 1 month maybe.This should add some context.

Apple has told at least one manufacturer in China to immediately halt production of iPhone 14 Plus components while its procurement team reevaluates demand for the product, which Apple has positioned as a cheaper alternative to its more expensive iPhone Pro models but equipped with a large screen, according to one of the people.

Two downstream Apple suppliers in China that rely on the parts and assemble them into larger modules are also cutting their production 70% and 90% respectively, the person said.

Off topic, but Apple should be ashamed of themselves for even releasing the 14 and 14 plus. Good thing that consumers are not falling for it.
 

jpiniero

Lifer
Oct 1, 2010
14,510
5,159
136
These are not monolithic. You have defects in the core and defects in the MCD bonding, and the two cases will not always align.. So no, there are at least 2 cuts needed, to maximize die utilization.

That's probably rare enough that you could just make a small SKU for mobile if that ends up being big enough.

Except N33 is 128bit with 8GB ram so it won't be a 7700.

If people are going to get hung up over memory capacity then maybe by the time it shows up on desktop 3 or 4 GB chips will be available. If they (revert?) back to a schedule like RDNA 2 then it will be August or September of next year before you will see N33 desktop. If the laptop version sells poorly then maybe you will see it sooner.

The 4 lanes is maybe not an issue with laptops since they will be using PCIe 5 but more ram would be helpful on desktop 4.0 or 3.0 systems I think.
 

Kaluan

Senior member
Jan 4, 2022
500
1,071
96
Hard to tell. Angstronomics didn't mention this info, only WGPs and shader engines.
I think 192 ROPs is realistic.
Ray accelerator I think will be doubled per WGP considering the massive shader increase. TMUs will be 12-16(+50-100%) per WGP in my opinion.
Interesting, thank you. Figured as well that it doesn't sound very intuitive to increase shaders by 2,4x, but boost the ray hardware by just 1,2x even if they are improved and you can clock much higher.

I'll napkin math some stuff and see what makes most sense for the texturing and raster units.
The 4 lanes is maybe not an issue with laptops since they will be using PCIe 5 but more ram would be helpful on desktop 4.0 or 3.0 systems I think.
Hm, wasn't N33 revealed to be 8x 5.0? 8 lanes is generous enough, especially if it helped them shave off a crucial few extra mm/2 off the die and reduce tracings/connectors costs lol
Only way I'd see it start degrading performance is on a 3.0 system in a game that overfills the 8GB (so mostly just 10th gen or older + native 4K... not the most common combo for a budget builder nowadays)
 

maddie

Diamond Member
Jul 18, 2010
4,723
4,628
136
That's probably rare enough that you could just make a small SKU for mobile if that ends up being big enough.



If people are going to get hung up over memory capacity then maybe by the time it shows up on desktop 3 or 4 GB chips will be available. If they (revert?) back to a schedule like RDNA 2 then it will be August or September of next year before you will see N33 desktop. If the laptop version sells poorly then maybe you will see it sooner.

The 4 lanes is maybe not an issue with laptops since they will be using PCIe 5 but more ram would be helpful on desktop 4.0 or 3.0 systems I think.
I have read this a few times and am still confused. Why are you writing about 4 lanes in response to N33, 128 bit and 8GB?
 

Gideon

Golden Member
Nov 27, 2007
1,608
3,573
136
Interesting, thank you. Figured as well that it doesn't sound very intuitive to increase shaders by 2,4x, but boost the ray hardware by just 1,2x even if they are improved and you can clock much higher.

I agree, it would be a mistake not doubling the unit count per WEU.

Raytracing is moving away from being a gimmic and working itself into core features of game-engines, like replacing rasterized lightning entirely (in Metro series) or speeding up more advanced lightning techniques (UE5 Lumen). Every profile title seems to have more and more of it.

And yes, in games you can buy today, raytracing is largely still a gimmic, but let's not forget how long the lifecycle of GPUs . Let's say someone wants to by "the best AMD has to offer" in September 2024. Even then, the only option would still be a N31 variant (probably a mid-cycle refresh), and that person almost certainly would want at least 2-3 years out of that purchase. Just adding 20% more Ray Accelerators (even when slightly tuned) would not be enough for that timeline. AMD already sorta is the laggard in the group (as even Intel has 2 ray traversal pipelines per Xe-core with can each do more).

What I would want for AMD to do is:
  1. Double the Ray Accelerator count per WEU
  2. Improve the units themselves. AMD's RT units currently do 4 Ray/Box and 1 Ray/Triangle intersections, while Intel does 12 and 1 (with 2x the "RT units" per "core"). As there is more and more geometry in games, I'd at least like to see the Ray/Triangle rate double to 2 per clock in at least most instances (Ray/Box rate IMO is less of an issue, but could also be increased slightly)
  3. Add some fixed-function hardware to speed up BVH traversal, like the competitors have (freeing up some CU resources)
I really hope we at least get point 1 or 2. Both would already be very good, with 3 as a bonus (but quite unlikely considering the die-sizes)
 
Last edited:

Timorous

Golden Member
Oct 27, 2008
1,532
2,535
136
That's probably rare enough that you could just make a small SKU for mobile if that ends up being big enough.



If people are going to get hung up over memory capacity then maybe by the time it shows up on desktop 3 or 4 GB chips will be available. If they (revert?) back to a schedule like RDNA 2 then it will be August or September of next year before you will see N33 desktop. If the laptop version sells poorly then maybe you will see it sooner.

The 4 lanes is maybe not an issue with laptops since they will be using PCIe 5 but more ram would be helpful on desktop 4.0 or 3.0 systems I think.

N33 is drop in compatible with N23 so will be 8 lanes.

N33 is going to be a cheap cheap part and with the spec it is perfect for 1080p and will be okay at 1440p. For that 8GB is all that is needed but it will be a 7600 tier part not 7700 tier.

N32 will fill in both x800 and x700 tier simply because that makes sense, 200mm of N5 is not a fat lot and then having the option of using 4 or 3 MCDs. Some of which will be forced if one of the IO links to the MCDs is defective. BOM wise it will be far cheaper than the N21 BOMs and it will be comparable to N22 BOMs.

Another way to look at is is a you can get about 107 N21 dies from a full wafer and 177 N22 dies from a full wafer, some of which will be defective (calculator I used did not include defect rate) for a total of 284 dies across 2 wafers.

With N32 you can have 295 dies from 1 wafer and 1719 MCDs from 1 wafer. That means from 2 wafers you can build 107 N32 based 7800XTs + 177 N32 based 7700XTs + have 11 N32 dies left and 760MCDs left.

Cost wise I believe 2 N7 wafers is a bit more than 1 N5 wafer, not sure if it is more than 1 N5 + 1 N6 wafer, but you are only really using a bit more than half the MCDs off of the N6 with that ratio so really 2 N5 wafers + 1 N6 wafer can probably build all the N32 cards that are required if the skew more towards 7700XTs. That is 3 wafers building nearly 600 GPUs at just under 200 GPUs per wafer which is better than N22 in terms of wafer efficiency and far far better than N21 for x800 series.

Also because 2 of the more popular segments are built using the same GCD it allows AMD much more flexibility it meeting needs depending on what product has more demand vs their forecasts.
 

biostud

Lifer
Feb 27, 2003
18,193
4,674
136
N33 is drop in compatible with N23 so will be 8 lanes.

N33 is going to be a cheap cheap part and with the spec it is perfect for 1080p and will be okay at 1440p. For that 8GB is all that is needed but it will be a 7600 tier part not 7700 tier.

N32 will fill in both x800 and x700 tier simply because that makes sense, 200mm of N5 is not a fat lot and then having the option of using 4 or 3 MCDs. Some of which will be forced if one of the IO links to the MCDs is defective. BOM wise it will be far cheaper than the N21 BOMs and it will be comparable to N22 BOMs.

Another way to look at is is a you can get about 107 N21 dies from a full wafer and 177 N22 dies from a full wafer, some of which will be defective (calculator I used did not include defect rate) for a total of 284 dies across 2 wafers.

With N32 you can have 295 dies from 1 wafer and 1719 MCDs from 1 wafer. That means from 2 wafers you can build 107 N32 based 7800XTs + 177 N32 based 7700XTs + have 11 N32 dies left and 760MCDs left.

Cost wise I believe 2 N7 wafers is a bit more than 1 N5 wafer, not sure if it is more than 1 N5 + 1 N6 wafer, but you are only really using a bit more than half the MCDs off of the N6 with that ratio so really 2 N5 wafers + 1 N6 wafer can probably build all the N32 cards that are required if the skew more towards 7700XTs. That is 3 wafers building nearly 600 GPUs at just under 200 GPUs per wafer which is better than N22 in terms of wafer efficiency and far far better than N21 for x800 series.

Also because 2 of the more popular segments are built using the same GCD it allows AMD much more flexibility it meeting needs depending on what product has more demand vs their forecasts.
As I posted earlier in the thread I think this:

RX 7500(N33) 6GB
RX 7600(N33) 8GB
RX 7700(N32) 12GB
RX 7800(N32) 16GB
RX 7900(N31) 20GB
RX 7950(N31) 24GB

Maybe a 3D cache version of both top N31 and N32.