Question Speculation: RDNA2 + CDNA Architectures thread

Page 84 - 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
2,705
6,427
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
 

CastleBravo

Member
Dec 6, 2019
120
271
136
There is so much demand right now for components / home electronics that I wouldn't expect supply to be able to meet demand initially, but if more than a handful of 'regular' people can get their hands on one on launch day, that would probably be a big win at this point.

Well, it is possible that he knows demand for RDNA2 will be exceptionally poor ;)
 

GodisanAtheist

Diamond Member
Nov 16, 2006
7,166
7,666
136
If Navi21 has right at 2x Navi10's 4k performance, it will be less than 2x on 1440 and 1080.

- The 5700xt experienced some non-linear performance drop off at 4K resolutions. My expectation/hope is that we see double performance at 1080/1440 and *better* than double performance at 4K as big-navi becomes less CPU bound.

This would not be unlike what we see with the 3080/3090 that really start stretching their legs at the 4K res.

AMD is just starting from so much further behind, which makes it less believable/likely to happen...
 

exquisitechar

Senior member
Apr 18, 2017
684
942
136
- The 5700xt experienced some non-linear performance drop off at 4K resolutions. My expectation/hope is that we see double performance at 1080/1440 and *better* than double performance at 4K as big-navi becomes less CPU bound.

This would not be unlike what we see with the 3080/3090 that really start stretching their legs at the 4K res.

AMD is just starting from so much further behind, which makes it less believable/likely to happen...
Yeah, it'll be more than 2x 5700XT performance at 4k.
 
  • Like
Reactions: Tlh97

Timorous

Golden Member
Oct 27, 2008
1,748
3,240
136
- The 5700xt experienced some non-linear performance drop off at 4K resolutions. My expectation/hope is that we see double performance at 1080/1440 and *better* than double performance at 4K as big-navi becomes less CPU bound.

This would not be unlike what we see with the 3080/3090 that really start stretching their legs at the 4K res.

AMD is just starting from so much further behind, which makes it less believable/likely to happen...

Is it really much further behind when they are starting with a 10.3B transistor die and RDNA performance per transistor is similar to Turing?
 

TESKATLIPOKA

Platinum Member
May 1, 2020
2,523
3,037
136
I am really starting to wonder If the info about 505mm2 and 80CU is actually correct. For 80CU, 128ROPs and 512bit Rdna1 GPU I got only 461mm2 and that is with the same transistor density.
I personally don't believe density will stay the same so either die size is wrong or the number of CU.
 

eek2121

Diamond Member
Aug 2, 2005
3,100
4,398
136
I am really starting to wonder If the info about 505mm2 and 80CU is actually correct. For 80CU, 128ROPs and 512bit Rdna1 GPU I got only 461mm2 and that is with the same transistor density.
I personally don't believe density will stay the same so either die size is wrong or the number of CU.
RDNA 2 features a slight rework to include stuff like RT.

There have now been multple trustworthy sources saying that:

Navi 21 - 80CU
Navi 22 - 60CU
Navi 23 - 40CU
Navi 24 - 20CU?
 

TESKATLIPOKA

Platinum Member
May 1, 2020
2,523
3,037
136
That doesn't necessarily mean they can't be wrong about the actual CU count, unless they have some insider info, which is in my opinion not very likely considering how tight lipped AMD is.
Or the CU count is correct, but die size is wrong.
 
  • Like
Reactions: Tlh97 and Olikan

uzzi38

Platinum Member
Oct 16, 2019
2,705
6,427
146
I am really starting to wonder If the info about 505mm2 and 80CU is actually correct. For 80CU, 128ROPs and 512bit Rdna1 GPU I got only 461mm2 and that is with the same transistor density.
I personally don't believe density will stay the same so either die size is wrong or the number of CU.
More than one person has corroborated the die size. AquariusZi is not the only person who says ~500mm^2 for Navi21. If he was, we'd wouldn't be so sure of it being this size.
 

TESKATLIPOKA

Platinum Member
May 1, 2020
2,523
3,037
136
More than one person has corroborated the die size. AquariusZi is not the only person who says ~500mm^2 for Navi21. If he was, we'd wouldn't be so sure of it being this size.
Then is 80CU wrong? Or you think density is the same and the missing 40mm2 is used up for better IPC and rt functionality?
 

TESKATLIPOKA

Platinum Member
May 1, 2020
2,523
3,037
136
Lower density and higher clock speed. Speed demon baby.
So instead of using high density libraries to increase density by ~25-50% to Renoir levels which can be used for more CUs, ROPs... they use high performance libraries to increase the frequency by ~20%? That's not a very good tradeoff in my opinion.
For example:
RDNA1 with 80CU, 128ROPs and 512bit GDDR6 is 461mm2.
If I increase density by 30% I end up with 120CU, 192ROPs and 512bit GDDR6 and the size would be ~451mm2. If I set 2Ghz for the first GPU then for the second GPU I would need 1.4-1.5Ghz for the same performance and at those clocks I think It would consume less power than the first one.
 
Last edited:

eek2121

Diamond Member
Aug 2, 2005
3,100
4,398
136
XBOXSX CU is actually slight smaller than n10


Yes, and if AMD uses HBM2, they could fit a full 8-core zen2 there... and would still have some space left

My understanding is that the GPU portion of the Xbox SoC is not directly comparable to RDNA2. They will be mostly similar, but there are also a number of differences.
 
  • Like
Reactions: Tlh97 and Olikan

Glo.

Diamond Member
Apr 25, 2015
5,803
4,777
136
So instead of using high density libraries to increase density by ~25-50% to Renoir levels which can be used for more CUs, ROPs... they use high performance libraries to increase the frequency by ~20%? That's not a very good tradeoff in my opinion.
For example:
RDNA1 with 80CU, 128ROPs and 512bit GDDR6 is 461mm2.
If I increase density by 30% I end up with 120CU, 192ROPs and 512bit GDDR6 and the size would be ~451mm2. If I set 2Ghz for the first GPU then for the second GPU I would need 1.4-1.5Ghz for the same performance and at those clocks I think It would consume less power than the first one.
"Large Caches" ;)
 

Glo.

Diamond Member
Apr 25, 2015
5,803
4,777
136
This guy is making money with clickbait and fakes. Dont know why anybody believes him after the Ampere fiasco. He was totally wrong on everything which was not leaked by Twitter.
Considering that he has photos of Quadro 6000 GPU that have not been released previously I would not post things like this, especially considering the amount of things he was correct on in terms of Ampere info ;).
 
  • Like
Reactions: spursindonesia

sontin

Diamond Member
Sep 12, 2011
3,273
149
106
A photo is nothing. We got leaks from the 3080 cooler in june. nVidia has sent Quadros to partners. It will be released in 1 1/2 weeks.

He was wrong with every "non leaked" point about Ampere. Go back and watch his videos again.

Dont trust such clickbait youtubers.