Test 1: Nominal Mode
First of all we ran some tests in the system nominal mode, when none of the components were overclocked. We only changed the memory frequency multiplier and timings. I have to say that during this test we tried to emulate the most typical operational conditions that is why we decided not to disable any processor technologies. Hyper-Threading, Turbo Mode and Enhanced Intel SpeedStep worked the way they were supposed to: the operating system saw our CPU as an eight-core one, its frequency dropped to 1.2 GHz in idle mode and under operational load of different intensity it increased to 3.46, 3.33 and 2.93 GHz.
First of all, we checked out the results in synthetic benchmarks, which measure the performance of the memory sub-system separately from the rest of the platform. Lavalys Everest 5.30 is a perfect example of a test application like that.
As you can see from the obtained results, memory sub-system settings do have some effect on the measures system parameters. As the memory frequency increases and the latencies drop, we see that the practical bandwidth starts growing and the practical latency – decreasing. As a result, the performance difference between DDR3-1067 and DDR3-1600 in Everest test reaches 18%.
We also used another synthetic test called MaxMem2. The peculiarity of this test program is that it not only measures the memory bandwidth and latency in single-threaded mode, but also can estimate the memory performance in case of multi-threaded operation.
Overall, we see almost the same situation as in Everest. However, the results in multi-threaded mode are in fact pretty interesting. According to the numbers shown on the corresponding diagram, the memory frequency appears way more important than timings, when multi-threaded memory access takes place. DDR3-1333 shows more than 20% advantage over DDR3-1067, while DDR3-1600 appears almost 40% faster than a similar system equipped with slow DDR3 SDRAM.
However, the above mentioned numbers are the results of synthetic benchmarks, which are not directly connected to the real-life memory sub-system performance. Therefore, let’s go over to the real applications. To estimate the average performance in popular applications we used CPC Benchmark 2007, which measures the performance in GIMP graphics editor, during video transcoding using Handbrake utility and under multi-threaded load created by simultaneous HD video playback and data compression in 7-zip archiving tool.
And this is where we see very clearly that in real life things are completely different from what we have just seen in synthetic benchmarks. In many cases fast DDR3-1600 memory provides only an imaginary advantage over DDR3-1333 and DDR3-1067 platforms: very often the results differ by less than 1%. Very often, but not always. For example, we see in the multi-threaded test that faster memory may still have a noticeable effect on performance. By raising the DDR3 SDRAM frequency only one step higher we can increase the system performance by the good 4-6%. And as you can notice, this frequency increase causes a more substantial performance improvement than lowering of the timings.