NVIDIA SLI: How Does It Work?
First of all, you should know that NVIDIA’s SLI means Scalable Link Interface rather than 3dfx’s “Scan Line Interleave”, so there are fundamental differences between the two technologies, despite the identical abbreviation. On the other hand, we can see some traces of the Voodoo Scalable Architecture in NVIDIA’s SLI. For example, one of the cards in an NVIDIA SLI configuration is a master and another is a slave, and the master is responsible for assembling the final frame and outputting it onto the monitor. Yet there are some significant differences: NVIDIA’s multi-GPU technology uses its own methods of distributing the load between the graphics cards, and these methods differ greatly from 3dfx’s line interleave.
NVIDIA SLI offers two operational modes for a higher performance: Split Frame Rendering (SFR) and Alternate Frame Rendering (AFR). A third mode when only one graphics card is working is added for compatibility with older games.
Split Frame Rendering (SFR) means that the frame is split in two and the two graphics cards each renders its own part of the frame. The gist of the technology is that the frame is not necessarily split in two equal parts – their size may vary dynamically depending on the complexity of the scene. In the latter case the efficiency and performance would be the highest as each graphics card bears almost the same load. NVIDIA refers to this mode as Symmetric Multi-Rendering with Dynamic Load Balancing (SMR). So far they are only talking about uniting two NVIDIA GPUs, but four and more GPUs may be supported in the future.
Alternate Frame Rendering (AFR) means the frames are being rendered alternately by each of the graphics card, similar to ATI’s MAXX.
The SMR mode is, however, the most thrilling thing. According to NVIDIA, the intelligent algorithms of dynamic load distribution make it possible to achieve a performance gain of 80-90 percent in the SLI mode. That’s a brave claim indeed considering that the efficiency of Video Array technology developed by Alienware is from 40 to 70 percent. The downside is that a scrupulous optimization of the drivers and applications is required to achieve such a big performance growth in the SLI mode. We can compare this to computers with multiple CPUs: their performance gain depends on how well the particular application can distribute its threads among different processors. Aside from the topic, we want to note that the number of programs optimized will depend on the popularity of multi-core central processors.
But let’s get back to NVIDIA’s SLI. We will see later on if the claim about an almost double performance gain has anything substantial in itself, but now we will discuss the technical aspects of the implementation of NVIDIA’s multi-GPU technology. Right now the SLI mode is supported by NV45 (GeForce 6800 Ultra/GT) and NV43 (GeForce 6600 GT) graphics processors. The die of each of these GPUs includes a special MIO unit responsible for connection and synchronization of the GPUs.
Considering the fantastically high bandwidth of the PCI Express interface they might have used it to transfer the necessary data instead, but NVIDIA wanted to get the utmost from SLI technology and minimize possible latencies and performance hits by equipping its new GPUs with a dedicated interface. The MIO slot located on the graphics card looks like a PCI Express x1 slot, but somewhat narrower. This slot is probably also responsible for the Frame Lock function, i.e. it avoids the frame tearing effect as a result of GPUs falling out of sync. By the way, the inexpensive mainstream GeForce 6600, although is based on the NV43 core, is officially devoid of SLI: the necessary slot is just not wired on the PCB.
When the graphics cards are plugged into the mainboard’s slots of the PCI Express x16 design, they are then connected with a special adapter. The adapter is a small PCB with two corresponding connectors.
Here’s how it looks when the cards are in the system and joined with the adapter:
It takes meeting a few requirements even to enable the SLI mode, not to mention to make it work correctly. You’ll learn about these requirements in the next section of this review.