Question Speculation: RDNA2 + CDNA Architectures thread

[uzzi38]

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

 

[Kenmitch]

Nah, they know better than to ask things they don't really wanna know the answer to
Carbonite.

You do at least change their outfits so they're holiday compliant?

 

[eek2121]

According to AMD, Q2 and Q3 is when the vast majority of console chips will be produced with ramp up in early Q2 and ramp down in late Q3. In Q4, there is a large back-off until again they start to ramp up in Q2 of 2021. This follows the seasonality of the console market in which the holiday season dominates for console sales so AMD has most output for consoles in Q2-Q3 so that the actual consoles can be ready in Q4.

So if AMD is launching Zen3 and Navi in early to mid Q4 as has been rumored, then they should have plenty of time to build inventory for launch followed by a quick ramp up in volume for the main holiday season, at least for the consumer market which has a much quicker time to market.

That is exactly what I was getting at. The contract between Sony/Microsoft and AMD spells out the numbers for the initial run. AMD will have had to dedicate a large amount of capacity for the initial run, and a much smaller amount of capacity for future production runs. The first production run is always the largest when dealing with stuff like this. AMD has enough parts in the retail channel that availability is *mostly* ok.

Zen 3 and Navi2X will likely be announce soon (end of month?) and they will both land in October or possibly the beginning of November.

 

[lobz]

You do at least change their outfits so they're holiday compliant?

I paid very close attention to make it look PC.

 

[DisEnchantment]

amdgpu commits indicate Sienna Cichlid is indeed Navi21.
Same patch with different commit message.

[PATCH] drm/amdgpu: Fix L1 policy violations (PSP) on sienna cichlid SRIOV

[PATCH] drm/amdgpu: Fix L1 policy violations (PSP) on Navi21 SRIOV

If you check the diff, the changes are the same

 

[jpiniero]

According to AMD, Q2 and Q3 is when the vast majority of console chips will be produced with ramp up in early Q2 and ramp down in late Q3. In Q4, there is a large back-off until again they start to ramp up in Q2 of 2021. This follows the seasonality of the console market in which the holiday season dominates for console sales so AMD has most output for consoles in Q2-Q3 so that the actual consoles can be ready in Q4.

Doesn't really work that way, especially with a launch.

With the PS4, they sold 4.2M by the end of the 2013.
By Early April, it was up to 7M.
August, 10M.
November, 14M.
End of 2014 18.5M.

If Sony bought enough wafers to get them 10M consoles, that will only last so long. They will need a steady stream of chips and not some ramp up and down, only adjusting down if demand doesn't hit their forecast.

 

[Stuka87]

Doesn't really work that way, especially with a launch.

With the PS4, they sold 4.2M by the end of the 2013.
By Early April, it was up to 7M.
August, 10M.
November, 14M.
End of 2014 18.5M.

If Sony bought enough wafers to get them 10M consoles, that will only last so long. They will need a steady stream of chips and not some ramp up and down, only adjusting down if demand doesn't hit their forecast.

Sony has already said that in order to be one of the first ones with the console, you have to prove your a sony fan, whatever that means. But, it could be due to limited quantities early on.

 

[Hitman928]

Doesn't really work that way, especially with a launch.

With the PS4, they sold 4.2M by the end of the 2013.
By Early April, it was up to 7M.
August, 10M.
November, 14M.
End of 2014 18.5M.

If Sony bought enough wafers to get them 10M consoles, that will only last so long. They will need a steady stream of chips and not some ramp up and down, only adjusting down if demand doesn't hit their forecast.

I was relaying what Lisa Su said in the last AMD earnings call.

Besides, your numbers actually prove what she was talking about. The PS4 launched on November 15, 2013 and sold 4.2M units in 1.5 months which is a 2.8M units/month sale rate. After the end of the year, it took them 3-3.5 months to sell an additional 3M units which is a 0.86M - 1M units/month sale rate. After the holidays they were selling at ~1/3 the rate. Then for the next 4 months, again 3M units or 0.75M units/month. Then when you get towards the holidays the sales rate ramps back up and peaks at 4.5M units in ~1.5 months or 3M units/month. This follows the exact timeline Lisa Su gave because you need the chips to be ready a few months before the consoles they go in are ready to be sold.

 

[beginner99]

Sony has already said that in order to be one of the first ones with the console, you have to prove your a sony fan, whatever that means. But, it could be due to limited quantities early on.

Or high prices early on. Sounds like $599 or higher initial pricing which makes sense given the actual hardware.

 

[uzzi38]

Alright, got a few things to post this fine morning. For starters:

https://twitter.com/x/status/1301712577110581249

Note it is talking about PPA, but still worth noting. Gives you some rough ideas about performance

https://twitter.com/x/status/1278419276492410880

I completely missed this a long time ago, so posting it now. Further proof to the fact that Navi10 isn't actually memory bottlenecked.

And lastly: https://www.pttweb.cc/bbs/PC_Shopping/M.1599147022.A.4CF

This is a post from the guy who leaked RDNA2 die sizes many months ago. Now he's stating something I've been a little worried about for a while.

Console orders are ramping up fast, and RTG seem to be prioritising fulfilling those over putting in additional orders for Navi21, despite the performance being good and the fact that AMD should still be able to extract pretty healthy margins from Navi21 (as the poster claims)

Expect low availability on launch is my bet. I wouldn't be surprised if the entire stock is just from the first batch of orders at that.

 

[Gideon]

Quoting a post from Ampere thread as it was about this Coreteks video:

Claiming RDNA2 only reaching RTX 3070 speeds (and the only selling point being supposedly 16GB of GDDR6).

If AMD's RDNA2 performance is at the RTX 3070 level, then wouldn't it technically compete with Nvidia's previous high end 2080 TI? I do wish Nvidia had been able to get the Ampere chips manufactured at 7 nm, which was one of the reasons I didn't bother with Turing. I also wonder just how big the next generation video cards are going to be. That would be funny if the missing coprocessor is a CPU socket because the video card is so big, it becomes it's own system.

I'm a bit sceptical about Coreteks claims (though dissapointing rumors about AMD's products have became true in the past). Why?

1. AMD obviously knew what their Big Navi performance target was going to be from the start (as they set it).
2. They knew where 2080 Ti stands for about 2 years now.
3. They also knew the next Nvidia card will be on Samsung 7nm EUV (as it was wildly reported at the time)
4. They still decided to market Big Navi as "high-end!" knowing it will release somewhere around 2020Q4 / 2020Q1.

If roughly 2080 Ti's perfomance is all they planned for, they must have had rocks in their head thinking this will be "high end" in 2021. There is no way Nvidia was gonna bring less than 30% performance uplift. They could have gotten most of that from the shrink.

Considering historic trends it would have been absolutely reasonable to model Ampere as at least 40-50% performance uplift vs Turing. Yes, Nvidia could have underdelivered, but you don't design your products on the premise that the competition will screw it up - see how well it turned out for Intel, or AMD in the past to that matter. On the other hand, expecting competiton to bring the best usually ends up with great products. See AMDs 64 Core "Rome" processors as an example, which were originally supposed to compete with Ice Lake Server from day one.

To top it all off while Ampere is slightly better than expected (mostly due to higher TDPs and lower prices), there is nothing that unique about 2070 reaching 2080 Ti perf.

• 970 was roughly as fast as 780 ti.
• 1070 was quite a bit faster than 980 ti.

Turin generation was the only outlier in this regard and considering Nvidia essentially went with the same node, it was somewhat to be expected (the only main difference between 16nm and 12nm is recticle limit).

TL;DR:
I hope Coreteks is dead wrong. If he's right AMD's project managers must have been dumb as a rock to market Big Navi as "High end"

 

[kurosaki]

The big navi will of course be at least double as fast as the 5700xt. Where does that put us on the scale? Beating the 2080Ti by quite a large margin. Way above 3070.

GPU 2019 Benchmarks - Compare Products on AnandTech

GPU 2019 benchmarks: Compare two products side-by-side or see a cascading list of product ratings along with our annotations.

www.anandtech.com

Anand bench

I dont believe that the IPC improvements have stood still either, so the question boils down to:
1: RT performance, how well does it fair in that arena compared to the competition?
2: Pricing. If the price of a 80CU, Navi similarly performing in RT costs 300-400 usd. It's going to be a blockbuster.

 

[DisEnchantment]

TL;DR:
I hope Coreteks is dead wrong. If he's right AMD's project managers must have been dumb as a rock to market Big Navi as "High end"

There are two things, I doubt he one of either.

• People who understand and/work with technology, skilled in various arts as they would say.
• People who have access to insider information.

From his explanation of things he has a superficial grasp of the subject he is talking about.
However I understand all of this comes from his passion for tech and I appreciate his willingness to acknowledge when he is wrong, unlike others.

 

[exquisitechar]

Quoting a post from Ampere thread as it was about this Coreteks video:

Claiming RDNA2 only reaching RTX 3070 speeds (and the only selling point being supposedly 16GB of GDDR6).



I'm a bit sceptical about Coreteks claims (though dissapointing rumors about AMD's products have became true in the past). Why?

1. AMD obviously knew what their Big Navi performance target was going to be from the start (as they set it).
2. They knew where 2080 Ti stands for about 2 years now.
3. They also knew the next Nvidia card will be on Samsung 7nm EUV (as it was wildly reported at the time)
4. They still decided to market Big Navi as "high-end!" knowing it will release somewhere around 2020Q4 / 2020Q1.

If roughly 2080 Ti's perfomance is all they planned for, they must have had rocks in their head thinking this will be "high end" in 2021. There is no way Nvidia was gonna bring less than 30% performance uplift. They could have gotten most of that from the shrink.

Considering historic trends it would have been absolutely reasonable to model Ampere as at least 40-50% performance uplift vs Turing. Yes, Nvidia could have underdelivered, but you don't design your products on the premise that the competition will screw it up - see how well it turned out for Intel, or AMD in the past to that matter. On the other hand, expecting competiton to bring the best usually ends up with great products. See AMDs 64 Core "Rome" processors as an example, which were originally supposed to compete with Ice Lake Server from day one.

To top it all off while Ampere is slightly better than expected (mostly due to higher TDPs and lower prices), there is nothing that unique about 2070 reaching 2080 Ti perf.

• 970 was roughly as fast as 780 ti.
• 1070 was quite a bit faster than 980 ti.

Turin generation was the only outlier in this regard and considering Nvidia essentially went with the same node, it was somewhat to be expected (the only main difference between 16nm and 12nm is recticle limit).

TL;DR:
I hope Coreteks is dead wrong. If he's right AMD's project managers must have been dumb as a rock to market Big Navi as "High end"

Coreteks makes Adored look like an expert. These performance claims will end up just like his Ampere RT coprocessor BS - something to laugh about after launch.

I can assure you that even Navi22 is too much for GA104.

 

[leoneazzurro]

This is a post from the guy who leaked RDNA2 die sizes many months ago. Now he's stating something I've been a little worried about for a while.

Console orders are ramping up fast, and RTG seem to be prioritising fulfilling those over putting in additional orders for Navi21, despite the performance being good and the fact that AMD should still be able to extract pretty healthy margins from Navi21 (as the poster claims)

Expect low availability on launch is my bet. I wouldn't be surprised if the entire stock is just from the first batch of orders at that.

This is not the only place where I heard about such a thing. Reason being, AMD prioritizing their mobile CPU lineup, and Zen 3 over discrete graphics as TSMC supply is quite tight until beginning of 2021. But on the business side of things, it is has a lot of sense as, margin wise, it is better to sell a Zen 3 16 core 4950x than a 80 CU Navi 21 board, for around the same price.

 

[kurosaki]

hing. Reason being, AMD prioritizing their mobile CPU lineup, and Zen 3 over discrete graphics as TSMC supply is quite tight until beginning of 2021. But on the business side of things, it is has a lot of sense as, margin wise, it is better to sell a Zen 3 16 core 4950x than a 80 core Navi 21 board, for around the same price.

Never reflected upon that. But. That of course makes a huuuge difference in margins.

 

[DisEnchantment]

Some good news for AMD and Open Source SW.

Oak Ridge Sponsoring GCC Compiler Improvements For NVIDIA + AMD GPU Offloading - Phoronix

www.phoronix.com

ORNL has been having a long running program called the "Software Readiness Program" to prepare for Exascale. One of the goals there was to replace CUDA specific code.
Now they are announcing a sponsorship to support GPU offloading in GCC. Of course this supports NV A100 and it will support AMD CDNA and all others accelerators that are spec compliant.

AMD has some basic GPU support in GCC but not as comprehensive as LLVM. Mentor Graphics was contracted by AMD some time ago to work on the GPU offloading but I suppose it was not comprehensive.

If you recollect one of the additional contracts that AMD won with the two Exascale projects was the development of ROCm.
I remember it was several hundreds of millions for that. ROCm is mainly LLVM based. So having this work for GCC now will be a big boost.
AMD's LLVM fork has around 30K+ lines of changes on top of LLVM11 to support ROCm. There is a major work ongoing to upstream all of that.
It looks like with LLVM12+ ROCm will work off the shelf with LLVM on debian based distros, with rocm-runtime and smi libs already available as debian packages and full kfd support for ROCm in amdgpu are getting merged (see Linux 5.9 commits).

Hopefully the GCC work is as extensive as AMD's work with LLVM.

 

[moinmoin]

Yes, margin wise consumer Radeon chips so far can't compete with any Zen chips. Which is exactly the reason why AMD would want to reach as high end as possible level with RDNA2 where even at lower than Nvidia prices the margins are just much higher. With datacenters now being occupied by CDNA, that's the only way consumer Radeon chips will ever be competitive with Zen margins.

As for a possible RDNA2 shortage, that's just on AMD's conservative planning from maybe two years ago. The constant backlog of Epyc chips as well as the current drought of Renoir chips show that that's something AMD in general has to face and isn't something specific to RDNA2 at all. At some point AMD has to rectify its capacity contingency planning to be less conservative, and I personally expect the monthly wafer orders currently occupied by the console chips will be what AMD will keep using for their other products, keeping the wafer orders at TSMC at a constant high level from that point on.

 

[DDH2]

Quoting a post from Ampere thread as it was about this Coreteks video:

Claiming RDNA2 only reaching RTX 3070 speeds (and the only selling point being supposedly 16GB of GDDR6).



I'm a bit sceptical about Coreteks claims (though dissapointing rumors about AMD's products have became true in the past). Why?

1. AMD obviously knew what their Big Navi performance target was going to be from the start (as they set it).
2. They knew where 2080 Ti stands for about 2 years now.
3. They also knew the next Nvidia card will be on Samsung 7nm EUV (as it was wildly reported at the time)
4. They still decided to market Big Navi as "high-end!" knowing it will release somewhere around 2020Q4 / 2020Q1.

If roughly 2080 Ti's perfomance is all they planned for, they must have had rocks in their head thinking this will be "high end" in 2021. There is no way Nvidia was gonna bring less than 30% performance uplift. They could have gotten most of that from the shrink.

Considering historic trends it would have been absolutely reasonable to model Ampere as at least 40-50% performance uplift vs Turing. Yes, Nvidia could have underdelivered, but you don't design your products on the premise that the competition will screw it up - see how well it turned out for Intel, or AMD in the past to that matter. On the other hand, expecting competiton to bring the best usually ends up with great products. See AMDs 64 Core "Rome" processors as an example, which were originally supposed to compete with Ice Lake Server from day one.

To top it all off while Ampere is slightly better than expected (mostly due to higher TDPs and lower prices), there is nothing that unique about 2070 reaching 2080 Ti perf.

• 970 was roughly as fast as 780 ti.
• 1070 was quite a bit faster than 980 ti.

Turin generation was the only outlier in this regard and considering Nvidia essentially went with the same node, it was somewhat to be expected (the only main difference between 16nm and 12nm is recticle limit).

TL;DR:
I hope Coreteks is dead wrong. If he's right AMD's project managers must have been dumb as a rock to market Big Navi as "High end"

Coreteks is bad, very bad.

I will completely destroy his claim in 1 sentence: a while ago an unknown AMD GPU popped up on a NVIDIA friendly openvr benchmark, besting the top 2080ti GPUs(think heavily overclocked) by 17%. The normal score for a standard 2080ti is 80, so this unknown AMD GPU beat the standard 2080ti by 30%, while being fed by an unreleased AMD *mobile* cpu

https://www.extremetech.com/computi...nown-amd-radeon-gpu-beating-rtx-2080-ti-by-17

30% is already on par with the 3080


Coreteks is bad, very very bad. Ignore him

 

[uzzi38]

He's not bad. Just mistaken I think.

https://twitter.com/x/status/1301839482287796224

8GB VRAM SKU is the big giveaway.

I'll just say I talk to him on Discord... it's not the first time he's gotten GPU dies completely confused. For the longest time he thought Navi10 had both GDDR6 and HBM2 controllers, insisting that a HBM2 SKU would launch for Apple at some point.

 

[marifire]

Long time reader, decided to register today, as this could be the best season for videocards in a very long time.

And Navi 10 was indeed launched later for Apple with HBM in the form of 5600M a few months back, great performance/watt if you ask...i think Coreteks usually has some true leaked info but just mixes some pieces.

Beeing optimistic and trusting Lisa, if Radeon wants to compete in high end again, they can do it based on improved RDNA1 and promised 50% performance/clock uplift, but they need 80Cus and high clock for RTX 3080/3090. 3090 is just 15-20% more powerful than 3080 this time. NVIDIA could release quite more expensive 3070 16GB and 3080 20GB, or call them Ti/Super, and also maybe some 3090 Ti or Titan with 84SM, but 3090 already has 82SM, so if Gaming Ampere 8nM full is 84SM, not a lot of margin there, not even in clocks considering power consumption.
AMD may need not just a small IPC gain but also say 20% more clock than navi10, 2300Mhz or so similar to PS5 clock...so the question is can they do it without sacrificing performance/power sweet spot as they were forced to in recent past launches? Remember 5700 and 5600 are both quite more efficient and lower voltage than 5700 XT.
What if that new Radeon RDNA2 is GDDR6 8GB/16GB 256-bit RX 6800 navi22 60Cus, close to or a bit higher than 3070/2080 Ti performance, and the "big navi" Navi 21 80/72Cus is HBM2 16GB/12GB RX 6900? Navi23 is expected later, maybe christmas, and should be great midrange, quite faster than RX 5700 too for around 249 USD, but both navi 22 and 21 should be released in october before consoles. If they manage 50% P/W increase and high clocks, could reach 225% more performance than RX 5700 XT, and thats around 45% more than 2080 Ti, so just between both 3080 and 3090.

 

[DDH2]

He's not bad. Just mistaken I think.

https://twitter.com/x/status/1301839482287796224

8GB VRAM SKU is the big giveaway.

I'll just say I talk to him on Discord... it's not the first time he's gotten GPU dies completely confused. For the longest time he thought Navi10 had both GDDR6 and HBM2 controllers, insisting that a HBM2 SKU would launch for Apple at some point.

Ill have to disagree with you there, he's bad.

His last vid claims amd have told partners they will now only compete with the 3070, implying previously they assumed they would compete with the 3080.

Firstly, For this to be true AMD would have had to assume the 3080 would have been the speed of the 2080ti.

Secondly, from the information i posted above, which is available for anyone to find and has been out for a while, completely contradicts his claim.

Finally, any analysis of the XBSX GPU would have also nulled his claim. 56cu running at 1.8ghz, which DF estimated to have the same performance as the 2080 in a 2 week old port of gears5 without any optimisation.

So either he is bad or he is willfully ignorant

 

[jpiniero]

You don't know when the 3080 competitor will be released. For all we know, it's only the 3070 competitor now, and the 3080 one will be released sometime next year.

IMO, AMD needs to say something before Ampere's release if they are releasing the 3080 competitor this year. If they don't say anything I think you have to assume that it's only the 3070 competitor.

 

[uzzi38]

Long time reader, decided to register today, as this could be the best season for videocards in a very long time.

And Navi 10 was indeed launched later for Apple with HBM in the form of 5600M a few months back, great performance/watt if you ask...i think Coreteks usually has some true leaked info but just mixes some pieces.

Beeing optimistic and trusting Lisa, if Radeon wants to compete in high end again, they can do it based on improved RDNA1 and promised 50% performance/clock uplift, but they need 80Cus and high clock for RTX 3080/3090. 3090 is just 15-20% more powerfull than 3080 this time. NVIDIA could release quite more expensive 3070 16GB and 3080 20GB, or call them Ti/Super, and also maybe some 3090 Ti or Titan with 84SM, but 3090 already has 82SM, so if Gaming Ampere 8nM full is 84SM, not a lot of margin there, not even in clocks considering power consumption.
AMD may need not just a small IPC gain but also say 20% more clock than navi10, 2300Mhz or so similar to PS5 clock...so the question is can they do it without sacrificing performance/power sweet spot as they were forced to in recet past? Remember 5700 and 5600 are both quite more efficient and lower voltage than 5700 XT.
What if that new Radeon RDNA2 is GDDR6 8GB/16GB 256-bit RX 6800 navi22 60Cus, close to or a bit higher than 3070/2080 Ti performance, and the "big navi" Navi 21 80/72Cus is HBM2 16GB/12GB RX 6900? Navi23 is expected later, maybe christmas, and should be great midrange, quite faster than RX 5700 too for around 249 USD, but both navi 22 and 21 should be released in october before consoles.

That's Navi12 that Apple use. Not Navi10.

 

[marifire]

That's Navi12 that Apple use. Not Navi10.

yes, you´re right but still Navi 1X HBM and it was expected much earlier than it finally launched.

 

[leoneazzurro]

5600M Mac version with HBM is very good, matching RTX2060 mobile with a 50W envelope.