Question Speculation: RDNA2 + CDNA Architectures thread

Page 203 - 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
All die sizes are within 5mm^2. The poster here has been right on some things in the past afaik, and to his credit was the first to saying 505mm^2 for Navi21, which other people have backed up. Even still though, take the following with a pich of salt.

Navi21 - 505mm^2

Navi22 - 340mm^2

Navi23 - 240mm^2

Source is the following post: https://www.ptt.cc/bbs/PC_Shopping/M.1588075782.A.C1E.html
 

insertcarehere

Senior member
Jan 17, 2013
639
607
136
Doesn't that clash with some earlier rumors that were expecting it to be clocked a lot higher? It should have the same 40 CU at a much higher clock speed along with the benefits of infinity cache. Unless they've decided to use more conservative clock speeds and power consumption, 20% comes across as underwhelming. If it's just using the Navi 21 clock speeds that both the XT cards use, that alone is almost 20% above the boost clock that the 5700XT got. The memory bus is smaller, but faster memory and the infinity cache should compensate for that.

A 20% performance boost still puts AMD squarely inline with Nvidia and 3060 Ti performance at 1080p, but it would cause a bit of a gap as the resolution increases. A $50 price difference would help them sell a lot of cards in normal circumstances, but I don't think it makes a lot of sense to do when there are as many supply issues.

If ~335mm^2 fully enabled Navi 22 on TSMC 7nm really can't beat partially enabled ~390mm^2 GA104 3060ti on SS 8nm that's not a good endorsement for the architecture.
 

GodisanAtheist

Diamond Member
Nov 16, 2006
6,716
7,006
136
If ~335mm^2 fully enabled Navi 22 on TSMC 7nm really can't beat partially enabled ~390mm^2 GA104 3060ti on SS 8nm that's not a good endorsement for the architecture.

- The RDNA2 arch has been punching well above its weight in rasterization workloads on a mm2 basis (more likely matching Ampere after all the Tensor stuff is removed). Would be a truly strange development for performance/mm2 to *regress* as the die got smaller.
 

insertcarehere

Senior member
Jan 17, 2013
639
607
136
- The RDNA2 arch has been punching well above its weight in rasterization workloads on a mm2 basis (more likely matching Ampere after all the Tensor stuff is removed). Would be a truly strange development for performance/mm2 to *regress* as the die got smaller.
Well, if Navi 22 is supposed to be a 40CU/64ROP part that's half the execution units vs Navi 21 at ~64% of the die size. While undoubtedly Navi 22 would be better at utilizing said units, that's still a pretty big deficit which will need to be made up.

GA104 is 48SM/96ROP vs GA102's 84SM/112ROP at ~62% of the die size for comparison, decrease in execution units is much more in line with die size decrease.
 

Mopetar

Diamond Member
Jan 31, 2011
7,797
5,899
136
The shader count makes less of a difference at 1080p and 1440p than it does at 4K. The 5700XT is already half the execution units on top of missing other features relative to the 6900XT, yet at 1080p it's slightly above 60% of the performance. At 1440p it falls to 55% and it isn't until 4K that it drops to 50%.

Also not all of Navi 22 is half. The memory bus and infinity fabric are 75% of Navi 21, but that's not necessarily important here. As good as Navi 21 is for lower resolutions it's not able to use all of the shaders. If there's a 6700XT that can hit higher clocks it should be a beast in 1080p.
 
  • Like
Reactions: Tlh97

beginner99

Diamond Member
Jun 2, 2009
5,208
1,580
136
Anything below RTX 3060 Ti for me for this GPU is straight up fail.
And that's why it won't be priced at just $350 more like $400 then add in tax and shortage fees and actual price in shops will be more like $500. I'm a managing my own expectations given the 6800(XT) pricing.
 

Glo.

Diamond Member
Apr 25, 2015
5,657
4,409
136
And that's why it won't be priced at just $350 more like $400 then add in tax and shortage fees and actual price in shops will be more like $500. I'm a managing my own expectations given the 6800(XT) pricing.
Nvidia may hit that 399$ MSRP point in 2021 fgor RTX 3060 Ti.

And THAT is a problem for AMD. AMD needs to bring better value in order for their solution to be considered better option. Look at 6800 XT. More efficient, faster in 90% scenarios, and 50$ cheaper.

This is the only way AMD can, at the moment compete with Nvidia.
 

Martimus

Diamond Member
Apr 24, 2007
4,488
152
106
I personally don't think it makes any sense either, but the reviews I linked seem to suggest that there is poor scaling or how else would 60 CU w/ decently higher clocks and the full 128 MB IC only perform only 50% better than a 5700XT? It's got better specs in every department.

I must be interpreting the data from the reviews incorrectly or something.
You said that it doesn't scale well, but your linear extrapolation assumes perfect scaling. You should be able to calculate scaling based on 6800XT to 6900XT scores, and use that to scale down at a better rate. Then you can extrapolate the data from the 6800 with the adjusted clock speeds.

Of course none of this will work in practice, as the clock speeds are actually different between all of these cards, with the 6800XT having the highest clocks in testing (although not by much over the 6900XT), but it will give you a good approximation. If I have time after my Stats test, I'll do the calculations for you.
 

Mopetar

Diamond Member
Jan 31, 2011
7,797
5,899
136
Nvidia may hit that 399$ MSRP point in 2021 fgor RTX 3060 Ti.

And THAT is a problem for AMD. AMD needs to bring better value in order for their solution to be considered better option. Look at 6800 XT. More efficient, faster in 90% scenarios, and 50$ cheaper.

This is the only way AMD can, at the moment compete with Nvidia.

If they're 10% better for the same price I think that's good enough. 3060 Ti is really at the edge of what's going to provide acceptable RT in lower resolutions, so I don't think it's as important for the ~$300 market.

I have a feeling that infinity fabric is going to give a good boost at 1080p and 1440p that might let AMD punch a bit above its weight class with their midrange cards.
 

Saylick

Diamond Member
Sep 10, 2012
3,084
6,184
136
You said that it doesn't scale well, but your linear extrapolation assumes perfect scaling. You should be able to calculate scaling based on 6800XT to 6900XT scores, and use that to scale down at a better rate. Then you can extrapolate the data from the 6800 with the adjusted clock speeds.

Of course none of this will work in practice, as the clock speeds are actually different between all of these cards, with the 6800XT having the highest clocks in testing (although not by much over the 6900XT), but it will give you a good approximation. If I have time after my Stats test, I'll do the calculations for you.
My point was that 50% more CUs, more effective bandwidth, higher average clocks, and what should be higher IPC would have led me to believe we'd get higher than 50% performance over a 5700XT, but that isn't the case: we only get 50%. Like, I said, I might be misinterpreting the data so if you have a more accurate calculation, I'd be more than happy to be corrected.
 

Martimus

Diamond Member
Apr 24, 2007
4,488
152
106
My point was that 50% more CUs, more effective bandwidth, higher average clocks, and what should be higher IPC would have led me to believe we'd get higher than 50% performance over a 5700XT, but that isn't the case: we only get 50%. Like, I said, I might be misinterpreting the data so if you have a more accurate calculation, I'd be more than happy to be corrected.
My point was that if it doesn't upscale perfectly, then it won't down scale perfectly either. If adding 50% more CUs and increasing clock speed by 20% doesn't increase performance by 60%, then doing the reverse won't decrease performance by 37.5%. It will have similar scaling downward and would only decrease scaling by 33.3% (if the 50% increase you state is correct). Of course you would need to adjust for higher clocks from there.
 
  • Like
Reactions: Tlh97

dr1337

Senior member
May 25, 2020
309
503
106
My point was that 50% more CUs, more effective bandwidth, higher average clocks, and what should be higher IPC would have led me to believe we'd get higher than 50% performance over a 5700XT, but that isn't the case: we only get 50%. Like, I said, I might be misinterpreting the data so if you have a more accurate calculation, I'd be more than happy to be corrected.
For comparisons sake when purely GPU compute limited we do in fact see 2x aka 100% the performance over navi10. However at 1080p yes this is only a 50% increase. Just as we've seen with ampere there are diminishing returns for increasing performance (at lower res) by adding more cores. Things like bandwidth, clockspeed, and the removal of CPU bottlenecks are far more important when it comes to lower resolution performance.
 

Attachments

  • average-fps_3840-2160.png
    average-fps_3840-2160.png
    90.5 KB · Views: 36
  • Like
Reactions: Tlh97 and scineram

Mopetar

Diamond Member
Jan 31, 2011
7,797
5,899
136
So it's pretty much in line with other results where RDNA2 is about equivalent to what Nvidia had with Turing. Not really surprising, but I think it kind of solidifies the point that if you care about RT at all this generation you should probably go with Nvidia. Even then the performance still isn't good enough and I think that the frame rates from a game that's over 20 years old shows that it's going to take another two or three generations of improvements to the technology before it's ready for mainstream use.
 

GodisanAtheist

Diamond Member
Nov 16, 2006
6,716
7,006
136
Given AMD's RT implementation, what can they do to improve their RT performance and bring it up to par with Nvidia (non-DLSS)?

Double up on RT cores per CU? Simply add another SE with a new array of CUs (scaling their dies up to NV size)?
 

Stuka87

Diamond Member
Dec 10, 2010
6,240
2,559
136
Given AMD's RT implementation, what can they do to improve their RT performance and bring it up to par with Nvidia (non-DLSS)?

Double up on RT cores per CU? Simply add another SE with a new array of CUs (scaling their dies up to NV size)?

The first thing is driver improvements. AMD has stated the drivers need a lot of work. And current games were all coded and optimized for nVidia's implementation (because it was the only implementation at the time).

I am not sure where the bottleneck is currently, so not sure what would be required to make it better.
 

Head1985

Golden Member
Jul 8, 2014
1,863
685
136
All true, the 6700 XT will be a much more "balanced" design in every way. It will be much closer to the 6800 than simply looking at the number of CUs would make one think.

The only way I can see it being only 20% faster than the 5700 XT is if they are conservative with clock speeds, which isn't going to happen. I don't think it will be $350 either, for sure.
I think 6700XT will cost 450usd and have 3060TI perf.Maybe 5% faster.Real MSRP 550euro.
MSRP in EU really does not matter because everything here cost 50% more than should.
 
  • Like
Reactions: Glo.

Mopetar

Diamond Member
Jan 31, 2011
7,797
5,899
136
The first thing is driver improvements.

There's a limitation to what drivers can accomplish and it's pretty evident that Nvidia has included a lot more specialized hardware for handling parts of RT workload that can't be done effectively with shaders. AMD might get better compatibility than they have now, but I wouldn't expect miracles with regards to performance improvements from drivers alone.

I think 6700XT will cost 450usd and have 3060TI perf.Maybe 5% faster.

If it's only the same as the 3060 Ti or just barely faster, I don't see them charging a $50 premium over it unless they have a boatload of cards and Nvidia doesn't.
 
  • Like
Reactions: Tlh97

lightmanek

Senior member
Feb 19, 2017
387
754
136
Just tested Quake II RTX on my overclocked RX6800XT and I'm happy with how it performs as I had very low expectations to start with.
Good news is, it works and it performs better than RTX2060 I had a chance to play on previously.

Full details with Upscaling AA:
1920x1080 = 79FPS on Demo1 map
2560x1440 = 48FPS on Demo1 map

I took few photos as screen grab was turning black screens form fullscreen mode.
20201217-001533.jpg
20201217-001625.jpg
20201217-001616.jpg
20201217-001654.jpg
20201217-001705.jpg


Top pictures are QHD and the last two are FHD
;)