I don't think it will happen that easy ...
If SiS ... err... I mean XGI succeeds to bring a solution worth to be take into consideration then the other players will consider it too .
As XGI's cards are the only present , real example of a multiGPU card I'm going to start off about them .
The GPUs have 80 mln transistors each which means that they'll have 160 mln transistors on each card . That's 30% more than nVIDIA has on their NV35 .
I will not compare these GPUs with the R3xx line from ATi as ATi uses 0.15 micron technology on their top of the line products and anyway the hardware implementation of the R3xx line is more than a wonder to me .. probably has to do with the cross licensing deals ATi has with INtel but that's another story .
So ... 30% more transistors translates into 30% more die size then it translates into ma more expensive GPU ( per wafer ) and also into lower yields .
Pay attention to these last 2 words "lower yields" that probably won't be the case with XGI as we are not talking about a 160 mln GPU but about more 80 mln ones ... less transistors higher yields so this differentiates XGI from nVIDIA in this case .
Conclusion 1 :
- the only disadvantage seems to be the fact that they have to pay more per wafer <or > get less chips (dual GPU card setups ) per wafer.
- the yields are likely to be higher than on nV35 for example ...
Now lets talk about the cost of the PCB ... NV35 uses a 256 Bit BUS PCB while the memory bus on the Dual XGI cards will be a 128 Bit one ... I don't think this will be a big cost issue ... most likely the price will be the same considering that the XGI Dual PCBs will be complex too .
The R&D costs fall down but not quite in the logic you would expect ....
Let's say ATi invested huge amounts of money and work into their R420 project ... the first 6 months from the launch we will have a refresh of the line with faster chips which will offer 15 to 20% more performance in the best case ... after 12 months ATi won't have to invest the same huge amount to develop a different architecture but it will launch a dual setup with the only R&D costs of the more complex PCB and some other issues like VPU interconnect BUS .
This kind of move will most like bring the 60%
(compared with the 6 month refresh - 92% improvement from the first product based on that architecture) performance delta desired from the new , yearly refresh of the product line .
The most important is that the R&D costs for the yearly refresh are now at 10% from the initial strategy .
Also , the manufacturing process is now mature so the cost per chip is also smaller to compensate from the total cost of the card ... with the more complex PCB .
So ... 24 months architecture cycle will also offer a more stable platform for the programmers and developers to work with .
This , in my opinion , can not go on forever like 3dfx thought : with 4x GPU setups and all .
Reasons :
-the amount of card memory ;
-the redundancy ratio of the setup ;
-cost per GPU ( a GPU will never cost like a low cost mobo chipset ) ;
-
the changes in the API and new implementations
For this ^ particular reason a R500 will most certainly appear even in this multi GPU scenario and ... just like the case with the changes in the architecture ... the performance will be greater than the older Dual setups .
Then ... we can all proceed to the next dual setup for another 24 months while waiting for DX11 .
The most important development in such a scenario :
Sure the first time , the dual setups won't bring anything extraordinary innovative BUT ... if this is the way to go ... from the second generation of such implementation the companies adopting such a method must
develop a method of SHARING the video memory in a most EFFICIENT way . This will be , in my view , the most important development as it would drive the price of the card something like 20% lower while bring more performance .
PCI Express will bring the necessary bandwidth for such devices along with the increased current supply for mid end solutions .
Conclusion 2 :
-I don't think anybody will ever use Dual card setups again , like in 2 slots in the mobo . It's too expensive and too inefficient as the second card won't have as much bandwidth as the main one and they'll have to use cables which are affected by loss of signal ... etc. .
Sounds like a step back to the Voodoo 2 days. Why does this have to come?