LOSTCIRCUITS

The SLI Legacy

Some 7 years ago, another bottleneck was successfully overcome by 3Dfx Interactive (before they changed their name) by using two separate graphics cards in pass-through mode to enable so-called scan line interleaving or SLI. The core of this technology is the splitting of the images into odd and even scan lines that are rendered independently by two separate graphics subsystems, each comprising a GPU and onboard memory. After the demise of 3dfx (after they changed their name), the technology was inherited by nVidia along with the rest of the 3dfx portfolio.

Scanline interleaving in its classic form was somewhat doomed as overall system performance increased since the PCI bus was no longer capable of managing the bandwidth required for high resolution rendering. The accelerated graphics port (AGP) with its single port specifications precluded any further interleaving scheme anyway and, for a number of years, SLI was confined to a role as Sleeping Beauty.

In retrospect, it is somewhat ironic that in the battle about the upcoming interconnect specifications some three years ago, nVidia was in the HyperTransport camp, along with AMD, whereas ATI supported Intel’s proposal of the 3rd Generation I/O (3GIO), which finally became industry standard under the name PCI Express. The ironic part of this is that PCI Express allows the use of several parallel PCI Express Graphics (PEG) interfaces, one of the prerequisites for the resuscitation of SLI. Predictably, it did not take nVidia very long to come up with the necessary hardware and driver support – enabled indirectly by ATI.

One thing to keep in mind when talking about the modern form of SLI is that we are no longer bound to scanline interleaving. On the contrary, the new forms of SLI don’t even support this mode and to set things apart from past accomplishments, nVidia has coined the new term scalable link interface for the old moniker. SLI in its current form supports two distinct modes, namely Split Frame Rendering (SFR) with load balancing, which means that any given frame is split horizontally with dynamic shift of the border to warrant equal distribution of complexity and content over both cards. The second mode, dubbed Alternate Frame Rendering (AFR) is somewhat simpler in that each card renders one frame and skips the next – which is rendered by the second card.

The first reports on performance analyses of SLI are in, the overall verdict appears that, despite some immaturity of drivers, the average performance gain is somewhere between 50 and 80%, depending on application and resolution. In terms of product life cycle, this is equivalent to approximately 6-9 months. One hypothetical drawback of SLI is that it requires on the driver level a remapping of graphics resources over two electrically identical PCIe Graphics slots, in short: PEG16X to 2x PEG8X. This limits the bandwidth of the system interconnect to 4 GB/sec per card.

Given the fact that for all practical purposes - especially gaming, there is no benefit even for AGP8x over AGP4x, this sacrifice appears acceptable. Moreover, at present, there are no systems capable of delivering more memory bandwidth than 6GB/sec even in theoretical streaming applications. That means that even the 4GB/sec per card would compete for memory bandwidth under concurrent access conditions and in so far, the entire thing is a moot point.

Next Page:    => Back to Speed Bumps =>