Initial Practical Experience With SLI
Our expectations and anticipation of SLI were enormous. Even back in March when putting together our PCI Express article , we speculated on future multi-rendering options that would be opened up by the new PCI Express bus. We were therefore quite thrilled when NVIDIA later made its first announcement that there would be such a solution.
With the arrival of the first SLI-capable motherboards our enthusiasm was quickly dampened however. It was literally raining crashes and the lab SLI configuration with the GeForce 6800 and 6600 GT PCIe cards either could not be recognized or the computer would crash after a few seconds of error messages. After many fruitless attempts to get an SLI system running we found out the reason: our cards were too "old". Our models were very early NVIDIA reference samples that according to the manufacturer do not yet function properly with SLI. As these cards had not been in stores for long anyway at the time we received them, no particular importance should be attached to this matter. With the latest 6800 Ultra and GT samples SLI worked right off the bat. That of course raises the question of whether it is at all possible to later purchase a second graphics card for SLI operation, perhaps even from a different card maker. At present we are unable to answer that question.
We also had big problems with the SLI motherboards. With the ASUS A8N-SLI it completely blew the BIOS every time we restarted the computer - a problem we encountered not only with our test board. Due to extreme testing time limitations, it was not possible to switch the board. There are only a few samples in existence anyway, though more than just those from MSI. The source of the problems turned out later to be a series of defective Bios components. A quick change-out made possible by the socket mount took care of the problems. After that the A8N-SLI Deluxe ran with no problems and at a good speed.
The MSI K8N Diamond, on the other hand, had major memory module problems with SLI operation. Our Kingston HyperX, set up for 2/2/2/6 DDR 400 timings, caused constant crashing. The computer would boot and run stably in pure desktop operation even with Prime95 with 2/2/2/6 timings, but crashed immediately in SLI operation with most games. In single operation, however, these problems did not occur. Only after setting the timing down to 2.5/3/3/7 did the MSI K8N Diamond run stably with SLI in operation. Our test sample also appears to have a problem concerning memory bandwidth. Even with 2/2/2/6 timings, Sandra displayed a memory bandwidth of only 4.9 GB/s, corresponding to only DDR 333! The ASUS A8N-SLI by contrast delivers just under 6 GB/s.
We also had to lower the memory timings for SLI operation somewhat for the ASUS board as well, as crashes occurred here too with 2/2/2/6 timings. With 2/3/3/6 timings the board then ran stably. We anticipate and hope that these problems are a result of the motherboards' revision status and will no longer occur with models that reach the stores. ASUS and MSI jostled each other over the past few weeks to be the first to introduce a working SLI motherboard. Under such conditions, problems are pre-programmed. What is striking though is that most of the instability occurred only in SLI operation.