As seen in previous nVidia nForce chipsets, the new nForce 4 is a single chip solution. This means that instead of dividing the chipset into a Northbridge (for video and memory input which communicates directly with the processor) and Southbridge (peripheral and drive input, communicates with Northbridge), all functions have been placed on a single integrated core logic circuit.

This design innovation helps to reduce data bottlenecks by eliminating the data bus between the Northbridge and Southbridge chipsets completely. Given the unique architecture of the Athlon 64 processor, in which the memory controller is on the CPU itself rather than the chipset, this kind of one-chip solution makes a lot of sense.

nVidia's nForce 4 platform is currently available in three distinct variants, with a fourth version apparently following in the near future.

The three main variants available now are:
nVidia nForce 4 - a socket 754 solution which supports both Athlon 64 and Sempron processors
nVidia nForce 4 Ultra - a solution for Athlon 64 and FX processors in the socket 939 formfactor
nVidia nForce 4 SLI (Scaleable Link Interface) - a solution for socket 939 Athlon 64 and FX processors, featuring dual PCI-Express slots for dual SLI operation with compatible video cards.

As you might expect, there are abundant similarities between each of these nForce 4 chipsets. Each of the three versions currently released handles different situations, but each also offers a similar feature set to the user, the biggest of which is PCI Express.

nVidia's new Scaleable Link Interface (SLI) technology is used to link two nVidia based cards together, splitting the rendering load between them to increase 3D performance. The technology requires a pair of compatible videocards (Nvidia Geforce 6600GT models and above) with SLI connectors (must be implemented by the video card manufacturer) and an Nforce 4 SLI chipset-based motherboard.

Typical PCI-Express-based motherboards use the PCI-Express x16 slot to interface with video cards. As you'd imagine, this provides 16 PCI Express lanes to the single card for a total available bandwidth of 8GB/s. The Nvidia Nforce 4 SLI solution provides two physical PCI-Express video slots, and uses a switch to divert 8 PCI-Express data lanes to serve each slot. A single card can also be used in either slot, and in this case the full 16 PCI-Express lanes are available. In a typical SLI solution, the cards themselves are also linked by way of an SLI cable attached to the special MIO 'video bus' connector on the top of each card.

In the nForce 4 motherboards we have seen, the SLI switch is implemented on a small card which must be physically switched around to go from 'normal mode' in which the full 16 lanes of PCI-Express goodness are available to a single card and 'SLI mode' in which 8 lanes are directed to each physical slot. We're not sure if this can be replaced with an auto-sensing switch or not, but we hope so.

Nvidia's SLI works by allowing the two graphical processors to share the rendering workload, governed by the Nvidia Detonator software drivers. The CPU passes all neccessary 3D information to the 'primary' GPU, which then shares the information with the second card via the video bus interface cable. This removes the overhead of synchronizing the two processors from the PCI-Express bus, allowing improved performance. The video bus link itself apparently runs at up to 10GB/s, though we doubt that this bandwidth is fully utilized.

Currently, the only nVidia SLI-compatible video processors are the Geforce 6600GT, 6800, 6800GT and 6800 Ultra. The graphical processors in each video card must be identical, as must the video BIOS revisions, though the cards can run at separate speeds (the SLI system will assume the lowest clock speeds for both cards). This means that it is going to be pretty much essential to have two identical cards from the same manufacturer to get SLI working correctly. Nvidia has introduced a certification program to ensure that users can find compatible products.

The actual SLI rendering process uses one of two modes: Alternate Frame Rendering (AFR) and Split Frame Rendering (SFR). AFR has each video card render a separate frame, while SFR, the method that has gotten more publicity, uses each GPU to render part of one frame. Interestingly, the choice of which method to use in which games is pre-programmed into the Detonator driver suite, meaning that if there is no existing profile for the game you are playing, SLI will not work with that game. In these cases, a compatibility mode is used, which cuts off the SLI process completely, using only a single GPU (and we'd assume only 8 PCI-Express lanes) for all rendering tasks. Nvidia claims that they have already created profiles for more than 100 of the most popular 3D games, and more will follow with Detonator driver updates.

The Split Frame Rendering mode is probably the most interesting part of Nvidia's SLI technology. Using the Detonator driver to balance and allocate the video load, each GPU shares about half of the rendering work for each frame, then the completed frame is assembled by the first primary GPU and output to the PCI-Express x16 bus. Obviously this will not be 100% efficient, as different parts of each graphical frame will vary in complexity and some overhead is added in assembling the frame at the end, but overall this method should result in a considerable performance increase. You can expect CPU load to increase as well, since the Detonator software is responsible for balancing the video load to each card at all times.

Alternate Frame Rendering mode, where a frame is rendered separately on each video card, should give even higher performance, but this technology cannot always be used on modern 3D games due to certain graphical effects which require multiple frames to be blended together. Split Frame Rendering has no such limitation as both cards are always working on a single graphical frame.

The major benefit of Nvidia's SLI is its ability to more fully utilize the massive bandwidth of the PCI-Express x16 video solution. A pair of GPUs can process information twice as fast (minus the overhead of the communication between them) and use the available bandwidth more efficiently, considerably boosting 3D performance. This should also enable users to get top-tier performance out of a pair of mid-range 6600GT cards. Interested users should note that having two videocards also considerably increases power consumption, and using a pair of 6800 Ultras will mandate a hefty power supply.