Latencies have a huge effect on realized memory subsystem performance, or "effective bandwidth" as you might term it. It is very hard to analyze bandwidth alone, as latencies will come into play any time your data objects are large enough to produce a high degree of row and column switching. The truth is the bandwidth is not truly changed by the latencies, but what we perceive as bandwidth is.
This is all really much ado about nothing though. The computer is not a device created to produce bandwidth. It is a device created to run programs. How well it does this in most cases is not hugely affected by raw bandwidth of the memory subsystem. There are exceptions to this rule, but in general both the effective latency characteristics of the memory subsystem along with the effective cpu power dwarf the effects of bandwidth on application perfromance.
Sandra buffered bandwidth is essentially just a lie, a lie that assumes we can actually realize the full potential bandwidth of a given memory type and speed, calculates what percentage of optimal performance it feels your memory subsystem operates at, and then multiplies the theoretical maximum by this percentage. As such is not a true performance metric, and should be essentially ignored. The unbuffered bandwidth figures are an actual measurement, where the buffered ones are a rating. And anyone who has read many of my posts knows how fond I am of ratings.
The important thing is to put any and all systhetic benchmark results in their proper context by comparing them to actual application performance. By doing this consistantly, you can develop an understanding of what synthetic benchmarks can tell you, and more importantly what they cannot. In this context bandwidth itself is shown to be a rather minor point, but as with any optimizing proceedure, if we can maximize bandwidth without paying other penalties (such as increased latencies) we produce an faster system as a whole.
Just like hard drive, cpu, video card, or basically any other PC performance area memory performance is the composite of the individual performance yielded in an essentially infinite variety of circumstances. It is difficult to place single-number values on such complex and variable quantities, even if the side of us that yearns for simplicity likes such attempts. So although the Sandra buffered bandwidth largely succeeds in removing the effects of latency upon the bandwidth figures, this is a pointless goal that moves further away from the real quantity to be maximized-application performance.
