Stepped Performance of Nvidia nForce 680i SLI
Most of Xbit Labs’ readers should already be familiar with the new overclocking terms that arrived with the new CPUs and chipsets, such as FSB wall, FSB hole and FSB strap. We have already had a chance to come across all these factors directly affecting overclocking results.
FSB wall is the maximum bus frequency supported by the given CPU sample. This frequency cannot be exceeded by reducing the clock frequency multiplier even if the mainboard can run at higher speeds. For example, Intel Core 2 Duo E4300 CPUs that we had in our lab couldn’t go beyond 400MHz FSB no matter what we did.
FSB hole is a frequency interval when the mainboard doesn’t work properly including inability to boot the OS or even start at all. At the same time the mainboard is working fine at higher or lower frequencies. For example, 400-450MHz FSB frequency range on Asus P5N-E SLI or the 450MHz FSB frequency on our today’s hero - abit IN9 32X-MAX Wi-Fi with the older BIOS version.
FSB strap. In this case it is the frequency when the chipset switched to a different work mode. At the same time the latencies increase and the performance drops, however, the system can operate at higher FSB speeds. Asus mainboards on Intel P965 Express chipset switch FSB strap somewhere after 400MHz bus frequency. E4300 processor didn’t hit this value thanks to its relatively high 9x multiplier, while E6300 processor had to get beyond 400MHz FSB to overclock to the same resulting frequency. So, we witnessed a paradox: although E4300 supported lower memory and bus frequencies it turned out faster than E6300 that lacked overclocking potential to make up for the increased latencies.
The above described situation referred to Intel P965 Express chipset, and what about the FSB strap frequency for the Nvidia nForce 680i SLI and can the FSB strap term refer to this chipset at all? During the tests of Asus Striker Extreme mainboard we discovered that the performance dropped when FSB frequency changed from 420MHz to 425MHz and only when the FSB frequency reached beyond 440-450MHz this negative effect could be eliminated. But maybe this performance drop is not really typical of the nForce 680i SLI chipset and is only the peculiarity of Asus Striker Extreme mainboard?
Therefore we decided to test the read speed from the memory with EVEREST Cache and Memory benchmark in the entire range of FSB frequencies supported by abit IN9 32X-MAX Wi-Fi (up to 450MHz with 25MHz increment). This benchmark is very dependent on the memory frequency, reacts actively to changes in latencies and allows detecting the FSB strap easily. Although the mainboard works unstably at 400MHz+ frequencies, this benchmark can stand it. The memory was running with fixed timing settings of 5-5-5-15-2T.
I would like to stress that the obtained performance values are not the actual chipset performance. No one will ever use such high timings with low memory frequencies. We increased the timing settings on purpose, in order to ensure that the system would work fine in the entire frequency range from the nominal 266MHz up to 450MHz. However, this approach will allow us to detect the performance drop, if it occurs. Theoretically, the FSB frequency increases synchronously with the memory and the dependence should be almost linear. And if there appears a sharp curve anywhere on the graph, it will indicate a change in chipset work mode, the enabling of FSB strap.
Well, the beautiful thing about theory is that it is all so clear and logical. The practical results we obtained were truly shocking: we didn’t expect them to change so unevenly.
The results of our benchmark suggested that the mainboard worked fine only at FSB frequencies that were multiples of 50: 300MHz, 350MHz, 400MHz, 450MHz, and at frequencies that were multiples of 25 we observed unexpected performance drops. An additional test session was necessary, so we performed the same tests with a smaller increment of 20MHz. Now we could draw some more substantial conclusions:
So, according to our theoretical expectations, the mainboard performance is growing almost linearly in the interval between 266 and 320MHz FSB. At 325MHz we see a sudden drop in performance. Look at the second graph - EVEREST doesn’t measure latency that well and when we repeat the tests a few more times the results differ very noticeably. However, in this case the latency graph corresponds completely to the performance graph: the drop on one corresponds to the spike on another. After that the performance keeps growing together with the memory frequency and FSB frequency up to 350MHz, and then comes another drop, a more massive one this time – from 360 to 380MHz FSB.