LOSTCIRCUITS

Status Confirmation - The Need for a Link Interface

So what else was there? We mentioned earlier that there is a need for one card knowing what the other is doing, especially with respect to synchronization and, by extension, for load balancing. Sure enough, this can be accomplished through the same PCIe slots, at least in theory, however, nVidia is going a different way by using the mysterious edge connector. This edge connector is essentially a full duplex I/O bridge connector, meaning that each input on one card is connected to an output on the other card. We used our trusted multimeter to ohm out the different traces and to no surprise, all “gold fingers” on the upper side of graphics card #1 are routed to the corresponding “gold fingers” on the lower side of the second card.

We traced all connections between the individual contacts and color coded them for your viewing pleasure. The green group maps 1:1 on both sides of the link interface, as does the purple group.

On the board itself we could trace some of the traces but there is no more conclusive evidence as to what the bridge is for than to look at its actual role in the SLI functionality in real life. In essence, what we wanted to know is whether SLI would work at all with the bridge removed and the short answer is yes. However, things are not that simple, as we reckoned before, the bridge is necessary for synchronization and exchange of status reports between the two graphics adapters. Ergo, with the bridge removed, and SLI enabled, we should still be able of running in multi GPU mode. However, we’d expect SLI to be somewhat out of whack and we in the load-balancing toolbar, together with some performance metrics, we have the perfect tools to show what is going on. In other words, with the cross talk between the cards screwed up we expect the two cards still working, however, against each other and with a substantial overhead going into cleaning up the screwed-up situation. Performance in that case could range anywhere from below a single card to somewhere in between single card and functional SLI mode.

       

The two GeForce 6600GT in action with the link connector in place and the "redirector" in dual graphics card orientation --- removal of the link interface results in an error message, however, multi-GPU processing is still functional, albeit not at optimal performance --- Load balancing, which in the optimal scenario should place the horizontal separation line in the center of the screen is completely off if the link interface is removed.

Average FPS in Aquamark3 with the GeForce 6600 GT running in single card configuration and in SLI mode. In the SLI** configuration the Link Interface was removed and load-balancing was basically non-existent (as shown above).

Enough about the intricacies of load sharing and bridges, what we are really interested in is the performance that can be gained by moving to an SLI system. Keep in mind that there are two different modes in which SLI can run, that is we have split frame rendering (SFR) in which each card renders one portion of the image to be displayed, and then there is alternate frame rendering (AFR) where every card renders one frame and then skips the next. What is necessary to keep in mind is that the two modes are predefined for each application and cannot be swapped. That is, an application such as FarCry using SFR cannot be forced to run in AFR mode.

SFR vs. AFR

Very briefly, each mode has its potential and real life advantages and disadvantages. Advantages for SFR are that each card will at least in theory only render a smaller portion of the output screen, meaning that theoretically the on-board frame buffer can be smaller. In reality, this is not the case, the entire geometry needs to be loaded, the same goes for the textures, and it is the clipping that reduces the load on the GPU by reducing the size of the frame to be rendered. If one frame is very asymmetric in terms of its contents (for example jungle at the bottom and blue skies at the top), the demarcation line can be shifted up from the center to even the distribution of what every card has to render. Unfortunately, this again requires that every card knows what the other one is doing, meaning a substantial overhead.

AFR, on the other hand does not provide any advantage in terms of the overall memory space of a single card. Both cards will have to contain the same amount of data and instructions, geometry setup and textures and in so far, performance will be governed by similar rules as in a single card scenario. On the other hand, there is absolutely no need for any rocket science when it comes to load distribution, it is basically murder by numbers – odd – even – odd – even…

Amongst the issues that have plagued early reviews of SLI configuration was the fact that the graphics cards examined were already too powerful to scale appropriately with doubling of resources. This is where the GeForce 6600 GT comes in as a new metric to look at SLI. Arguably, the 6600 GT is by itself rather powerful, yet, there is still a world of difference to the high end cards from nVidia and ATI. On the positive side of things, the difference is not restricted to the performance but inevitably has found its way into the cost associated. At US$ 200,- a pop, these cards are affordable and, even more important, they are available. Let’s see what these little gems can do, in a pedestrian single configuration as well as in SLI mode.

Next Page:    => GeForce 6600 GT and Competition =>