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

So AMD still hasn't figured out how to downclock the memory in multi-monitor setups. :???:
According to the review on ComputerBase, Nvidia's hardware downclocks only when two LCDs are connected (e.g. with 3 LCDs it consumes even more power than Radeons).
 
The investment is an AMD GPU. You can't even get it separately, which would help with adoption. Truaudio might help AMD sell GPUs as a neat bullet feature though. I bet the hardware is miniscule on 28nm so cost-benefit may turn out to be glorious for them.

Your point isn't clear to me. You painted it as if truaudio would suffer because people would need to spend more to make it useful (they don't. Which is different to G-sync, where you need to have a compatible monitor on top of it) or as if people will lose their investment if they decided to use Truaudio (they won't, it should be compatible with any dedicated audio hardware).

Sure, they need a compatible AMD GPU, but that is a entirely different issue when we can foresee that AMD will integrate truaudio on their APU.

IF you needed to buy a standalone card to get it then your point would be strong (you would need to spend extra for it), but it's probably cheap enough that AMD will integrate it on their entire portfolio so marketshare shouldn't be a problem in the future.
 
That schedule is from 2009. Hardly a predictor for current products considering all the mess AMD has been in.

Yeah, the current limits for bus width and clock rates have been exhausted. GDDR6?

I think some explanation can be derived from the slide itself.

It was betting on a foundry other than TSMC supplying wafers. So maybe the interposer process was supposed to be offered by GloFo. GloFo hasn't happened at all for GPUs in 2012 or 2013.
 
This might more properly belong in a Pirate Islands thread, but I think AMD will have interposers in 2014. It was supposed to happen in late 2012, but evidently it didn't.
(...)

But if you look at Hawaii, it's pretty clear to me that it needs to happen for the next big GPU. I mean, suppose its 20nm replacement has 64 CUs (4096 shaders), where is it going to get its bandwidth? AMD is already running a 512-bit bus with 5 GT/s RAM. Presumably they could push that to 6, maybe even 6.5 GT/s, but that would only be 30% more bandwidth at a very significant power cost. The alternative would be a wider bus, but 768 bits? That would be pretty damn costly, not to mention very large on the die.

So it seems to me that interposers need to happen concurrently with the jump to 20nm.

You're assuming that the next high-end GPU from AMD needs a lot more than 320GB/s. There's a good chance it won't.

Titan still sits at the top with only 288GB/s, and the GTX 680 still does very well with just 192GB/s.
Moreover, many recent games/benchmarks don't see much of a difference between the GTX 670 and the GTX 660 Ti, where both use the GK104 with one SMX disabled but the later has 25% less memory bandwidth and ROPs.

With 512bit, AMD got a bus so wide that they probably just paired it with the cheapest GDDR5 they could find and called it a day. And this was for a flagship GPU designed for 4k resolutions.
Furthermore, they've got a lot of headroom with the memory we have available today. Pair a Hawaii with the 7GT/s GDDR5 found in the current GTX770 and the memory bandwidth reaches almost 450GB/s.

So I think AMD is fine with their next top-end solution (1 year from now?) if they just stick to the 512bit bus and GDDR5.



What you could be asking is about their next mid/high end solution. How to get a mid-end GPU (20nm successor to Pitcairn?) to get the same performance as a R9 290X and fits into a cheap-ish PCB with a 256bit memory controller.
That's where the interposer might come. And before that, the APUs should be even more critical to get that upgrade.
 
Furthermore, they've got a lot of headroom with the memory we have available today. Pair a Hawaii with the 7GT/s GDDR5 found in the current GTX770 and the memory bandwidth reaches almost 450GB/s.
The size of Hawaiis IMC could be reduced by lower speeds, so 7GT/s is probably not stable (5.5-6Gbps seems to be a more possible range).

So I think AMD is fine with their next top-end solution (1 year from now?) if they just stick to the 512bit bus and GDDR5.
If they go for ~350mm² @ 20nm 512-Bit >6Gbps GDDR5 are probably not possible, through pad limitations.

Maybe we will see something equal at AMD: http://www.freepatentsonline.com/8489839.html
... IMCs on seperated dies, connected with broad interfaces on package, with lower pad-sizes (see Crystalwell with 512-Bit).
 
The size of Hawaiis IMC could be reduced by lower speeds, so 7GT/s is probably not stable (5.5-6Gbps seems to be a more possible range).


If they go for ~350mm² @ 20nm 512-Bit >6Gbps GDDR5 are probably not possible, through pad limitations.

Maybe we will see something equal at AMD: http://www.freepatentsonline.com/8489839.html
... IMCs on seperated dies, connected with broad interfaces on package, with lower pad-sizes (see Crystalwell with 512-Bit).


You'd have to find some proof for both of your assumptions...
 
You'd have to find some proof for both of your assumptions...

The first one is not an assumption, it's pretty much what Dave said. But Curaçao features 5.6 GT/s memory, and I think I've seen overclocked versions with >6 GT/s, so I would assume that the upper bound for PHYs of this size is somewhere between 6 and 6.5, probably closer to the former for a high-volume product.

Now, 6 GT/s memory would provide 20% more bandwidth, and that may well be fine for most games (so long as you don't reach very high definitions) but beyond 4K, there's a reason AMD went all out with bandwidth: compute. This is becoming a substantial part of AMD's graphics business, and I suspect it is (part of) the reason why AMD is making big GPUs again.

As for the second assumption, RV770 suggests that fitting a 512mm² bus on a ~350mm² chip would be difficult, but probably possible: http://techreport.com/r.x/radeon-hd-4870/die-shot.jpg
However, you might have to make it a rectangular design; which doesn't seem to bother Intel much.

2.5D stacking should increase bandwidth, save power and silicon area as well, and the combined effect would be greatly increased performance (cf. Iris Pro). If it can be done a a reasonable cost, it's a very big win.
 
Oh, dear, what's happening. AMD released the card at 550$ and nvidia responded today by slashing from 650 to 500$ the prices of GTX 780.

Waiting for AMD response- Radeon 290X for the real price of 369$
 
Nothing unexpected. Nvidia could not keep those prices. Titan completely obsolete. Talk about value loss. And its not the first time this has happend to Nvidia.
 
Nothing unexpected. Nvidia could not keep those prices. Titan completely obsolete. Talk about value loss. And its not the first time this has happend to Nvidia.

It's expected but incredible aggressive for a market leader.

Being 10 percent cheaper, having a 3 card game bundle(the shield discount is not much of an incentive) on top of the Nvidia brand is killer. Add in that there is stock and the overclocking headroom and AMD needs 290 now. The paper launch of 290x(a few thousand cards world wide) really hurt potential sales.

This is going to hurt Nvidia's margins. I think what hurts more than the die size of the chip is the cooler. That cooler looks insanely expensive.
 
It's expected but incredible aggressive for a market leader.

Being 10 percent cheaper, having a 3 card game bundle(the shield discount is not much of an incentive) on top of the Nvidia brand is killer. Add in that there is stock and the overclocking headroom and AMD needs 290 now. The paper launch of 290x(a few thousand cards world wide) really hurt potential sales.

This is going to hurt Nvidia's margins. I think what hurts more than the die size of the chip is the cooler. That cooler looks insanely expensive.

For what i have see on "leaked " bench of the 290 ( non x ), it match easely the 780 .. what do you want Nvidia do ? watch AMD release a 290x faster of the 780 and a 290 who match it for way less money.

We can speak about noise and cooling, but you will got a lot of "AIB " GPU's with different coolers enough soon. And i can allready bet the difference in price with reference models will not be high.

Im not so sure their margin are really hurts, and in all case, the 770 have allready been produced and sold since along time ( thoses are GK104 ), for the 780, they are allready sold since some months.. And i believe they think they will keep high margin on the TI version, who will cover the drop on price of the 780.

lets not forget, there's different things in question right now, AMD have been too aggresive lately in gaming developpement, price of the 290x ( we was not expect it at this price ), mantle, console hardwares etc etc.. Outside G-sync, it was allready surprising to see Nvidia trying to come with a library " ( of known component for game developpement made by Nvidia), Carmack insisting continusly on NvOpenGL extension ( without saying AMD is providing similar extension for OpenGL ). They are both aggressive. AMD was doing a Bundle, Nvidia will do a Bundle. AMD show "Game API", Nvidia will show something related, if AMD is at the same performance level, Nvidia will release a new card and for be competitive in price will decrease is price. AMD is showing 4K performance, Nvidia will show 4K performance.. its like that it work. Im really not suprised by it ( less of the guys and girls who have buy a 780 some weeks ago. )
 
Last edited by a moderator:
It's expected but incredible aggressive for a market leader

Why would you call a company with lower market share (actually nvidia has the lowest graphics market share amongst AMD, Intel and themselves) a market leader? Leader in what?

Anyway, current pricing for AMD won't work and they need something aggressive too...
 
Last edited by a moderator:
Say, hand out the bare chips for do-it-yourself replacement [solder iron for free!]? :LOL:
But, hu, that would be something ... sockets for GPUs ...
 
Why would you call a company with lower market share (actually nvidia has the lowest graphics market share amongst AMD, Intel anf themselves) a market leader? Leader in what?

Anyway, current pricing for AMD won't work and they need something agressive too...

Steam hardware survey usually shows a substantial marketshare advantage for nVidia. I guess they still win in the discrete card market and most important of all, the market of people who actually buy graphics cards for gaming.
 
Steam hardware survey usually shows a substantial marketshare advantage for nVidia. I guess they still win in the discrete card market and most important of all, the market of people who actually buy graphics cards for gaming.

They have the advantage in discrete market share on last quarter, AMD have even loose some points. But AMD was still on the 7000 series. in general they are in the 55-45 , 60/40%.. it was 38-62 this time.. its not like Samsung who have 3 times the market share of Apple lol ( their second ).... ( or windows vs Linux )... Its go up, its go down. When i read some peoples thoses last time on different forums, its like if AMD had 10% of the market and Nvidia 90..
 
Last edited by a moderator:
Back
Top