AMD: R8xx Speculation

[Razor1]

hmm the rv870 is smaller then the rv770

 

[Silent_Buddha]

If true it basically means AMD is conceding the enthusiast sector to Nvidia, but I don't see how in the world Nvidia could even compete against this in the performance/mainstream sector.

Regards,
SB

 

[Jawed]

hmm the rv870 is smaller then the rv770

Apparently there's no such thing as RV870.

What's been shown could be a D3D11 replacement for RV770/790 or it could be a replacement for RV740 or it could be a half-way GPU that sits between the replacements for RV740 and RV770/790.

Jawed

 

[Silent_Buddha]

Apparently there's no such thing as RV870.

What's been shown could be a D3D11 replacement for RV770/790 or it could be a replacement for RV740 or it could be a half-way GPU that sits between the replacements for RV740 and RV770/790.

Jawed

True all that we know for sure is that it's a DX11 GPU, but not which one. If this is the mainstream/budget chip however, that's quite an accomplishment to have that basically ready to ship day and date with Rv870.

Regards,
SB

 

[JoshMST]

Yeah, I too was surprised when I posted the analysis at PCPer that the chip would be that small. I was expecting a 240 to 260 mm sq. chip as well.

However, I do think that it will be a bit more powerful than we are fearing. A lot of DX11 functionality was already integrated into previous designs (or so the rumors say), and I believe my estimate of 1.2 billion transistors is a lot closer than what other's (such as Anand's 1 billion) are saying. Then again, I have not always been as conservative as I probably should have been in previous speculation...

Anyway, considering that the RV670 was AMD's high end chip on the 55 nm process, and how small it was (compared to the R600 and G80 on older processes) I can honestly say I see this as a further example of AMD's chip philosophy for graphics. The first DX11 chips will have all the functionality, but with only a slight performance increase over current DX10 parts (such as the HD 4890). This will allow AMD to sell these parts at the $250 range (and make some decent money with fair margins).

Looking beyond this... will AMD have a refresh part on 40 nm? Or will it shuffle off the more complex design to GF's 32 nm bulk in 1H 2010? Considering that this DX11 part is already in silicon, we have to assume the design flows for the next gen part are nearly complete, how easy would it be to port that over to GF's 32 nm bulk design rules? Perhaps "easy" is not the word... but how long of a time would it take to implement such a change?

 

[Jawed]

True all that we know for sure is that it's a DX11 GPU, but not which one. If this is the mainstream/budget chip however, that's quite an accomplishment to have that basically ready to ship day and date with Rv870.

40nm appears to have mucked up a lot of stuff so this "relatively small" chip might be the only D3D11 warrior, whereas RV790's replacement was originally supposed to have been readying for launch about now.

RV770 was designed to be about the size it turned out to be but 2 clusters were added. The margin between it and this 40nm chip, something like 80mm², is so huge I find it hard to believe that this chip is meant to be the new king of AMD's hill.

Jawed

 

[LordEC911]

40nm appears to have mucked up a lot of stuff so this "relatively small" chip might be the only D3D11 warrior, whereas RV790's replacement was originally supposed to have been readying for launch about now.

RV770 was designed to be about the size it turned out to be but 2 clusters were added. The margin between it and this 40nm chip, something like 80mm², is so huge I find it hard to believe that this chip is meant to be the new king of AMD's hill.

Jawed

Answer to your question?

I guess the main thing to come away with is that there is no reason for AMD to show all their cards. Show enough to scare certain people and hype others, get a read off their response.

 

[Ailuros]

Whatever it is meant for, it sounds like a nice blow against Intel's LRB die show

 

[3dilettante]

If the performance gains are still modest, it is good news for Intel, since the peak performance will be a much lower bar to meet for Larrabee.

It would be fun to imagine a world where AMD could wodge three of those evergreen chips together.

 

[nAo]

It's a very wide chip, a 25%-50% ALUs increase plus a 10% clock increase would do wonders. I wouldn't underestimate it just because it's not that big.

 

[Ailuros]

It's a very wide chip, a 25%-50% ALUs increase plus a 10% clock increase would do wonders. I wouldn't underestimate it just because it's not that big.

I don't underestimate it at all; it's the competitions turn to prove in the foreseeable future that size does matter in terms of performance. Otherwise history will just repeat itself.

If the performance gains are still modest, it is good news for Intel, since the peak performance will be a much lower bar to meet for Larrabee.

It would be fun to imagine a world where AMD could wodge three of those evergreen chips together.

Hopefully desktop LRB will be able to beat that part, otherwise it'll be more embarrassing than any of us could imagine.

 

[3dilettante]

It's a very wide chip, a 25%-50% ALUs increase plus a 10% clock increase would do wonders. I wouldn't underestimate it just because it's not that big.

That missing ~80 mm2 of die space would have been a lot of performance margin to play with in the face of unspecified Intel and Nvidia competitors.

It may well be a decent performance bump over its forebears, as R600 was generally better than what it replaced.


The chip's small size does point towards some prioritization of die space, which would be critical for a Fusion chip that might only allocate space for one or two of those chip quadrants to the GPU.

 

[Ailuros]

We still don't know what kind of RV8x0 chip this is; it could be 870 but it also could be only a mainstream part.

 

[nAo]

That missing ~80 mm2 of die space would have been a lot of performance margin to play with in the face of unspecified Intel and Nvidia competitors.

Their architecture is very area efficient that it might very well be that this part is mostly bandwidth limited, and they simply preferred to not use anything more than a 256 bit bus.

The chip's small size does point towards some prioritization of die space, which would be critical for a Fusion chip that might only allocate space for one or two of those chip quadrants to the GPU.

Or even less

 

[fellix]

The chip's small size does point towards some prioritization of die space, which would be critical for a Fusion chip that might only allocate space for one or two of those chip quadrants to the GPU.

Small size and uber-dense transistor cells -- yay! Could it be that the ROPs have been finally ditched?

 

[Lukfi]

Just a wild theory here. We know that AMD had fully functional samples of RV740 in February, if not even earlier. By that time they probably also knew that TSMC botched the 40nm process. That was while they had "RV870" and "RV840" in development for the 40nm process (I know that they don't use those codenames internally, but people around here do). Then they learned that nVidia is several months behind them with their DX11 offering, so they decided to delay RV870 and use a different process node for it (who knows, either they went safe with 55nm, or risked GlobalFoundries' 32nm, but either way needed time to redesign the chip for the different process). RV840 was sufficiently small to have relatively good yields on 40nm though, so they started manufacturing it.

 

[neliz]

Newest Terrytweet.

they don't consider it a rock, but a Missile!

DX11 missile was launched in Taiwan - Mental note - get Catalyst drivers ready. Read and enjoy http://budurl.com/x3ly

That url leads to PCPer's piece.

 

[Davros]

couldnt find anywhere suitable to add his and didnt want to start a new thread
so here goes
Computex Taipei 2009: Even though the forthcoming “Clarkdale” based Socket 1136 Core i5 processors with on-chip GPU are not due to go on sale until sometime early next year, they had a little system running out at their Nangang booth

anyone know if this gpu is dx11 it would be a bit daft to release a dx10 gpu in 2010

 

[Davros]

Then they learned that nVidia is several months behind them with their DX11 offering,

found this quote "This steals some of the thunder out of nVidia’s claim that they will be the first to market with DX11 parts."

are they still caiming this ?

 

[3dilettante]

Their architecture is very area efficient that it might very well be that this part is mostly bandwidth limited, and they simply preferred to not use anything more than a 256 bit bus.

That would be quite the sweet spot to hit.

Just the right size to get the most out of the bandwidth of the bus, but also a size where adding additional bandwidth saving measures would have made adding more ALUs too large an additional die expenditure.