AMD: Volcanic Islands R1100/1200 (8***/9*** series) Speculation/ Rumour Thread

[CarstenS]

If you just scale with linear transistor density, my guess is that the chip will be uncoolable by standard means. Rememeber, you get a theoretical/ideal 49% reduction in die-size (or a corresponding increase in transistor count).

Given that Tahiti XT at 365 mm² consumes at a conservative estimate 250 Watts, you might look at around ((550/1.3)/365)x2x4.312M i.e. 10bn Transistors and thus at roughly (9994M/4312M)*0,7*250 ie. 405 Watts.

 

[A1xLLcqAgt0qc2RyMz0y]

I'm quite sure the fact that they're not doing them doesn't mean they wouldn't know how to make one.

Please explain, as you state, that AMD knows how to make a big die (550mm2) GPU why they refuse to do so.

 

[Alexko]

Please explain, as you state, that AMD knows how to make a big die (550mm2) GPU why they refuse to do so.

In this case it would be incumbent upon you to explain how making a 550mm² die is fundamentally different from making a 420mm² one.

 

[3dilettante]

Would they make money doing so?
Nvidia put the software and developer investment into building the necessary relationships with the professional and computing segment.
The history of better tools and compatibility allows them to sell for a higher price, which justifies the expense of a high-end die with lower volumes than the mainstream market.
It's something AMD has made some efforts to meet, but it's an uphill climb.

AMD would know how to build a big die like that, but why would the market pay the premium for it?

 

[Wynix]

I think they made the decision to only take on the GTX780, this way they can have a much smaller die size, thereby increasing yields.
If they wanted to beat Titan or a full GK110 they would likely have to push past 500mm, which would be much more expensive.

As someone who has no intention of purchasing a $999 card, i believe they made the right call.

The matter about recycling Tahiti though...

 

[lanek]

Oh SHI***
It will slower, than 780
AMD does not know how to make big GPUs 500>mm2.

Im not sure with all the transistors and parts disabled on the 780 we can still speak about the die size of it. ( DP units, 3 SMX etc etc ).. What will be the size of a 780 if it was build from the scratch with the same specifications, instead of been a disabled GK110 ? Less SP, higher clocks. ( Retail OC 780 models match easely and pass over Titan, less SP higher clock. )

 

[silent_guy]

Saying 'AMD does not know how to make big die' is as stupid as saying 'Nvidia doesn't know how to produce with good yields.'

 

[Alexko]

Would they make money doing so?
Nvidia put the software and developer investment into building the necessary relationships with the professional and computing segment.
The history of better tools and compatibility allows them to sell for a higher price, which justifies the expense of a high-end die with lower volumes than the mainstream market.
It's something AMD has made some efforts to meet, but it's an uphill climb.

AMD would know how to build a big die like that, but why would the market pay the premium for it?

I think that's exactly the reason why AMD hasn't been making very large GPUs. They typically don't generate very high volumes in the consumer market, so they need the high margins of the professional market to justify the expense.

But, if Hawaii is indeed ~420mm² as Matt Skynner suggested, and if AMD's pro market share has indeed been rising continuously for the last few quarters, as executives mentioned during conference calls, and if this trend is maintained, we might see very large chips from AMD within a generation or two.

 

[A1xLLcqAgt0qc2RyMz0y]

In this case it would be incumbent upon you to explain how making a 550mm² die is fundamentally different from making a 420mm² one.

Nope, sorry FAIL.

A question as an answer to a question. Looks like a classic side step.

 

[silent_guy]

Please explain, as you state, that AMD knows how to make a big die (550mm2) GPU why they refuse to do so.

Because they think a portfolio without a big die makes more sense from a commercial and/or resource point of view.

And if they change their mind this time around then that because they've changed their mind on it. None of this is clear cut. There's a lot of opinion and guessing involved in deciding this stuff because you're trying to predict the future.

AMD knows how to make big dies: 360mm2 is big by all common standards. To make a 450mm2 die, you need to run the same kinds of analysis to make sure it works. Stop this kind of silliness.

 

[Albuquerque]

Nope, sorry FAIL.

A question as an answer to a question. Looks like a classic side step.

You're asking an absurd question, and then somehow treating the obvious answer like a five year old with your fingers plugged into your ears.

If you know how to make a 400mm die, then you know how to make a 500mm die just like you know how to make a 300mm die. The rhetorical question was spot-on, unless you have some anecode to the contrary.

 

[Dave Baumann]

What's the difference between making a 500+mm die on one process node and, say, a 300mm on the next node? From an engineering perspective the former is easier.

 

[Davros]

no no no, 500mm on the next node, powerrrrrrrrrrrrrr (to quote clarkson)

 

[Wynix]

What's the difference between making a 500+mm die on one process node and, say, a 300mm on the next node? From an engineering perspective the former is easier.

I hope you don't mind me being a bit anal, feel free to point out any errors.

Cayman = 389mm
Tahiti = 365mm

Now if we estimate the same reduction in die size that Tahiti received compared to Cayman(6.16%) then we get a die size of 342.5mm

Also Hawaii is estimated to be between 424-440mm and not 500mm.
So a 342.5mm chip compared to a 440mm chip comparison seems more appropriate(if we are being conservative).

Now if we take TSMC's claims of 20nm being 1.9x the density of 28nm then the 28nm chip would have to be around 650mm to equal the 342.5mm chip; That is almost 50% bigger then what hawaii will be.
Even taking into account a size similar to Cypress(334mm) doesn't change things a great deal.

Don't get me wrong, i think 20nm was never a viable option and won't be one until mid 2014.

 

[Psycho]

What is Cayman doing in that story? If anything you should compare Tahiti to Cypress (early 28 vs early 40), and Hawaii to Cayman.

 

[Wynix]

What is Cayman doing in that story? If anything you should compare Tahiti to Cypress (early 28 vs early 40), and Hawaii to Cayman.

You are correct, i realized that after i typed it up and was too lazy to change it, but like i said it does not change things much.

334mm x 1.9 = 634.6mm, compared to the other figure of 650 does not make much of a difference.

Hawaii to Cayman

This is an interesting idea;

334>389 = 16.5% increase
365>440 = 20.6% increase

So the die size jump from Tahiti to Hawaii is only slightly more than the jump from Cypress to Cayman.

 

[Pete]

Nope, sorry FAIL.

A question as an answer to a question. Looks like a classic side step.

What is he side-stepping, a question for which you already have an answer and are just waiting to hear? Because your question was answered by both 3dilettante and silent_guy several hours before this outburst of yours.

Do you think AMD is unable to scale their GPU architecture to 550mm^2, that their perf/mm^2 will drop off? Because that's a different question than whether they can just make a bigger GPU with reasonable yields. (Though silent_guy already answered both.) I'm not sure what would make you assume they can't. They can scale Cape Verde to Pitcairn to Tahiti. Why would you think the hurdle to the next die size up is a question of smarts and not money (in that, they can make it, but is it the best use of their engineers and will there be a big enough market for it)?

 

[Dave Baumann]

I hope you don't mind me being a bit anal, feel free to point out any errors.

The point of the comment was that there is no "magic" to a big die, its just more transistors. Saying that "you can't do big dies" is basically saying "oh, you're stuck at x number of transistors and you have no future going forward". In other words, an absurd statement.

 

[Wynix]

The point of the comment was that there is no "magic" to a big die, its just more transistors. Saying that "you can't do big dies" is basically saying "oh, you're stuck at x number of transistors and you have no future going forward". In other words, an absurd statement.

Oh absolutely, i completely agree; most people completely ignore the costs associated with enormous dies, you eventually hit a point where the costs outweigh the benefits.

 

[Dave Baumann]

http://finance.yahoo.com/news/amd-webcast-gpu-14-product-120000041.html