We start off the benchmark sections with a look at bandwidth and latency through the eyes of a few different programs. First up is the standard memory benchmark programs of recent memory, Lavalys Everest and SiSoft Sandra.

With the adjustments to how we handle tRD (Performance Level) for each set of results and the change to the DDR2 results, we weren't sure what was going to happen in the benchmarks. From these initial bandwidth tests though it looks as if there hasn't been much change to how things stack up. One item of note though is the drop in DDR2 bandwidth from the lower frequency that we previous benched at putting the G.Skill specified ratings well ahead of the FireStix in every measurement. The last three red results, however, appear to have been altered very little from previous testing where we left tRD the same for all results. The three overclocked results still scale with frequency but there does appear to be smaller gaps between the numbers.

The same pattern has carried over in the ScienceMark bandwidth and PCMark 05 memory suite numbers that we saw in Everest and Sandra. This includes the fact that the 6-6-5 results easily surpass both the DDR2 and DDR3 spec'd results which was previously not the case in other reviews. This could perhaps be the result of the 6-6-5 timing set clocking higher with these G.Skill HZs then any other kit but is also definitely influenced by the tighter tRD timing of 5.

The final set of numbers we will look at specific to memory testing are the latency results from Everest and Sandra. This is where the tighter tRD we used for the three overclocks really shows up and the overclocked results simply walk away from the top two blue and grey bars. We can also see that the spec'd results from the G.Skill HZs provide lower latencies than the DDR2-1000 setup which was a bit of a curve ball. We were expecting to see a bit more of a gap but going the other way. Perhaps it is a difference between the P35 and X38 chipsets but it appears to show the fact that frequency really does matter for latency, like the rest of the results indicate.