Have you ever seen a track meet? It takes skill to "Hand Off The Baton" which basically means that both runners have to be running at the same speed for one runner to hand off the baton to the second runner and not lose any paces.
When the CPU and the DRAM are moving at the same speed then the data is handed off in the same way. On a 1:1 transfer there is no (theoretical) loss in transfer speed. However, there is ram that can not reach the speed of the CPU. This is especially so when overclocking.
For the purposes of simplicity please interchange FSB with HTT for Intel vs AMD. Although there are differences in the transport, for this explanation it will suffice.
A divider is what allows you to run the CPU FSB and DRAM (memory) at different clock speeds. Ideally, your CPU FSB and memory should run at the same clock speed (i.e., synchronous). However, it is possible to run the CPU FSB and memory at DIFFERENT clock speeds (i.e., asynchronous), but only within certain limits of the motherboard chipset. The available CPU FSB/DRAM settings are defined as a "ratio", or sometimes called dividers.
Let us stick to the actual reference speeds of FSB and DRAM. In other words an Intel 800MHz FSB is actually a 200MHz FSB quad pumped and a 533MHz FSB is actually a 133MHz FSB quad pumped. Let us stay with the initial clocks. So for example, suppose you have an Intel CPU w/ FSB of 133MHz (clock). You want to use a memory of PC3200, which is 200MHz clock. By default, your BIOS will typically configure your system for "synchronous" operations by downgrading the memory clock speed to 133MHz to match the CPU. This is called a CPU/DRAM ratio (or divider) of 1:1 (synchronous). Not very appealing when you consider that you've just lost 66MHz of memory clock speed. That PC3200 is effectively now PC2100.
Now suppose you'd like to fully exploit that PC3200 and run it "asynchronously" to the CPU, in other words, at its fully rated clock speed of 200MHz. To do that, the BIOS must provide an appropriate "divider", or CPU/DRAM ratio. Chipsets will typical provide support for a very limited # of dividers (1:1, 3:4, 4:5, 5:6, etc), and it varies chipset to chipset. In order to support, in this case, a CPU FSB of 133MHz (clock) and memory speed of 200MHz (clock), we need an appropriate divider. In this case, if the motherboard supports a 4:6 divider, the configuration is supported (and here's the calculation):
133MHz / 4 = 33MHz
33MHz * 6 = 200MHz
The motherboard uses the divider to tell you what relationships are configurable for the CPU FSB and memory, and you simply do the math as above to see what's possible. If you don't have in this case the 4:6 divider, you can't run a CPU FSB of 133MHz (clock) and memory speed of 200MHz (clock), end of issue.
But sometimes you can workaround the problem. Let's suppose we overclocked the CPU, to 166MHz . By doing so, we can now take advantage of the 5:6 divider, then we have:
166MHz / 5 = 33MHz
33MHz * 6 = 200MHz
By overclocking the CPU FSB we've manage to not only run the memory "at spec", that is PC3200 or a 200MHz (clock), we've even increased the CPU FSB for better performance. Naturally, this doesn't address the issue of CPU stability. But it illustrates how overclockers use the CPU/DRAM ratio (or dividers) to configure the motherboard to their liking.
Let's say the CPU is not stable @ 166MHz clock. We still have the 5:6 divider available, so we'll simply increase the CPU FSB a few increments at a time. As we do, we'll be also increasing the DRAM clock speed thanks to the divider. Perhaps we find that the CPU is stable until, say, 144MHz (clock). We apply the formula again and get:
144MHz / 5 = 28MHz
28Mhz * 6 = 172MHz (or, 172MHz * 2 (DDR) * 8 (bits) = PC2752)
This will not give us as much DRAM performance as we would like, however it is closer to PC3200 than before! Just for theory we can consider using the 4:6 divider instead:
144MHz / 4 = 36MHz
36MHz * 6 = 216MHz !!!
Now we've now overclocked the memory as well. This is a good example of why overclockers often buy performance RAM, with speeds of PC3500, PC4000, etc. The extra headroom in that memory allows the overclock, with the assurance of stability. When one runs the DRAM faster than the FSB there are some people who believe that they are achieving extra performance running dram at higher speeds than the CPU. To me this is like the secondary runner at the track meet running faster than the primary runner. Now with benchmarks (that are not real world) it might seem that the second runner is giving benefit because of the cache correction and latency but in reality this is not happening to my view.