Test results: 1333MHz FSB

Now we install a processor with a 1333MHz FSB. As far as memory bandwidth is concerned, dual-channel DDR2-667 memory is sufficient at this frequency. As standard modifications of DDR2 memory cannot even get close to this FSB frequency, we'll focus on DDR3.

The speed of reading data from memory still grows together with memory frequency (up to 1333MHz), even when timings are increased disproportionately (CL7 in DDR3-1333 versus CL5 in DDR3-1066). But memory operating at 1600MHz does not yield performance gains, and reduction of absolute timing values is of no help here.

However, relative memory write rates are a tad different, but only in the last case, when raising memory frequency to 1600MHz increases performance.

Results of the memory read latency test are closer to our theoretical computations of timings: modes with lower absolute timings are at advantage here. As a result, higher-clocked memory will always be faster, but only because it has lower timings.

Multi-threaded reading is still a tad faster, while multi-threaded writing is a tad slower. Test results coincide with those for single-threaded memory access to the same degree.

You will hardly be surprised that our synthetic test results hold true in real applications; generally speaking, the question was whether DDR3-1600 memory with lower timings would manage to outperform DDR3-1333. Practical results civilly skirted this question, leaving us to evaluate the statistical measurement error. We may say that these modes demonstrate the same speed.

Now what concerns real differences in real applications. 7-Zip apparently gives preference to the NVIDIA chipset, so we have two comparison options: Intel X48 with DDR3 is faster than DDR2-667@4-4-4-12-2T by 5.5% at best. NVIDIA nForce 790i Ultra demonstrates similar results in comparison with the slowest DDR3 mode. If we analyzed unofficial modifications of DDR2 memory, we could get a higher performance gain with Intel X48, because DDR2 would work faster with it, and memory frequency would be specified regardless of its type. Maximum performance gain (from available nominal ones) with X48 in Doom 3 was almost 7%. It's lower with the NVIDIA chipset, but its minimal mode is faster.

This group of tests confirms our conclusion about advantages of the faster memory. We just cannot pinpoint the upper limit for sure: the frequency of 1333MHz is sufficient. Well, at least you shouldn't worry about performance drops with DDR3-1600 memory operating at normal timings.