Test 2: Overclocking
Although we didn’t discover any obvious advantages in using high-speed memory during the first part of our today’s test session, it doesn’t mean anything yet. High-speed memory like DDR3-1600 SDRAM is targeted primarily for computer enthusiasts and will most likely find its way into overclocked systems. In fact, during our previous high-speed memory tests we often noticed that its frequency potential can best be revealed during CPU overclocking. That is why the second part of our test session will be dedicated to performance analysis in an overclocker system.
In order to get as close to real-life conditions as possible, we overclocked our Core i7-860 processor to 4.0 GHz. Since the maximum multiplier for this processor is 22x under any load, we had to raise the base clock from 133 to 182 MHz in order to hit 4.0 GHz mark.
Turbo Mode technology, which changes the processor clock frequency multiplier dynamically, was disabled in this case, because this way we could achieve highest possible overclocking.
The increase in clock generator frequency caused certain changes in the list of supported memory frequency modes that can be set using available multipliers. Instead of DDR3-800, DDR3-1067, DDR3-1333 and DDR3-1600 SDRAM, the CPU automatically got support for DDR3-1092, DDR3-1456, DDR3-1820 and DDR3-2184 memory. The latter is obviously hardly possible at this point, because there are no memory modules in the market with these specifications. The remaining three modes are quite possible, which means that the currently available overclocker as well as regular memory modules can easily be used in LGA1156 systems with overclocked processors. We checked out all three of these modes in our second test session using various timings configurations that could be applied in each of these modes.
As usual, synthetic benchmarks come first.
As we have expected, during CPU overclocking the scattering of practical bandwidths and latencies of the memory sub-systems based on modules with different frequencies and timings grew bigger. Which is, in fact, not surprising at all, since the increment between different DDR3 SDRAM modes increases when you raise BCLK frequency.
We see a similar picture in another synthetic test called MaxMem2.
I would like to point out that processor overclocking speeds up the memory sub-system, even if the memory itself keeps working at the same frequency. For example, if we compare our results for DDR3-1092 SDRAM against the performance of DDR3-1067 from the previous part of our test session working at a close frequency, we will notice a significant increase in real practical bandwidth and lowering of the memory latency. There is nothing surprising here: it all comes from L3 processor cache and memory controller overclocking following the BCLK frequency increase.
However, the results of synthetic benchmarks are not so interesting for us, because they only give us an idea of theoretical advantages of high-speed memory in the most favorable testing conditions. As for the real-life performance, we can only explore it in real applications. Just like in the previous part of our tests, we will start with CPC Benchmark 2007.
Those applications where memory speed didn’t have that big of an influence on the system performance in nominal mode, do not seem to react that much to the memory sub-system settings during overclocking. However, in those tests where we did see certain influence, high-speed memory can work a small wonder. For example, during multi-tasking test the performance difference between the fastest and the slowest DDR3 SDRAM mode is over 15%, which is roughly speaking twice as much as the performance difference between the two closest processor models within a family. And even if we look at the total score in CPC Benchmark 2007, we definitely shouldn’t disregard memory sub-system settings. However, hitting the highest memory frequency at any price is hardly the best strategy here. In particular, you can significantly improve your system performance by simply lowering the timings of the memory working at lower frequency.