AMD confirms R680 is two chips on one board

[Rufus]

Interesting they're using a PCI-e switch between the two RV670s. So much for synergy between AMD & ATi. HT would've been the obvious choice there. I guess it was cheaper to out-source it.

PCIe switch is the only sane way to do it. You need something talking PCIe off-board, and you need to talk to the two 670s. Dropping in a PCIe switch means it's basically just plug-and-play, fairly straightforward. Doing something like HT would mean redesigning the 670 to have an HT interface and making a PCIe->HT switch, and would gain you next to no performance.

ATI, Nvidia, and AMD all have enough PCIe and HT experience (from chipssets) to whip up whatever is needed. There's just no reason to.

 

[ShaidarHaran]

It's just disappointing not to see any fusing of the two companies technologies is all.

 

[AlexV]

No. D9E is scheduled for March. R680 is scheduled for the last week of January.

Cheers, didn't know that. But that doesn't place me totally out of left field with my statement. If it's scheduled for release in the last week of January that'll most likely mean availability(as in general case, not some boards that fly off the shelves for madmen like me ) somewhere around the half of February at best. Which leaves it about what...2 months tops before it goes head to head with the D9E?It'll still be hard, as it'll probably trade blows with the Ultra and the general mindest will be:just wait for the D9E.

If Quad CF actually ends up functional, and is faster than a single D9E, these may still sell for CF configurations for those who absolutely hate Intel mobos based on sucky nV chips(like me, again, and the 780i shows no promise of turning things around so we'll still be there), and use Intel based mobos who only give CF. But wheter or not QuadCF will actually end up being a realistic alternative is still up in the air, IMO, and I'm simply not getting a very positive vibe from the "AFR is preferred rendering method" noises. That means latency and chances of being screwed by devs(like Crytek, who just decided that they really really had to use data dependencies inter-frames and screw AFR up)...we'll see.

 

[Shtal]

So does it mean it won't be equal to two HD3870's in Crossfire, but more like overclocked, 512MB HD3850's in a "Crossfire-on-a-stick" configuration ?

With HD3870 approaching sub-200 dollar levels, two of them might end up costing as much or even less than this one, while providing additional performance (due to extra memory bandwidth, especially important in high resolutions).

Doesn't seem like anything that would stir my interest, based on these preliminary observations...

Source said the card is using 0.7ns GDDR4 memory.

http://www.techpowerup.com/

Maybe their would be two versions one with GDDR3 and the other with GDDR4.

Dual GPU
1. Radeon HD 3850X2 GDDR3
2. Radeon HD 3870X2 GDDR4

 

[CJ]

http://www.techpowerup.com/

Maybe their would be two versions one with GDDR3 and the other with GDDR4.

Dual GPU
1. Radeon HD 3850X2 GDDR3
2. Radeon HD 3870X2 GDDR4

There will be two versions alright. A 1GB version and a 512MB version.

 

[Shtal]

There will be two versions alright. A 1GB version and a 512MB version.

I was about to edit my post to say same thing: But I guess you reply quicker.

 

[Shtal]

If Quad CF actually ends up functional, and is faster than a single D9E,

Maybe single 3870X2 will keep up with single D9E.

Edit: In some benchmarks ATI Crossfire gain higher average FPS compare to Nvidia SLI.

 

[AlexV]

Maybe single 3870X2 will keep up with single D9E.

Edit: In some benchmarks ATI Crossfire gain higher average FPS compare to Nvidia SLI.

Maybe Santa will actually say hello this year...I've been waiting to see him since I was a kid. Then again, maybe not

 

[Shtal]

Maybe Santa will actually say hello this year...I've been waiting to see him since I was a kid. Then again, maybe not

It's not my fault that you didn't seeing Santa

 

[AlexV]

It's not my fault that you didn't seeing Santa

Yes it is, because you're making me say nasty things about the 3870X2 with your faith in it.

 

[nexus_alpha]

You would have to know the 'field of view' of the lense and then the distance the camera was placed from the board to be able to reverse calculate the precise length due to perspective barrelling.

The dvi connection and backplates are of standard length they can be used to measure the board length.

 

[AnarchX]

There will be two versions alright. A 1GB version and a 512MB version.

Do you talk about effective memory size or size of memory on PCB?

 

[hoom]

Erm, or you could use the picture of the board sitting flat for comparison maybe?

PS: Why is the PCIE switch so huge? Just a bigger process?

 

[satein]

Erm, or you could use the picture of the board sitting flat for comparison maybe?

PS: Why is the PCIE switch so huge? Just a bigger process?

Yes, I think so. It might be the 3rd party component available on the market that is suitable to the task.

There are also two possibilities I could read... First is if anything goes well, AMD/ATi might have got its own PCIe switch chip soon... Second is the next chip will not need the PCIe switch chip... thus this 3rd party component will be only for this HD 3870x2

 

[hoom]

First is if anything goes well, AMD/ATi might have got its own PCIe switch chip soon... Second is the next chip will not need the PCIe switch chip... thus this 3rd party component will be only for this HD 3870x2

I was thinking much the same.

 

[Jawed]

Yes, I think so. It might be the 3rd party component available on the market that is suitable to the task.

It looks like a PLX PEX8548, with 48 lanes (i.e. 16 lanes to northbridge and 16 lanes to each GPU):

http://www.plxtech.com/products/expresslane/pex8548.asp

based on its dimensions. This is sad, because this switch is PCI Express 1.1 not 2.0, while RV670 is gen 2.0, apparently throwing away half the available bandwidth that each GPU could support.

The gen 2.0 part, PEX8648:

http://www.plxtech.com/products/expresslane/pex8648.asp

is considerably smaller, 27x27mm.

Jawed

 

[Arnold Beckenbauer]

It looks like a PLX PEX8548, with 48 lanes (i.e. 16 lanes to northbridge and 16 lanes to each GPU):

http://www.plxtech.com/products/expresslane/pex8548.asp

based on its dimensions. This is sad, because this switch is PCI Express 1.1 not 2.0, while RV670 is gen 2.0, apparently throwing away half the available bandwidth that each GPU could support.
...

Jawed

I'm sure, it's not cheap. All Gemini Radeons use this PLX bridge chip (or an older version, eg GeCube X1650 Gemini with 32 PCIe lanes).

 

[ShaidarHaran]

PS: Why is the PCIE switch so huge? Just a bigger process?

Larger process, and I/O doesn't scale as well as logic with smaller geometry.

 

[Arun]

Uhm, you guys DO realize you're looking at the package, not the chip, right? This isn't a flip chip...

 

[ShaidarHaran]

Uhm, you guys DO realize you're looking at the package, not the chip, right? This isn't a flip chip...

So how big is the die? That's still a mighty big package... (that's what she said)