Recently motherboard BIOSes have usually been mentioned due to new mobo models and CPU revisions, and the majority believes the latest versions makes it all better or at least nor worse. But it turns out that not always a new BIOS means better performance. Sometimes even a mere second memory module slows down the whole system or a faster CPU starts working slower than its predecessor. As an example, here's the story about some AMD Athlon 64 Socket 754 motherboards and their BIOSes.

The first is ABIT KV8-MAX3 on VIA K8T800 with BIOS v2.2
Athlon 64 3700+ (as well as the new 3400+ and Sempron 3100+) got strangely low result in RMMA max. memory write test, just about 2,500 MB/s. At that memory read was close to nominal 3100+ MB/s. At that the old 3400+ (2.2 GHz, 1 MB L2) indicated correct read and write results. We found out that new processors have CG stepping for which AMD introduced a new memory controller setting named 2T Command Rate. This is officially required to use more than 2 memory modules at 400 MHz (the previous C0 allowed only 2 and decreased clock rate with 3). It turned out that despite there was only 1 or 2 memory modules, optimized settings enabled "by SPD" timings and automatically enabled 2T Command Rate that resulted in significant memory write performance decrease (despite it has no relation to SPD). Of course, these 20% loss wasn´t that obvious in real apps, but the fact that Athlon 64 3400+ 2.2 GHz 1Ì L2 worked better than Athlon 64 3700+ 2.4 GHz 1Ì L2 in Canopus ProCoder, WinRar and 7zip was at least alerting. Besides, latencies were also increased by 7-11%.

The same was with MSI K8T Neo-FIS2R (BIOS 1.2).
The second example was Gigabyte K8NSNXP on nForce 3 with BIOS F5. After optimized settings were activated (Auto memory clock), the second memory module resulted in clock rate decrease from 200 MHz to 157 or 160 MHz depending on CPU clock. Note that this BIOS had a separate 2T Command Rate setting (shown only with CG CPUs) and the Auto setting usually worked correct (could be disabled/enabled). The second, less serious trouble of K8NSNXP was ignoring memory interleave mode with two modules. All AMD64 CPUs can work in this mode. For RMMA synthetic tests interleave mode made 3-10% difference. Actually interleaving is bankwise, so even a single two-bank module can be slightly "overclocked". (That is actually done by ABIT KV8-MAX3).