Question Speculation: RDNA2 + CDNA Architectures thread

Page 64 - 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
5,575
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
 

leoneazzurro

Senior member
Jul 26, 2016
905
1,430
136
Yes, I was speaking about the actual fans, not the holes. You can clearly see that the blades on the new card are almost touching the hole's edge, while on the Radeon VII there is a quite noticeable gap between the blades and the hole's wall. So that's some millimeters per side (probably 5 per side, amounting for 10mm bigger diameter)
 
Last edited:

kurosaki

Senior member
Feb 7, 2019
258
250
86
The big problem, as I see it, is that they still havent managed to find a solution for pushing some of the air outside the back of the card. Radial fans, we do not want them back, but look at the 30xx-cards. Thay have managed to do something Ive been wanting to do forever. Pushing air out of the back of the computer with an axial fan. How hard could it be, to manage to do what Nvidia has recently managed to do. Some one in AMD/ATI has to hire me as a consultant ASAP to discuss airflow desigs! :D
 

coercitiv

Diamond Member
Jan 24, 2014
6,151
11,686
136
Pushing air out of the back of the computer with an axial fan. How hard could it be, to manage to do what Nvidia has recently managed to do.
Let's wait and see just how efficient that Nvidia cooler is with pushing hot air out the back of the card. Keep in mind that fan blows air in 2 directions, to the back of the card and the opposite direction, towards the center of the card.

Untitled-1.jpg

To me the Nvidia card reminds me of the Sapphire Pulse Vega 56 with it's half size "nano" PCB and strong emphasis on cooling through the "free" half of the heatsink. The fans on the front of the card may look identical, but the amount of heat they displace from the heatsink certainly isn't.

Untitled-1.jpg
 

kurosaki

Senior member
Feb 7, 2019
258
250
86
Let's wait and see just how efficient that Nvidia cooler is with pushing hot air out the back of the card. Keep in mind that fan blows air in 2 directions, to the back of the card and the opposite direction, towards the center of the card.

View attachment 29868

To me the Nvidia card reminds me of the Sapphire Pulse Vega 56 with it's half size "nano" PCB and strong emphasis on cooling through the "free" half of the heatsink. The fans on the front of the card may look identical, but the amount of heat they displace from the heatsink certainly isn't.

View attachment 29867

As long as you have positive pressure in the computer, the air wants to move out the easiest way possible. I just dont get it why they still have not managed to build a cooler that uses axial fans, and pushes the air outside the box.
 

Timorous

Golden Member
Oct 27, 2008
1,532
2,535
136
Vega 20 was a 331 sq mm die with 4 HBM2 stacks. Navi 21 is likely to be a 500 sq mm die with 2 HBM2E stacks.

I am really starting to doubt 500mm^2 with 80CUs. It just does not seem credible.

AMD have already had Renoir at > 60M xtors/mm^2 and that clocks well with good power characteristics.
NV have also had GA100 at 65.6M xtors/mm^2 on 7nm and that is a huge 54B transitor chip that clocks at 1.4Ghz in a 400W envelope.

I really think NV21 will have density closer to 60M xtors/mm^2 than 40M and given that it would make 21B xtor die fit in a 350mm^2 area which is similar in size to the R7 die. This would fit with the bracket and screw hole positioning on the fortnite render looking very similar to the R7. If NV21 is 500mm^2 and has Renoir/GA100 density then it is likely to have around 30B transistors and that is a lot to work with.
 

leoneazzurro

Senior member
Jul 26, 2016
905
1,430
136
It depends on what the transistors are used for. I.e. advance power management, clock gating, improved clocks...
Moreover, if the card is around 275/300W as speculated, a die too small will be too thermally dense to be properly cooled.
I.e. the 3080 may have around up to 250W of pure GPU thermal power to get rid of. But this is on a 600+mm^2 area. If the same power must be extracted from a slightly above 300 mm^2 die, well... let's say it's not as easy.
 

jpiniero

Lifer
Oct 1, 2010
14,510
5,159
136

PS5 SoC yields are bad, under 50%. GPU frequency is I'm guessing the culprit. Wonder if AMD told Sony that 2.2 was very doable, and turns out it's not.
 
  • Like
Reactions: kurosaki

Timorous

Golden Member
Oct 27, 2008
1,532
2,535
136
It depends on what the transistors are used for. I.e. advance power management, clock gating, improved clocks...
Moreover, if the card is around 275/300W as speculated, a die too small will be too thermally dense to be properly cooled.
I.e. the 3080 may have around up to 250W of pure GPU thermal power to get rid of. But this is on a 600+mm^2 area. If the same power must be extracted from a slightly above 300 mm^2 die, well... let's say it's not as easy.

Renoir is 150mm^2 and clocks to 4.45+ Ghz on the CPU + 2.4+ Ghz on the GPU + 2+ Ghz on Infinity Fabric and that got upto 160W in furmark. (Using Toms Hardware numbers)

300W board power would actually be lower heat density than that OC'd Renoir APU so I don't see that being an issue.
 

leoneazzurro

Senior member
Jul 26, 2016
905
1,430
136
Of course it's always AMD's fault lol. Or maybe Sony felt they were going to be inferior to Xbox and asked AMD to push frequencies and this is the result...
Also, it was said in a PS5 conference that PS5 CUs are different respect to desktop RDNA2's CUs.. with the latter being beefier than the former. So it must be seen also if their "customization" were taken a bit too far.
 

leoneazzurro

Senior member
Jul 26, 2016
905
1,430
136
Renoir is 150mm^2 and clocks to 4.45+ Ghz on the CPU + 2.4+ Ghz on the GPU + 2+ Ghz on Infinity Fabric and that got upto 160W in furmark. (Using Toms Hardware numbers)

300W board power would actually be lower heat density than that OC'd Renoir APU so I don't see that being an issue.

Maybe it's an issue when you are guaranteeeing it running 24/7 for at least one year and on a large number of cards with a stock cooler instead of running a single test on a single lucky sample with a cooler that is not designed by you and probably beefed up...
 

Timorous

Golden Member
Oct 27, 2008
1,532
2,535
136
Maybe it's an issue when you are guaranteeeing it running 24/7 for at least one year and on a large number of cards with a stock cooler instead of running a single test on a single lucky sample with a cooler that is not designed by you and probably beefed up...

I am not convinced. If NV21 has similar density to NV10 then that is less than Ampere. That really does not seem credible. It may not be all the way to 60+M xtors/mm^2 but I do expect north of 50M.
 
  • Like
Reactions: spursindonesia

leoneazzurro

Senior member
Jul 26, 2016
905
1,430
136
How many ways of counting transistors are there?

A lot


It depends of what you are actually counting, only functional units as schematics or actual implementation, and for which type of circuitry.
In this article is explicitly stated that, for the same design, the "transistor count" may vary, according to what the producer actually counts, as much as 33-37%
 
Last edited:

Zstream

Diamond Member
Oct 24, 2005
3,396
277
136

PS5 SoC yields are bad, under 50%. GPU frequency is I'm guessing the culprit. Wonder if AMD told Sony that 2.2 was very doable, and turns out it's not.
Source?
 

A///

Diamond Member
Feb 24, 2017
4,352
3,154
136
The reporter who reported Sony's issues seems to be a gray wolf. Apparently they've been banned from multiple agencies in the past due to publishing unsubstantiated articles in the past. This is according to something I read off of Reddit's investing section. Sony's doing some custom chip work for their PS5, coupled with reducing CUs and pushing clocks as far as they can go up per their announcement weeks ago, I'm not surprised if there is some tangible credibility here. Sony may not repeat MS's mistake with pricing, but it looks like due to their own self-imposed shortcomings, MS will take the lead.
 

A///

Diamond Member
Feb 24, 2017
4,352
3,154
136
Let's wait and see just how efficient that Nvidia cooler is with pushing hot air out the back of the card. Keep in mind that fan blows air in 2 directions, to the back of the card and the opposite direction, towards the center of the card.

View attachment 29868

To me the Nvidia card reminds me of the Sapphire Pulse Vega 56 with it's half size "nano" PCB and strong emphasis on cooling through the "free" half of the heatsink. The fans on the front of the card may look identical, but the amount of heat they displace from the heatsink certainly isn't.

View attachment 29867
I was thinking this. It predates the Vega 56 design.

9pRXRUg.png
 
  • Like
Reactions: Tlh97 and coercitiv

Helis4life

Member
Sep 6, 2020
30
48
46
The reporter who reported Sony's issues seems to be a gray wolf. Apparently they've been banned from multiple agencies in the past due to publishing unsubstantiated articles in the past. This is according to something I read off of Reddit's investing section. Sony's doing some custom chip work for their PS5, coupled with reducing CUs and pushing clocks as far as they can go up per their announcement weeks ago, I'm not surprised if there is some tangible credibility here. Sony may not repeat MS's mistake with pricing, but it looks like due to their own self-imposed shortcomings, MS will take the lead.
That's interesting. See of Sony coroborates anything.


Article says as low as 50%, not less than. Also says 11 million chips made, doesn't specify a time period though. Current us fiscal year ends 31st September.

I was under the impression the chips were made to order based on monthly projections, rather than bulk manufacturing. How long would it take to sell 11 million units?
 

Hitman928

Diamond Member
Apr 15, 2012
5,178
7,629
136
That's interesting. See of Sony coroborates anything.


Article says as low as 50%, not less than. Also says 11 million chips made, doesn't specify a time period though. Current us fiscal year ends 31st September.

I was under the impression the chips were made to order based on monthly projections, rather than bulk manufacturing. How long would it take to sell 11 million units?

The PS4 sold just over 4 million from November through the end of the year in its launch year. It then had sold a total of 10 million units by Aug 10 of the following year. The Nintendo switch sold about 15 million units in 10 months which I believe made it the fastest selling console ever and came in at a lower price point which obviously will help sales. So if Sony had really ordered 17 million chips to be ready by March 2021 (according to the article), that seems very aggressive to me. That means they were planning on selling faster than the switch within the first year of launch and most likely doing that bulk of those sales after the holidays. I don't have any "sources" but that doesn't seem realistic.

As low as 50% parametric yield also seems unrealistically low. I mean, it's not a new process and AMD has already put multiple CPUs and GPUs on this process. Even if Sony asked AMD to essentially overclock the GPUs, I highly doubt AMD wouldn't have been able to tell them they'd get really terrible yields of they did. I could understand it being lower than expected but that 50% number just seems made up.
 

Zstream

Diamond Member
Oct 24, 2005
3,396
277
136
Uh? The article? You can read the english version of the article above if Google Translate isn't your thing.
Calm down, the source of the article is just random jibberish. No way Sony would build out 15m million units in a SILO and "wait" for buyers. This type of thing does not happen, and would be counter productive to refresh cycles or revisions of the SoC. You build as many as you project within a 6-8 month period, and work on revisions for the rest. Yields will get better as time progresses.

Let's for fun say Sony predicted 15 million consoles for sale. That might have been pre-covid, and with the world economy the way it is, no way in heck that many will be sold. It's definitely wishful thinking.
 
  • Like
Reactions: Lodix

jpiniero

Lifer
Oct 1, 2010
14,510
5,159
136
No way Sony would build out 15m million units in a SILO and "wait" for buyers.

It makes sense for consoles since while you don't know what demand will be, there's little downside to overproducing since you know you will sell it eventually. Well, as long as yields are good. MS has said the Series X SoC yield is good.
 

MrTeal

Diamond Member
Dec 7, 2003
3,554
1,658
136
The big problem, as I see it, is that they still havent managed to find a solution for pushing some of the air outside the back of the card. Radial fans, we do not want them back, but look at the 30xx-cards. Thay have managed to do something Ive been wanting to do forever. Pushing air out of the back of the computer with an axial fan. How hard could it be, to manage to do what Nvidia has recently managed to do. Some one in AMD/ATI has to hire me as a consultant ASAP to discuss airflow desigs! :D

As long as you have positive pressure in the computer, the air wants to move out the easiest way possible. I just dont get it why they still have not managed to build a cooler that uses axial fans, and pushes the air outside the box.
Because using an axial fan to push air through a heatsink that then hits a PCB, has to compress and take a 90 and travel along the PCB along a <1" path out the back of the case is really inefficient. Pulling in air through the axis and pushing it out radially is the whole reason radial fans exist. Blower cards aren't loud because a radial fans is a fundementally bad design, they're loud because they need to run at high RPM to overcome the static pressure of pushing through the heatsink and venting out a 075"x3.5" slot at the back of the card while still moving enough air to keep the card cool. Radial fans are inherently better in high backpressure applications at the expense of flowrate, so they do tend to get noisy when they get cranked up to move enough air to cool 300W from a small fin array.
 

kurosaki

Senior member
Feb 7, 2019
258
250
86
air through a heatsink that then hits a PCB, has to compress and take a 90 and travel along the PCB along a <1" path out the back of the case is really inefficient. Pulling in air through the axis and pushing it out radially is the whole reason radial fans exist. Blower cards aren't loud because a radial fans is a fundementally bad design, they're loud because they need to run at high RPM to overcome the static pressure of pushing through the heatsink and venting out a 075"x3.5" slot at the back of the card while still moving enough air to keep the card cool. Radial fans are inherently better in high backpressure applications at the expense of flowrate, so they do tend to get noisy when they get cranked up to move enough air to cool 300W from a small fin array.
Tomorrow we'll see if the Axial blower is any good or not. I believe Nvidia has quite a silent and cool design coming up.
*As long as there is a positive pressure in the case, I believe we wont have large problems moving the air out via a GPU. So, while an axial blower seems hard to pull off in a negative pressure case, I think it's quite the oposite in a positive pressure one.
 
Last edited: