NVIDIA's new SLI uses a high-speed digital link between the graphics cards, and the cards talk to one another using a proprietary inter-GPU communications protocol. Support for scalability resides in the MIO portion of the NV40 GPU, alongside the units responsible for various forms of video output, like the TMDS transmitters and RAMDACs.
SLI support resides in the display portion of NV40. Source: NVIDIA.
NV40's scalability support isn't limited to just two GPUs, either—larger groups of GPUs are possible. The first (and perhaps only) implementations of the new SLI will be limited to two GPUs, however. The cards will operate in a deeply politically incorrect master/slave configuration. The master card will produce the video output, while the slave card feed its output data to the master. The slave card need not be connected to a monitor at all.
One of the big challenges for those assembling NVIDIA SLI systems will be finding the right motherboards that will support two graphics cards. The only workable motherboard options so far appear to be workstation-class dual Xeon boards. I have a hunch we should see some more specific information about these motherboards announced today. These boards have a pair of PCI Express X16 slots, but the electrical connections to the second PCI Express slot are only eight lanes, or X8 bandwidth. That's still gobs of bandwidth—effectively double what you'd get from AGP 8X—so it shouldn't be a problem. NVIDIA says the card in the full PCI-E X16 slot will be the master card, and the other will operate as the slave.
The folks from NVIDIA claim SLI technology will be completely transparent to applications, including games. The company has a number of patented and patent-pending algorithms its graphics drivers will use to distribute the workload between the cards. Fundamentally, though, the load-balancing scheme is simple: split the screen into two parts, and let each card render its own portion.
Workloads are split between GPUs dynamically. Source: NVIDIA.
The actual split between the cards will vary depending on the graphics loads involved. Sometimes the split may be 50/50, sometimes 70/30. NVIDIA says its load-balacing algorithms will key on multiple limiting factors in graphics, including fill rate, geometry loads, and pixel shader processing power. The CPU overhead for all of this magic, claims NVIDIA, should be around 1%.
The performance prospects are awe-inspiring. Based on some testing in 3DMark03 and the Unreal 3 engine, NVIDIA anticipates that a pair of cards running in SLI should achieve 1.87 times the performance of a single card. Considering that the GeForce 6800 Ultra is a contender for the title of "fastest graphics card on the planet," a GeForce 6800 Ultra SLI rig may cause soiled britches.
Before you buy a 72-pack of Depends at Sam's Club, though, remember that SLI can only deliver such performance gains when certain kinds of graphics bottlenecks are limiting performance. In the days of Voodoo 2 SLI, the one and only real gain from SLI was more fill rate, which means more pixel-pushing power for painting higher resolution images at higher frame rates. NVIDIA's SLI can, of course, offer more fill rate for insanely high screen resolutions and uber-sampled antialiasing, but it can also help crunch through additional polygon complexity or more intensive pixel shader effects. Just realize that the application has to be graphics bound, not CPU limited, in order to benefit from SLI.
