Question Speculation: RDNA3 + CDNA2 Architectures Thread

Page 158 - 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
5,573
146

GodisanAtheist

Diamond Member
Nov 16, 2006
6,717
7,010
136
Did RDNA 2 slow down?

It just seems less competitive now vs Nvidia.

relative-performance_3840-2160.png

-Testbed changed for the new review. NV's drivers have actually become "heavier" than AMD's for this last cycle (likely earlier but AMD was just uncompetitive from a hardware standpoint in the Navi 10 and GCN4 days) so moving from a 5800x to a 13000 series Intel is likely allowing NV to stretch it's legs a bit more.

Would be WONDERFUL if some reviewer tested more cards with more logical mid-range and prior gen setups so people running 8700k or 1600x processors with dumpster tier DDR4 memory got a more realistic picture of how new cards perform for them but... hey I'm not the one putting in the work so how much can I complain.
 

GodisanAtheist

Diamond Member
Nov 16, 2006
6,717
7,010
136
In a universe where N32 is a 6900/6950xt competitor, N33 with half the CUs, half the bandwidth, and on 6nm is sure not going to be 6750/6800 speed.

- I think Chiplets are putting a lid on clock scaling, which is a problem N33 won't have as a monolithic chip.

N33 with a 17% IPC increase + a 15% clock speed bump (~2700mhz game clock) should bring a 6600xt-esq part up to ~6750XT levels, at least at lower resolutions (6750 is ~25% faster and 6800 is ~45% faster).
 

TESKATLIPOKA

Platinum Member
May 1, 2020
2,329
2,811
106
- I think Chiplets are putting a lid on clock scaling, which is a problem N33 won't have as a monolithic chip.

N33 with a 17% IPC increase + a 15% clock speed bump (~2700mhz game clock) should bring a 6600xt-esq part up to ~6750XT levels, at least at lower resolutions (6750 is ~25% faster and 6800 is ~45% faster).
100*1.17*1.15 = 134.5
IPC increase is already questionable. If that 2700MHz is game clock is also questionable.
In 3Dmak It should be around desktop N22 based on what I saw. Will It really happen? We will see after CES 2023.
At least I expect they will announce N33 there.
 
Last edited:

insertcarehere

Senior member
Jan 17, 2013
639
607
136
- I think Chiplets are putting a lid on clock scaling, which is a problem N33 won't have as a monolithic chip.

N33 with a 17% IPC increase + a 15% clock speed bump (~2700mhz game clock) should bring a 6600xt-esq part up to ~6750XT levels, at least at lower resolutions (6750 is ~25% faster and 6800 is ~45% faster).

IPC increases and clock speed bumps aren't "free" in terms of Silicon transistors and die size.

Your performance increases would be reasonable here if Navi 33 turns out to be significantly larger than Navi 23...but as the former is supposedly ~200mm^2 on TSMcN6 that wouldn't be the case, there shouldnt be appreciably more transistors on Navi 33 vs Navi 23.

A 6750XT-matching Navi 33 implies AMD improving PPTransistor/PPA iso-process by 30+% from Navi 23 to Navi 33. This would be unprecedented (5700xt -> 6600XT does not get near this) and be so far removed from what we're seeing with Navi 31 that calling them both RDNA3 would be very misleading.
 

biostud

Lifer
Feb 27, 2003
18,193
4,674
136
- I think Chiplets are putting a lid on clock scaling, which is a problem N33 won't have as a monolithic chip.
Only if you don't design it properly. As I've compared it to before, the GTX 480 which had had massive heat/clock problems because of a bad design, which got fixed with the GTX 580. Obviously it is not the same, but maybe similar. In compute only tasks some have reported very high frequencies, so the chip is capable of reaching these clocks under the right (or in this case wrong) conditions. Why it is not possible in gaming is the great question.
 

TESKATLIPOKA

Platinum Member
May 1, 2020
2,329
2,811
106
IPC increases and clock speed bumps aren't "free" in terms of Silicon transistors and die size.

Your performance increases would be reasonable here if Navi 33 turns out to be significantly larger than Navi 23...but as the former is supposedly ~200mm^2 on TSMcN6 that wouldn't be the case, there shouldnt be appreciably more transistors on Navi 33 vs Navi 23.

A 6750XT-matching Navi 33 implies AMD improving PPTransistor/PPA iso-process by 30+% from Navi 23 to Navi 33. This would be unprecedented (5700xt -> 6600XT does not get near this) and be so far removed from what we're seeing with Navi 31 that calling them both RDNA3 would be very misleading.
The IPC(PPC) will come mainly from dual issue. You will need higher clocks to get ~30% higher performance to be on par with N22.

In some games RDNA3 has a very nice improvement in IPC(PPC).
Screenshot_22.png
If they can improve performance in other games, then it shouldn't be that bad.

If you shrink N23 to N6, then It would be ~200mm2.
Your first sentence is true, but just because die size would be comparable, doesn't mean that transistor density or transistors per die would be the same.
P%C5%99edstaven%C3%AD-GPU-architektury-AMD-RDNA-3-08.png

The RDNA 3 architecture is apparently designed to achieve very high transistor density, with 165 % more transistors per square millimetre than RDNA 2 on the 7nm process. This is a much larger increase than the transition from 7nm to 5nm process alone would allow. In the RDNA 2 architecture, a CU pair had 215 million transistors in a 4.33 mm² area, while in RDNA 3 a CU pair has 331 million transistors in just 2.50 mm².
If we say from those +165% -> 75% come from the better process, then 100*1.75=175 => 265/175=1.51.
RDNA3 should have ~50% higher density than RDNA2 even on the same process, at least for WGPs.
 
Last edited:
  • Like
Reactions: Kaluan

Heartbreaker

Diamond Member
Apr 3, 2006
4,222
5,224
136
The IPC(PPC) will come mainly from dual issue. You will need higher clocks to get ~30% higher performance to be on par with N22.

In some games RDNA3 has a very nice improvement in IPC(PPC).
View attachment 73458
If they can improve performance in other games, then it shouldn't be that bad.

If you shrink N23 to N6, then It would be ~200mm2.
Your first sentence is true, but just because die size would be comparable, doesn't mean that transistor density or transistors per die would be the same.
P%C5%99edstaven%C3%AD-GPU-architektury-AMD-RDNA-3-08.png


If we say from those +165% -> 75% come from the better process, then 100*1.75=175 => 265/175=1.51.
RDNA3 should have ~50% higher density than RDNA2 even on the same process, at least for WGPs.

Transistor density increase is from removing the memory controllers (analog) and SRAM cache from the main die, neither of which scale very well and are low density.
 
  • Like
Reactions: Lodix

TESKATLIPOKA

Platinum Member
May 1, 2020
2,329
2,811
106
Transistor density increase is from removing the memory controllers (analog) and SRAM cache from the main die, neither of which scale very well and are low density.
If you had read what I posted, then you would know that AMD made a significant improvement in transistor density, which has nothing to do with what you wrote, and It's also not just because of the better process.
 
Last edited:
  • Like
Reactions: Kaluan

Timorous

Golden Member
Oct 27, 2008
1,532
2,535
136
IPC increases and clock speed bumps aren't "free" in terms of Silicon transistors and die size.

Your performance increases would be reasonable here if Navi 33 turns out to be significantly larger than Navi 23...but as the former is supposedly ~200mm^2 on TSMcN6 that wouldn't be the case, there shouldnt be appreciably more transistors on Navi 33 vs Navi 23.

A 6750XT-matching Navi 33 implies AMD improving PPTransistor/PPA iso-process by 30+% from Navi 23 to Navi 33. This would be unprecedented (5700xt -> 6600XT does not get near this) and be so far removed from what we're seeing with Navi 31 that calling them both RDNA3 would be very misleading.

Vega 20 -> Navi 10

5700XT is about the same perf as Radeon 7 but manages that with 75% of the transistors and 75% of the die size on the same N7 node.
 
  • Like
Reactions: Kaluan

Glo.

Diamond Member
Apr 25, 2015
5,658
4,416
136
Vega 20 -> Navi 10

5700XT is about the same perf as Radeon 7 but manages that with 75% of the transistors and 75% of the die size on the same N7 node.
And RDNA1 architectural redesign was the same type of redesign for GCN architecture that RDNA3 is for RDNA1 and 2. Its improved on architectural level. Alus should do more, with less resources.
 

Kaluan

Senior member
Jan 4, 2022
500
1,071
96
Transistor density increase is from removing the memory controllers (analog) and SRAM cache from the main die, neither of which scale very well and are low density.
Uhh... nope.

Neither Infinty Cache MALL or memory controllers are inside the WGPs in any RDNA2 design, no one does GPU designs like that.

It's like saying Matisse/Vermeer/Raphael's L3 and MC are inside the Zen cores.

Fjza5VjWACMxCpy.jpeg.jpg
Credit: Fritzchens Fritz/Locuza
 
  • Like
Reactions: Lodix

Heartbreaker

Diamond Member
Apr 3, 2006
4,222
5,224
136
Uhh... nope.

Neither Infinty Cache MALL or memory controllers are inside the WGPs in any RDNA2 design, no one does GPU designs like that.

It's like saying Matisse/Vermeer/Raphael's L3 and MC are inside the Zen cores.

View attachment 73487
Credit: Fritzchens Fritz/Locuza

They are measuring transistor density for the WHOLE die, which includes memory controllers and a big cache, which are low density. Remove those from the die, and the die transistor density increases dramatically.

When we get down to the lower tier dies that don't have chiplets, I guarantee you are going to see the transitor density fall significantly.
 

Kaluan

Senior member
Jan 4, 2022
500
1,071
96
They are measuring transistor density for the WHOLE die, which includes memory controllers and a big cache, which are low density. Remove those from the die, and the die transistor density increases dramatically.

When we get down to the lower tier dies that don't have chiplets, I guarantee you are going to see the transitor density fall significantly.
Straw man.

Who's "they"? No one is doing that, not AMD and not the people who brought it up.
No one was arguing N33 die won't be less dense than N31. You were replying to a post about the much smaller and much denser WGPs of RDNA3 saying that they're only smaller because of chiplets (and node shrink). That's false, not to mention technically wrong.

RDNA3 is obviously not just a die shrink and "chiplet-ization" of RDNA2 lol
 
  • Like
Reactions: TESKATLIPOKA

TESKATLIPOKA

Platinum Member
May 1, 2020
2,329
2,811
106
They are measuring transistor density for the WHOLE die, which includes memory controllers and a big cache, which are low density. Remove those from the die, and the die transistor density increases dramatically.

When we get down to the lower tier dies that don't have chiplets, I guarantee you are going to see the transitor density fall significantly.
What's your point exactly? N33 will have better density than N23 and not just because of using N6 instead of N7, that's evident from what I posted above.

You don't have to guarantee anything. It's clear N33 will have worse density than N31 or N32, because one is using N6 and the others N5 process. Removal of cache and memory will just increase the difference, but this is not what we originally discussed.
What was originally discussed was If N33 with a smaller die size can perform as N22, then for some reason you brought up chiplets.
 
Last edited:

insertcarehere

Senior member
Jan 17, 2013
639
607
136
Vega 20 -> Navi 10

5700XT is about the same perf as Radeon 7 but manages that with 75% of the transistors and 75% of the die size on the same N7 node.

Yes it did, Vega 20 was also a significantly more compute-focused GPU than RDNA1 (2:1 FP64, HBM2) and was well.. Vega, improving by similar magnitudes with RDNA 2 as a baseline is a very different story.

If you shrink N23 to N6, then It would be ~200mm2.
Your first sentence is true, but just because die size would be comparable, doesn't mean that transistor density or transistors per die would be the same.
P%C5%99edstaven%C3%AD-GPU-architektury-AMD-RDNA-3-08.png


If we say from those +165% -> 75% come from the better process, then 100*1.75=175 => 265/175=1.51.
RDNA3 should have ~50% higher density than RDNA2 even on the same process, at least for WGPs.

The discussion on transistor density is a detour from the supposed performance for N33. If they're able to make ~200mm^2 N33 on N6 perform like ~330mm^2 N22 on N7, then something's not adding up with AMD's decisions for N31/N32:
Capture6.JPG

1672224194735.png

Using Computerbase data for consistency, the difference between Navi 23 to Navi 22 is roughly similar to difference between Navi 21 to Navi 31. Therefore if we assume Navi 33 performs like Navi 22, a ~500-550mm^2 Navi 21-sized RDNA3 (monolithic or split up in chiplets) should perform similar to what Navi 31 does now, all on N6. Instead what we actually got was ~300mm^2 of TSMC N5 in GCD alone and 200+mm^2 of MCDs in N6 for this level of performance.

Now, AMD aren't gonna use more expensive TSMC N5 wafers for the hell of it when they can use N6 and auto-win on BOM instead so it's difficult to reconcile what we see with N31 and how that relates to performance discussions for N33.
 

Attachments

  • Capture5.JPG
    Capture5.JPG
    33.3 KB · Views: 9
Last edited:
  • Like
Reactions: Lodix

Timorous

Golden Member
Oct 27, 2008
1,532
2,535
136
Yes it did, Vega 20 was also a significantly more compute-focused GPU than RDNA1 (2:1 FP64, HBM2) and was well.. Vega, improving by similar magnitudes with RDNA 2 as a baseline is a very different story.

You said it was unprecedented. Obviously it is not. Without checking numbers Maxwell and or Pascal might be candidates as well.


The discussion on transistor density is a detour from the supposed performance for N33. If they're able to make ~200mm^2 N33 on N6 perform like ~330mm^2 N22 on N7, then something's not adding up with AMD's decisions for N31/N32:
View attachment 73527

View attachment 73528

Using Computerbase data for consistency, the difference between Navi 23 to Navi 22 is roughly similar to difference between Navi 21 to Navi 31. Therefore if we assume Navi 33 performs like Navi 22, a ~500-550mm^2 Navi 21-sized RDNA3 (monolithic or split up in chiplets) should perform similar to what Navi 31 does now, all on N6. Instead what we actually got was ~300mm^2 of TSMC N5 in GCD alone and 200+mm^2 of MCDs in N6 for this level of performance.

Now, AMD aren't gonna use more expensive TSMC N5 wafers for the hell of it when they can use N6 and auto-win on BOM instead so it's difficult to reconcile what we see with N31 and how that relates to performance discussions for N33.

It will depend on N33 clockspeeds. Running N31 at 3.2 Ghz saw a 14% increase in performance over reference in CP2077.

Reference 7900XTX operates at around 2.6Ghz on average. Reference 6950XT is around 2.4Ghz. this is an 8% increase.

If N33 sees the same 8% clockspeed increase with a 17% IPC increase then it should be about 26% faster than the 6650XT or just a bit behind 6700XT.

If they have fixed whatever is causing N31 problems with high clocks at reasonable power consumption in games then it could have clocks as much as 33% higher. This would mean 55% more performance than the 6650XT which is about on par with the 6800 (162/103 since you set 6600XT as 100% not 6650XT).

These are my lower and upper bounds, quite a wide spread tbh.

On average it will probably be 6700XT to 6750XT but some games might trend lower if ILP is not extracted and where the driver can extract it 6800 performance might appear.
 
  • Like
Reactions: Lodix and Mopetar

TESKATLIPOKA

Platinum Member
May 1, 2020
2,329
2,811
106
Yes it did, Vega 20 was also a significantly more compute-focused GPU than RDNA1 (2:1 FP64, HBM2) and was well.. Vega, improving by similar magnitudes with RDNA 2 as a baseline is a very different story.



The discussion on transistor density is a detour from the supposed performance for N33. If they're able to make ~200mm^2 N33 on N6 perform like ~330mm^2 N22 on N7, then something's not adding up with AMD's decisions for N31/N32:
View attachment 73527

View attachment 73528

Using Computerbase data for consistency, the difference between Navi 23 to Navi 22 is roughly similar to difference between Navi 21 to Navi 31. Therefore if we assume Navi 33 performs like Navi 22, a ~500-550mm^2 Navi 21-sized RDNA3 (monolithic or split up in chiplets) should perform similar to what Navi 31 does now, all on N6. Instead what we actually got was ~300mm^2 of TSMC N5 in GCD alone and 200+mm^2 of MCDs in N6 for this level of performance.

Now, AMD aren't gonna use more expensive TSMC N5 wafers for the hell of it when they can use N6 and auto-win on BOM instead so it's difficult to reconcile what we see with N31 and how that relates to performance discussions for N33.
N31 is ~525mm2(300mm2+225mm2) in total. N31 with the same specs but being a monolith would be ~450mm2 or maybe a bit less.
I don't understand why you think a 500-550mm2 5nm RDNA3 would perform like N31.

BTW, It's not like this is the final performance of N31, drivers will improve performance who knows by how much, but It's not a secret dual issue doesn't do much in many games and in some you see >20% gain.
Screenshot_22.png

I don't think anyone expects the same performance as N22 at 4K. I myself was always thinking about 1080p or maybe even 1440p and at those resolutions the difference is less than 141/103=1.37 or 37%.

If N33 can perform 30% faster than N23 will depend on clocks and IPC gain.
 
Last edited:
  • Like
Reactions: Kaluan

TESKATLIPOKA

Platinum Member
May 1, 2020
2,329
2,811
106
It will depend on N33 clockspeeds. Running N31 at 3.2 Ghz saw a 14% increase in performance over reference in CP2077.

Reference 7900XTX operates at around 2.6Ghz on average. Reference 6950XT is around 2.4Ghz. this is an 8% increase.

If N33 sees the same 8% clockspeed increase with a 17% IPC increase then it should be about 26% faster than the 6650XT or just a bit behind 6700XT.

If they have fixed whatever is causing N31 problems with high clocks at reasonable power consumption in games then it could have clocks as much as 33% higher. This would mean 55% more performance than the 6650XT which is about on par with the 6800 (162/103 since you set 6600XT as 100% not 6650XT).

These are my lower and upper bounds, quite a wide spread tbh.

On average it will probably be 6700XT to 6750XT but some games might trend lower if ILP is not extracted and where the driver can extract it 6800 performance might appear.
Only 14% more performance with 23% higher clock speed is only 61% scaling.
If performance doesn't scale linearly with clock speed then N33 wouldn't be 26% faster with 17% better IPC and 8% higher clock speed, It would be 23% faster.
With 33% higher clock speed, It would be 100*1.17*[(100+33*0.61)/100] = 140.5% or 40.4% faster.
Then let's not forget about BW being a bottleneck, at least in this second case.
 

insertcarehere

Senior member
Jan 17, 2013
639
607
136
N31 is ~525mm2(300mm2+225mm2) in total. N31 with the same specs but being a monolith would be ~450mm2 or maybe a bit less.
I don't understand why you think a 500-550mm2 5nm RDNA3 would perform like N31.

My point there was that, assuming Navi 33 performs similar to N22 like some posters here claim, it'd be reasonable to expect that the architecture should scale such that a N21-sized RDNA3 GPU (~500-550mm^2) should perform like N31 does now, all on N6 (~7nm node). But as it is AMD requires 5nm with similar area to reach that level of performance.

BTW, It's not like this is the final performance of N31, drivers will improve performance who knows by how much, but It's not a secret dual issue doesn't do much in many games and in some you see >20% gain.

Given your comparison is relative gains between RDNA2 vs RDNA3, this requires:
- Future AMD drivers constantly improve RDNA 3 performance by a bigger margin than RDNA 2 performance.
- Future Games are developed such that they generally advantage of AMD's implementation of dual-issue (Nvidia also implements 2x FP32, albeit somewhat differently AFAIK)

Those possibilities can't be dismissed but are very far from certain...

I don't think anyone expects the same performance as N22 at 4K. I myself was always thinking about 1080p or maybe even 1440p and at those resolutions the difference is less than 141/103=1.37 or 37%.

That difference was at 1440p, which should be plenty fair resolution target for a "midrange" card circa 2023.

If N33 can perform 30% faster than N23 will depend on clocks and IPC gain.

I agree.
 

TESKATLIPOKA

Platinum Member
May 1, 2020
2,329
2,811
106
My point there was that, assuming Navi 33 performs similar to N22 like some posters here claim, it'd be reasonable to expect that the architecture should scale such that a N21-sized RDNA3 GPU (~500-550mm^2) should perform like N31 does now, all on N6 (~7nm node). But as it is AMD requires 5nm with similar area to reach that level of performance.
Sorry, didn't notice you were talking about N6 and not N5 monolith.

Given your comparison is relative gains between RDNA2 vs RDNA3, this requires:
- Future AMD drivers constantly improve RDNA 3 performance by a bigger margin than RDNA 2 performance.
- Future Games are developed such that they generally advantage of AMD's implementation of dual-issue (Nvidia also implements 2x FP32, albeit somewhat differently AFAIK)

Those possibilities can't be dismissed but are very far from certain...
RDNA3 is a new architecture, so I don't think It's far-fetched to expect higher gains from driver updates than RDNA2.
We will have to wait for the next big Adrenaline update, maybe someone will retest It.

That difference was at 1440p, which should be plenty fair resolution target for a "midrange" card circa 2023.
Ok, it wasn't visible originally and the other was at 4K, so I assumed It was the same. I don't think N33 can achieve 37% better performance than N23 at 1440p.
 

GodisanAtheist

Diamond Member
Nov 16, 2006
6,717
7,010
136
It would be quite a hoot if N33 actually came with 1024 dual pumped SPs and it's being rumored out as 2048 similarly to how the higher N3x dies had their SP count doubled thanks to the dual pumped SPs.

If that's the case, given RDNA3 transistor bloat a 200mm2 turd that barely matches N23 let alone N22 would be in line for the tenor of this gen's releases.
 

TESKATLIPOKA

Platinum Member
May 1, 2020
2,329
2,811
106
It would be quite a hoot if N33 actually came with 1024 dual pumped SPs and it's being rumored out as 2048 similarly to how the higher N3x dies had their SP count doubled thanks to the dual pumped SPs.

If that's the case, given RDNA3 transistor bloat a 200mm2 turd that barely matches N23 let alone N22 would be in line for the tenor of this gen's releases.
N33 has 32CUs or 16WGPs, the same as N23 and that one had 2048 stream processors. N33 will have the same number of them, but dual issue(dual pumped).
Performance will be higher than N23.

If It had only 16CU(8 WGP) adding up to 1024 dual pumped SP, then It wouldn't match N23, but It wouldn't need 200mm2 die either.
 
Last edited:

Kronos1996

Junior Member
Dec 28, 2022
15
17
41
AMD needs to make up 10% in drivers just to match their launch numbers. What a soul-crushing waste of consumer trust the RDNA 3 launch was. They would’ve had almost no bad press if they were just honest about the launch performance.

I digress, this is an interesting generation and I’m curious to see where it goes. The rumors pointed to a very confident AMD and now we have what we got. I’m very curious about what other products and SKU’s they might be working on though.