SLI - SH/SC Wiki

What is SLI?

by Dr. Fred

You might remember "SLI" as a technology used by 3dfx to get dual Voodoo 2 cards working in parallel in the same system. Back then the acronym stood for "Scan Line Interleave", as one card was responsible for even fields on the screen while the other was responsible for odd fields. This offered a pretty significant performance boost and allowed two Voodoo 2 cards to perform almost as well as a single Voodoo 3 in certain games.

Unfortunately, while this was a rather clever and inventive design, the advent of AGP made its continuation impossible: AGP is a port, not a bus, and as such only allows a connection from one point to another. Since there is no specification for an AGP switch, the only way to have dual AGP slots is to have a dual-CPU system with two northbridges. Considering the cost of SMP hardware tacked on to the price of two high-end video cards, an AGP-based SLI solution would simply not be economically viable.

With PCI-Express, however, the situation is reversed. It is now possible to have several high-bandwidth slots on a single, relatively inexpensive desktop motherboard, and as such it is only logical for SLI to make a comeback.

The new SLI, now pioneered by nVidia, stands for "Scalable Link Interface" and works rather differently from the 3DFX system. For one it seems to require chipset support to function, and unsurprisingly the only desktop chipset that currently offers compatibility with SLI is nVidia's own nForce 4.

Where on a normal PCI-Express motherboard one would have a single 16X slot accompanied by three or more 1X slots, SLI nForce 4 motherboards have two 16X slots with a peculiar SODIMM-like connector in between. The reason for this turns out to be fairly simple: instead of having two full 16X slots, which would require a massive 32 lanes just for the video, SLI boards have two physical 16X slots with only eight lanes each. The connector simply acts like a huge jumper that switches between an 8X/8X setup, where each physical 16X slot gets eight lanes, and a 16/0 one, where the primary 16X slot gets all sixteen lanes and the secondary one is disabled.

While halving the number of available PCIe lanes for each video card might seem like it would adversely affect bandwidth, an eight-lane slot still allows around 4000MB/s total, which is already twice as much as AGP8X. Therefore, video card bandwidth requirements would have to more than double before a single card could saturate an 8X slot (or a 16X physical slot with eight lanes).

Where the 3DFX SLI simply shared lines on the screen between GPUs, nVidia's SLI is implemented at the driver level and can run in two distinct modes that spread the load more evenly across GPUs: Split Frame Rendering (SFR), and Alternate Frame Rendering (AFR). To understand these modes, you need to know a little about how SLI handles rendering in the first place.

Raw geometry data is sent from the CPU to the first GPU (GPU1), which then forwards the appropriate load to the second GPU (GPU2). Both GPUs then receive texture data independently through the PCIe interface, and each render their individual frames (or frame slices). GPU2 then sends the contents of its frame buffer back to GPU1, which assembles the frame and sends it to the display. Transactions between GPUs are handled through an external SLI bridge, which offers much higher bandwidth and lower latency than simply using the PCIe interface.

With SFR, the driver determines load and assigns a slice of the frame to each GPU. Both GPUs then work simultaneously, until GPU2 forwards its frame slice to GPU1, which then sends the whole frame to the display. This is a relatively low-latency method which is great for fast-paced games and twitch FPSes (e.g. UT2K4, Counterstrike). The driver having to determine load for each frame does introduce a significant amount of overhead, however, which cuts down the theoretical maximum performance gain to 80%.

With AFR, both GPUs simply render alternating frames, which reduces overhead while increasing latency. As such, AFR is best suited for slow-paced games (e.g. Doom 3) where a slight lag between user input and video output is relatively unimportant.

Naturally, the fact that SLI relies on GPU1 for all display communication means that users looking to run more than two monitors on the same system will still have to invest in a third PCI-based solution (or PCIe 1X-based, if PCIe 1X video cards ever pop up).

So what are the practical implications of SLI? Well, first of all, SLI requires two identical video cards from the nVidia 6600GT, 6800, 6800GT or 6800 Ultra series to function, so using your old Radeon 9700 with that new GeForce 6800 is a no-no. SLI does offer interesting upgrading opportunities, however, as it is possible to buy, say, a 6600GT card now and add another one down the road to obtain performance somewhat close to a single 6800GT.

Simple arithmetic shows this to be a poor financial choice, however: a $150 SLI motherboard with a $175 6600GT adds up to $325; adding another 6600GT, assuming they drop to around $150, would raise the price to $475. This is nearly $100 more than the price of a single ATI X800XL ($290) and a standard nForce 4 board ($90) at the time of this writing, and the X800XL would offer better overall performance than two 6600GTs.

SLI is however an interesting option for top-end systems, as one can stack a couple of 6800GTs in a single machine and obtain performance that should compete favorably with the next generation of hardware. Here again, though, eventually replacing your 6800GT with a next-gen $350 card would probably yield better performance overall. The 7800GTX that came out recently is also priced at $599, and offers performance than a pair of ~$350 6800GTs.

It will be interesting to see what the future holds for SLI: ATI are working on their own version of SLI, which is called "Multi Rendering", and SLI compatibility will trickle onto the Intel platform with the nForce 4 Intel Edition chipset. Whether SLI is to remain a curiosity for high-end users, or to become a genuine upgrading path for those on a budget remains to be seen, however.