Question Speculation: RDNA2 + CDNA Architectures thread

Page 60 - 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
5,572
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
 

Timorous

Golden Member
Oct 27, 2008
1,532
2,535
136
With 80CUs I really don't see it being 500mm².

As I have said with Renoir density you can fit 21B xtors into around 350mm² which is large enough for 80CUs, TMU updates for Ray tracing, 384 bit bus and a bit more cache. We also know that Renoir clocks well and has an excellent power profile.

That would make so much more sense than some 500mm2 80CU part because AMD could use fewer wafers to build the same number of GPUs allowing for more Renoir parts to be built for the laptop OEMs which is a far more important market for AMD.
 

Kedas

Senior member
Dec 6, 2018
355
339
136
About the possible 128MB cache, the Latency (not bandwidth) could be much more important for all the extra more CPU like operations.

Maybe it's mainly used for the ray-tracing part of the die.
So basically having more bandwidth left for the GDDR6 operations.
 

A///

Diamond Member
Feb 24, 2017
4,352
3,154
136
RT data would change as much as cached scene data. I think not unless I'm missing something here. In any case, RDNA2 will be a disappointment in one way or another.
 
May 17, 2020
122
233
86
I mean, the pieces of the puzzle kinda makes sense...
1- N21 have 16 slices of L2, this means 256 or 512 memory width... 512 seems very unlike, also we have that photo leak with256-bit

2- a small density increase, and this huge cache can fit with 80Cu... also, Ray Tracing loves huge caches

3- N21 support xGMI, we were all puzzled why...
xGMI is to have i think infinity fabric links like an equivalent ot NVLink, i have reminded about this slide. It's maybe only for HBM2(e) versions. But NAVI 2x is not intented for datacenters only CDNA is for datacenters and NAVI 2x not planned neither to be in chiplets :

IF%20v1.jpg
 
Last edited:

Kedas

Senior member
Dec 6, 2018
355
339
136
RT data would change as much as cached scene data. I think not unless I'm missing something here. In any case, RDNA2 will be a disappointment in one way or another.
That depends on many things in their ray tracing design. In any case, it's speculation.
 

A///

Diamond Member
Feb 24, 2017
4,352
3,154
136
That depends on many things in their ray tracing design. In any case, it's speculation.
Even before Koduri was at AMD, and long after he'd left ATI, they still messed up. Of course, Raja is now at Intel and this evening he tweeted some stupid photo of a container of capsicum. Intel hired some of the worst food scraps in the industry; Raja Koduri, Chris Hook who departed Intel not long ago, and Heather Lennon who also left Intel back in December and have now settled at Nuvia Inc. Chris's Linkedin Profile has nothing set for where he works but given the location, it's in Sunnyvale. Not a lot of big names to figure out where he works. If I had to guess, Juniper.

It's still hard to take AMD seriously when it comes to graphics. Even if they got rid of the above clowns, they still have people like Herkelman and his dumb jebaited gag or Frank Azor with his goofy tweet using an Annie reference. Which is sad, because I've listened to podcasts and vlogs where Azor has been a guest and he's not some silly trickster.

I don't want AMD to rush RDNA2's software, but be real, they don't really stand a chance. Even if the product is amazing and matches Nvidia, they need to price lower because how bad their software is or has been historically, and the lack of similar features other than RT which is in the air at the moment.
 
  • Like
Reactions: Tlh97 and Mopetar

Kedas

Senior member
Dec 6, 2018
355
339
136
Using history to predict the future is nice but If you use their history to predict their future then Zen would never have existed, so every new design is an opportunity for them to do it (almost) right.
Even an 1% hit rate on the cache will still happen even it's rare, increasing the cache increases the hit rate much. (yes I'm playing with the meaning here)
 

A///

Diamond Member
Feb 24, 2017
4,352
3,154
136
Using history to predict the future is nice but If you use their history to predict their future then Zen would never have existed, so every new design is an opportunity for them to do it (almost) right.
Actually I would have predicted Zen. You're either really young or need to brush up on your AMD history. Historically, AMD had one-upped Intel at every corner and whooped their butt for years. They brought the 64bit extension to 32bit, and killed Intel's IA-64 goals dead in the water. At one point, AMD was in most datacenters. Intel's major save from the Pentium 4 era was their Core 2 Duo. I won't even go into the hot mess that Pentium D was, emphasis on HOT MESS. AMD has always been an innovater. That is a fact.

However, to expect a 50% increase in performance over RDNA... that's dreaming. I'm sure it'll be good, but the environment that NVidia have developed over time isn't there. This isn't really an AMD fault either. If you're old enough to have owned ATI cards, you'd know they were terrible in software, too.

I knew about Zen long before it came out. If you paid attention to the tech press, you would have, too. Maybe you did, I don't know. I was impressed with Zen. I was impressive with Zen+. I was and still am impressed by Zen2. I know Zen 3 is going to be very good. However, my big question is not whether TSMC can provide in the future. They have to or they lose their major customers. I'm wondering if AMD can continue delivering excellent performance gains per generation.

I don't have a lot of confidence in Intel's rumored 11th gen desktop, Rocket Lake. I have almost zero confidence in Alderlake. I'm sure they'll deliver it, eventually, but as for it being a beast? Heck no.
 
Last edited:

Olikan

Platinum Member
Sep 23, 2011
2,023
275
126
xGMI is to have i think infinity fabric links like an equivalent ot NVLink, i have reminded about this slide. It's maybe only for HBM2(e) versions. But NAVI 2x is not intented for datacenters only CDNA is for datacenters and NAVI 2x not planned neither to be in chiplets :
That's exacly why we were puzzled...
Navi2x do have xGMI
o4mz5p4t8s251.jpg
 

linkgoron

Platinum Member
Mar 9, 2005
2,286
810
136
However, to expect a 50% increase in performance over RDNA... that's dreaming. I'm sure it'll be good, but the environment that NVidia have developed over time isn't there. This isn't really an AMD fault either. If you're old enough to have owned ATI cards, you'd know they were terrible in software, too.
A 50% increase over RDNA in what way? For the same die size? For what?

If AMD creates a ~500mm^2 card with a new architecture - anything below a 100% improvement over a half-the-size and older architecture 5700xt would be a failure IMO. However, according to recent leaks, even a card with twice the performance of a 5700xt would probably be just about equal to the 3080 performance wise. Hopefully AMD manages to beat the 3080 by 5-10% or so.

Blame their arrogance towards AMD.
I know that this is the RDNA thread, but I'm not sure if it's arrogance or just that Samsung really failed to deliver. I still think that the only reason we're getting GA102 so soon is that GA104 failed to deliver the expected performance (as well as the whole stack, obviously). When was the last time Nvidia released 320/350 watt cards or (according to leaked benchmarks) mediocre perf/watt improvements with a node shrink? As we've already seen that GA100 is on TSMC, I'm wondering if we'll see a TSMC 7nm Ampere refresh in 2021.
 

Kuiva maa

Member
May 1, 2014
181
232
116

Assuming the same very rough 10% increase the 5700XT sees from moving to a better API applies to this card, it seems to be slightly below 2080Ti and I would assume 3070. That seems from lackluster (if it is 6700 series) to pretty horrendous (if it is 6900 series). But I will stop here, too many assumptions and unknown variables to get a clear picture.
 

Glo.

Diamond Member
Apr 25, 2015
5,657
4,409
136

linkgoron

Platinum Member
Mar 9, 2005
2,286
810
136
Wait, so more Navi # means less performance?!
Within the same "family" - Polaris and Navi 1x were the same. Polaris 10 was 480, Polaris 11 was 460. Navi 10 is 5700/5600, Navi 14 is 5500/5300.

IIRC Raja Koduri stated that the numbers just mean that the "lower" number GPU R&D just started earlier.
 
Last edited:
  • Like
Reactions: Mopetar

GodisanAtheist

Diamond Member
Nov 16, 2006
6,716
7,004
136
Within the same "family" - Polaris and Navi 1x were the same. Polaris 10 was 480, Polaris 11 was 460. Navi 10 is 5700/5600, Navi 14 is 5500/5300.

IIRC Raja Koduri stated that the numbers just mean that the "lower" number GPU R&D just started earlier.

-I suppose we can assume that the R&D starts with the biggest part, and then they start work on paring the die down into smaller and smaller parts, hence the larger the # the smaller the die.
 
  • Like
Reactions: Tlh97 and moinmoin

ozzy702

Golden Member
Nov 1, 2011
1,151
530
136
I really want to believe that, but I'll temper my expectations for now. Expect the worse and all your surprises will be pleasant ones

Yep, I hope I'm wrong, but it's been a long time since AMD delivered that kind of competition at the high end and I don't expect the 3080 and 3090 to see solid competition. 3070TI on down, absolutely. Time will tell, I'd love to have viable options from AMD on the GPU front. I can't remember the last time I recommended Intel for CPUs.
 

GodisanAtheist

Diamond Member
Nov 16, 2006
6,716
7,004
136
Yep, I hope I'm wrong, but it's been a long time since AMD delivered that kind of competition at the high end and I don't expect the 3080 and 3090 to see solid competition. 3070TI on down, absolutely. Time will tell, I'd love to have viable options from AMD on the GPU front. I can't remember the last time I recommended Intel for CPUs.

-The last time AMD caught NV by surprise, it was with the HD4000 series vs the GTX2xx series when NV had bloated their dies up with GPGPU functionality that added little to gaming (at the time).

I feel like (and I think the sentiment is generally) that we're in a similar moment here where NV has bloated their dies supporting a cool but narrowly adopted feature and AMD is in a good position to come in and offer 85-90% of the performance for a much more competitive price.

If AMD is going going to surprise, now is their time to shine. I understand deeply and completely how this feeling slams head long into "but we've been hurt with fake hype and false promises before" so no judgement on feeling down on AMD.
 
  • Like
Reactions: Mopetar and Tlh97

ozzy702

Golden Member
Nov 1, 2011
1,151
530
136
-The last time AMD caught NV by surprise, it was with the HD4000 series vs the GTX2xx series when NV had bloated their dies up with GPGPU functionality that added little to gaming (at the time).

I feel like (and I think the sentiment is generally) that we're in a similar moment here where NV has bloated their dies supporting a cool but narrowly adopted feature and AMD is in a good position to come in and offer 85-90% of the performance for a much more competitive price.

If AMD is going going to surprise, now is their time to shine. I understand deeply and completely how this feeling slams head long into "but we've been hurt with fake hype and false promises before" so no judgement on feeling down on AMD.

I hope that's the case. Over the years I've owned a 4850, 4870, 4890, 6870, 7950, 7970, 390, 470, 480, 580... so it's not like I've been an NVIDIA fanboi, I just haven't had a compelling reason to run an AMD GPU in my main box for quite some time and have been disapointed time and time again over the past few years by AMD's releases.

I'm looking forward to seeing what AMD brings us with the Zen3 and RDNA2. It may be the first time in the past 15 years that I run a red red combo!