There was a review i've seen long time ago that showed the performance delta was much greater between 2-3-2-5 and 2-2-2-5 than between 2.5-2-2-5 and 2-2-2-5. I cant remember where it was though. It was tested on a P4 system I believe.
The performance delta should be 2-3% I imagine, not more. Of course 3% here, 2% there, 4% another place and it all adds up (that is the logic here). But spending a lot of money extra on 2 vs 2.5 is not the best choice. In the time of rising ram prices, it is better to buy MORE ram than faster speed/latency ram. In the time of low prices, you should buy faster ram/lower latency.
If you have A64, since the memory controller is onboard, the difference should matter even less adn it has been shown that running memory at 1T vs 2T is a lot more important for A64 systems.
Finally, PC2100-PC3500 ram with cas 3 latency => usually poor overclockers, that is why many shy away from them. Also sometimes ram can run 2.5-3-3-6 on an AMD system and 2-3-3-6 on an Intel system, and other times you can buy the ram rated at 2.5-4-4-7 and it will run at 2-3-3-6 or vice versa so it all depends. Of course in theory 2-2-2-5 should be the fastest for Intel systems, 2-2-2-11 for Nforce2 systems and 2-2-2-10 1T for A64 systems.
Also Duvie and articles on the web proved that running lower latency (say 250FSB:200mhz RAM 2-2-2-5 (PC3200)) produces just as fast if not a faster system for P4 as opposed to running higher memory speed and higher latency (ie. 250FSB:250mhz RAM 3-4-4-8 (PC4000). So CAS timing alone wont affect performance as much as "overall" slow latency ram.