IntroductionAt this year's CeBIT the most controversial and most debated new hardware was NVIDIA's Quad-SLI, which was awarded "best product of CES" earlier this year. By now everybody should know about SLI which combines the rendering power of two video cards to achieve greater performance, or better image quality through higher Anti-Aliasing Levels. The logical evolution of this dual solution is using four GPUs. However, since there are no chipsets/motherboards which support PCI-E x16 on four slots, NVIDIA had to come up with a way to make their idea happen.
In the past we have seen dual GPU solutions on one PCB from manufacturers like Gigabyte or ASUS. But these were just GeForce 6600 designs which were not using such complex PCBs. The 7900 GTX is the biggest and baddest NVIDIA card around, of course it has a completely different range of requirements when it comes to power and signal stability.
So putting two GPUs on one PCB was out. The clever people at NVIDIA worked up a rather simple but ingenious solution - just stack two video cards, let them use only one slot connector and interconnect them with an own PCI-E bus. Each card has a little bridge device on it which connects the second card to the main PCI-E slot into which the card is plugged.
Add a second card, connect the two cards each with a SLI bridge and you have four interconnected GPUs which need only two slots.
However, since the power requirements of this little electrical heating are so immense, most Quad SLI systems are using 900W PSUs. At the NVIDIA press conference one NVIDIA spokesman was "actually pretty happy with the power consumption, we are only using 680W PSUs".
This brings us to another problem. Since the consumed power does not do any physical work, well except for the fans, it is dissipated as heat. And there's lots of it. NVIDIA uses a low-profile fan on each GPU. This means that each video card has its own fan, so the total number of coolers is four.
But to stay within their "thermal budget" NVIDIA had to reduce the clock rates of the 7900 GX2 down to 500 MHz Core and 605 MHz memory from the 7900 GTX's original 650 MHz / 800 MHz.
Each SLI card has 512 MB of video memory per GPU. Of course this means that one card consisting of two GPUs has 1 GB video memory onboard. But you can not just say "1 GB video card". The memories are completely separated. Each GPU has only access to its own 512 MB memory. There's no dynamic distribution based on need, like it is happening in your computer's main memory. This means that you can not use more than 512 MB video memory at one time in your application.
You really get something for your money. The PCB is the biggest consumer video card PCB I have seen in a long time. The small red card is a Sapphire X850 Pro.