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

 

[maddie]

Sapphire?

AMD

 

[Vope45]

Sapphire?

Igor mentioned we would likely see Asus and Sapphire Navi21 first while Vega 20 is AMD only. Who leaked Vega 20 ? RGT.

 

[dzoni2k2]

Igor mentioned we would likely see Asus and Sapphire Navi21 first while Vega 20 is AMD only. Who leaked Vega 20 ? RGT.

Vega 20?

Anyway I guess it could also be controlled AMD leaks.

 

[moinmoin]

What is happening. This is a complete mess.

You all are supposed to know nothing. Just wait the remaining 5 days, it's not that long.

 

[soresu]

COME ON AMD! RELEASE THE WHITEPAPERZ!

I WANT TO KNOW EVERYTHING ABOUT RDNA2!

We still lack a full deep dive on Zen 3, I don't think we will see one for RDNA2 next Wednesday either.

Though I'm happy to eat my words of course.

MORE INPUT STEPHANIE, MORE!!!

 

[Glo.]

https://twitter.com/x/status/1319674037686161411

Update Big Navi Firestrike Ultra benchmark

6800XT AIB: ~12800

RTX 3090 FE: ~12400
RTX 3080 FE: ~10600

-> 6900XT: 14000??

This is getting out of control.

 

[Veradun]

https://twitter.com/x/status/1319674037686161411

This is out of control.

*fixed*

 

[HurleyBird]

It's 3DMarks though. It's gonna be slightly worse in actual games.

It's going to be *different* in actual games, not necessarily worse.

 

[Konan]

It'd be nice to see the TS and TSE scores I think they are more indicative of knowing where things end up. See FSU and 5700XT as an example of outlier results vs 2070S vs real world non synthetic benchmarks.

 

[zinfamous]

6800XT is the 72 CU part.

this is what I'm thinking. 6900_ parts might be very limited supply out of the gate, and may be the only full-die versions of "Big Navi"

I could see 6900"XT" (with board partners), and 6900XTX (AMD only, very limited) as only 2 SKUs for 6900, at least for the first year.

6800 SKUs are 72CU with a wider range of SKUs, and maybe that 64CU SKU as the lowest tier of "Big Navi"? ....though my brain tells me that "Big Navi" would be reserved for 72 and 80 CU dies only.

This based on recent history of AMD delivering a limited stack for each generation. Rumors or whatever, either from within or outside of AMD promising full-stack release for RDNA2, like always, mean nothing until we actually see them. That hasn't happened in what, 4 straight generations (counting Polaris refresh as one "generation").

Even if it really is in their plan for this generation, production forces might be hindering them now.

 

[Ottonomous]

These leakers need to be a bit more professional and post something conclusive

 

[soresu]

https://twitter.com/x/status/1319674037686161411

This is getting out of control.

Reminds me of this line from Hunt for Red October....

Edit: The burning fire adds some to it ironically....

 

[Glo.]

This based on recent history of AMD delivering a limited stack for each generation. Rumors or whatever, either from within or outside of AMD promising full-stack release for RDNA2, like always, mean nothing until we actually see them. That hasn't happened in what, 4 straight generations (counting Polaris refresh as one "generation").

If we would use AMD's history, they would never come up with Zen architecture.

As I have said. I would not consider those leaks to be real, if they would come from twitter, or forum no-names.

But those leaks come from reputable sources, with access to AIBs, and those scores for benchmarks leaking are, most likely, from AIBs.

If it isn't apparent for anybody by this point, RDNA2 release is nowhere near anything close to Polaris, Vega releases. We have not seen such secrecy before, we have not seen such efficiency before, from Consoles, we have not seen such performance from leaks, before.

So why do we are still in disbelief that AMD could possibly develop a good architecture?

 

[zinfamous]

this is what I'm thinking. 6900_ parts might be very limited supply out of the gate, and may be the only full-die versions of "Big Navi"

I could see 6900"XT" (with board partners), and 6900XTX (AMD only, very limited) as only 2 SKUs for 6900, at least for the first year.

6800 SKUs are 72CU with a wider range of SKUs, and maybe that 64CU SKU as the lowest tier of "Big Navi"? ....though my brain tells me that "Big Navi" would be reserved for 72 and 80 CU dies only.

This based on recent history of AMD delivering a limited stack for each generation. Rumors or whatever, either from within or outside of AMD promising full-stack release for RDNA2, like always, mean nothing until we actually see them. That hasn't happened in what, 4 straight generations (counting Polaris refresh as one "generation").

Even if it really is in their plan for this generation, production forces might be hindering them now.

...and so if those most recent leaks are true (only just reading through them after posting that above^), and 6800Xwhateverlet'scallitT is the true 3080 competitor, maybe even beating it at around 5% or more on release, then it doesn't make *a lot* of sense for AMD to focus on super premium 6900 versions, which they will instantly sell out of, at any price, however many they make, if it does indeed beat the 3090.

They have much history of being supply limited (or simply release limited) at the very top of their stack than otherwise, so I think it makes more sense to expect a very rare 6900 whatever, probably two versions, and the XTX being something nuts like a 320W+ beast that runs hot and loud, like we know AMD loves to do, just because, while 6800XT is the efficient, properly volted and clocked version that everyone wants because while it runs far more efficient at lower voltage, you can probably still overclock the crap out of it and still get *really close* to an AIO 6900XTX for like $500 less. ....I mean, this is what they do, so why not do it again?

 

[Mopetar]

Still doesn't make sense as to why the 6800xt is a cut down die, why AIB partners would only get a cut down die when most dies produced would be full 80cu dies, why there is no 6900 non xt and why there would be a 20% cut down die(64cu).

Even if most chips could be sold as 80 CU parts AMD might bin a lot of them as a 72 CU part if disabling part of the hardware allows for better clocks.

It makes sense for there to be at least one 80 CU part, and possibly two if there's a big difference in clock speeds. Call those the 6900 and 6900X (or XT).

A more mainstream 72 CU die is easy to ensure supply for and better average clock speed for if they do bi that way. Call that a 6800X. The add a final bin for what will be a 6800. That could have as few as 64 CUs and is likely the bin for dies with a damaged memory controller that would make it a 192-bit bus. If they do have an abnormally large cache it could also be the catch-all for anything too damaged to be sold as anything else while still salvaging the silicon and creating something to slot in above the 6700 series.

 

[Vope45]

So why do we are still in disbelief that AMD could possibly develop a good architecture?

Probably because people are cautious of AMD and their infamous brand image.

 

[soresu]

and the XTX being something nuts like a 320W+ beast that runs hot and loud, like we know AMD loves to do

AMD loves more market share at as cheap as they can make their reference cards to be - I have little doubt that they grit their teeth at the notion of being talked of as the volcanically hot and loud option on the market.

They understand that people want quieter cards which is why they have the new reference designs with 2-3 proper fans, though you have to wonder just how much of the previous lackluster reference designs weren't due to AIB's wanting to make more money off their own differentiated aftermarket designs.

When AMD weren't performing well enough to compete against nVidia they had little leverage to tell the AIB's what to do, perhaps they do now.

 

[DDH]

this is what I'm thinking. 6900_ parts might be very limited supply out of the gate, and may be the only full-die versions of "Big Navi"

I could see 6900"XT" (with board partners), and 6900XTX (AMD only, very limited) as only 2 SKUs for 6900, at least for the first year.

6800 SKUs are 72CU with a wider range of SKUs, and maybe that 64CU SKU as the lowest tier of "Big Navi"? ....though my brain tells me that "Big Navi" would be reserved for 72 and 80 CU dies only.

This based on recent history of AMD delivering a limited stack for each generation. Rumors or whatever, either from within or outside of AMD promising full-stack release for RDNA2, like always, mean nothing until we actually see them. That hasn't happened in what, 4 straight generations (counting Polaris refresh as one "generation").

Even if it really is in their plan for this generation, production forces might be hindering them now.

But see this is the part that currently seems unreconcilable with 6800xt being navi21xt, 72cu part.

With TSMCs reported defect density in its 7nm process of 0.09, the ratio of good dies to defective dies is around 2:1.

https://www.anandtech.com/show/16028/better-yield-on-5nm-than-7nm-tsmc-update-on-defect-rates-for-n5
This report indicates and the graphs show that for 7+ nm defect density is at least as good as 7NP, if not better due to less stages in the lithography process.

So, for every 72cu die shipped to an AIB partner, amd are stock piling 2 full 80cu dies?


Below is a list of AMD gpu releases. For each segment, apart from the VII, the full die is always labeled as XT, with the cut XL or previously PRO. The XTX chip
if there has been one has remained in the same segment as the XT die.



Why is AMD seemingly changing this around this time, releasing a cut down 72cu die as an XT, keeping the full 80cu die as xtx, and shifting the XTX die to a new product segment, 6900xt

Then there is the report that the 6900xt will be limited in supply, but why would that be the case if it is the 80cu full die of the navi21 chip, when from TSMCs own reported defect density we can conclude that 2 full dies are created to 1 partial die. If the 80cu die is limited, then the 72cu is even more limited.


What does reconcile these differences is that the 6800xt is a different chip altogether, a 64cu ~400mm2 full die, while the 6900xt is the 80cu ~500mm2 full die. This way AMD can choose to produce less 80CU dies, and will be limited in supply as rumour suggests.

The leaked numbers though dont see to line up with a 64cu chip at 2.2 ghz however

 

[Veradun]

We're quickly approaching our destination.

Brace yourselves, leaks are becoming torrential.

 

[soresu]

This based on recent history of AMD delivering a limited stack for each generation. Rumors or whatever, either from within or outside of AMD promising full-stack release for RDNA2, like always, mean nothing until we actually see them. That hasn't happened in what, 4 straight generations (counting Polaris refresh as one "generation").

I feel like this gen it could have more to do with node crush.

Too many fabless companies are relying on TSMC N7 for volume and AMD already demands plenty for Zen 2, Navi 1x and now Zen 3.

I think they do want a full stack as they promised for RT in graphics, it's just going to take a while to build up the necessary stock for the different dies to make a full range of SKU's.

Then of course Rembrandt will come much later to round it all out with a nice APU option for mobile and NUC's.

 

[moinmoin]

If it's 72 CUs for the high-end mainstream card AMD may have taken a page from Sony's strategy of enabling 36 CUs on a 40 CUs die for PS5. There may well be math for this approach (disabling the worst 10% of all CUs) to make business sense.

 

[Graphenewhen]

But see this is the part that currently seems unreconcilable with 6800xt being navi21xt, 72cu part.

With TSMCs reported defect density in its 7nm process of 0.09, the ratio of good dies to defective dies is around 2:1.

https://www.anandtech.com/show/16028/better-yield-on-5nm-than-7nm-tsmc-update-on-defect-rates-for-n5
This report indicates and the graphs show that for 7+ nm defect density is at least as good as 7NP, if not better due to less stages in the lithography process.

So, for every 72cu die shipped to an AIB partner, amd are stock piling 2 full 80cu dies?


Below is a list of AMD gpu releases. For each segment, apart from the VII, the full die is always labeled as XT, with the cut XL or previously PRO. The XTX chip
if there has been one has remained in the same segment as the XT die.

View attachment 32278

Why is AMD seemingly changing this around this time, releasing a cut down 72cu die as an XT, keeping the full 80cu die as xtx, and shifting the XTX die to a new product segment, 6900xt

Then there is the report that the 6900xt will be limited in supply, but why would that be the case if it is the 80cu full die of the navi21 chip, when from TSMCs own reported defect density we can conclude that 2 full dies are created to 1 partial die. If the 80cu die is limited, then the 72cu is even more limited.


What does reconcile these differences is that the 6800xt is a different chip altogether, a 64cu ~400mm2 full die, while the 6900xt is the 80cu ~500mm2 full die. This way AMD can choose to produce less 80CU dies, and will be limited in supply as rumour suggests.

The leaked numbers though dont see to line up with a 64cu chip at 2.2 ghz however

That's a fair assessment, but I seem to remember a similar thing happening with Vega? Plenty of Vega 64s but limited Vega 56s, even though the 56 was supposed to be the sweet spot and could easily overclock to 64 performance?

 

[zinfamous]

If we would use AMD's history, they would never come up with Zen architecture.

As I have said. I would not consider those leaks to be real, if they would come from twitter, or forum no-names.

But those leaks come from reputable sources, with access to AIBs, and those scores for benchmarks leaking are, most likely, from AIBs.

If it isn't apparent for anybody by this point, RDNA2 release is nowhere near anything close to Polaris, Vega releases. We have not seen such secrecy before, we have not seen such efficiency before, from Consoles, we have not seen such performance from leaks, before.

So why do we are still in disbelief that AMD could possibly develop a good architecture?

You catastrophically misunderstand me, as not once did I ever suggest such a thing. Not even close. I am not commenting so much about performance, but about what they can release to market, and how the market responds to those releases. They are still as supply limited as everyone will be and, whether or not you are correct (and I tend to think you are), there is literally no reason to accept rumors and leaks that go against the well-known norm. It doesn't matter how different AMD is now, from 10, even 5 years ago when Zen was brand new, RDNA is looking like the early stages of Zen, just now being able to predictably compete. ...so I think that's where we are.

Plenty of people disbelieved everything being leaked about Zen (I didn't, so picked up a pile of shares at $4.60 at the time, much to my "being very smart"--that's a joke; no one is actually that smart. It's really just luck), and while I *think* we are at a similar moment with RDNA2, there is no logical reason to assume the situation repeats. Even with Zen, almost 5 years ago now, they are only just now catching up to Intel in terms of market. nVidia is not nearly as far behind as Intel has been, (and I mean realistically).

You mistake that I am betting against them. I most certainly am not.

 

[Glo.]

Probably because people are cautious of AMD and their infamous brand image.

Which is?

Kicking Intel's butt when they .... up?

So why can't AMD possibly do the same thing to their biggest GPU competitor, when their biggest GPU competitor is doing the same thing, as Intel?

Its only human mind problem, that it can see only where the hockey Puck was, instead of where it goes.

 

[Zstream]

When are these things supposed to be sold online?