I had quite a luck in getting nice RAM chips few years back. I understand that DDR at 1600MHz is considered bit outdated today, but all I want now is to demonstrate how much the memory timings matter to overall system performance.
ASUS M5A99FX PRO R2.0
AMD Phenom II x6 1090t @ 3800 Mhz / NB 2800Mhz / HT 2600 Mhz
A-Data XPG 2.0 /w Elpida Hyper MNH Chips @ 1600Mhz
Some of you may recognize Elpida Hyper chips, as those were first DDR3 to get through 2000Mhz. Usually, latencies were loosen, voltage increased from 1.9v (which was "stock for Intel"*) up to 2.1v with additional cooling.
The setup above has certain limitations. Dualchannel configuration at 1600MHz should theoretically allow 25.600 Mb/s, however Read/Write test from memory would not go above 11 000mb/s. CPU is limited by Northbridge and Hypertransport bus. PCI-E 2.0 x16 is limited to 8000mb/s in one direction, while real transfers CPU>GPU would be around 5.500mb/s.
Positive on the configuration is that when you add up all bandwidth utlization from CPU and PCI-E, total RAM bandwidth is still higher. Therefore increasing in frequency has little impact on total performance, but latencies still do affect total performance.
For measuring the impact of DRAM Timings i used LinX 0.6.4, with problem size 9992 / 771mb (Smaller the block, the bigger impact of short latencies, this one is "medium" sized). No other processes are running on system, while LinX is set to run on 6 physical cores, with priority above normal. During the 2nd and 3rd test were RAMs set to 1.80v. Its above normal, but less when compared to 1.9v those chips were used to run. In theory, even when Bandwidth of RAM is not changing, shorter latencies should allow more IOPS through RAM.
For the tested frequency of 1600Mhz was common setting of 9-9-9-27 for cheaper Dimm modules. Good modules usually allowed 7-7-7-21. If you were really lucky you could go as low as 6-6-6-18, and maybe slightly lower.
Because its Windows 7 x64, there might be tasks executing in the background. To make results reliable, the test ran 10x for each configuration.
First test was performed at "stock" settings for the mainboard. 1600Mhz 9-9-9-27 2t
Best case scenario was 9.382 seconds and 70.9113 Gflops.
Worst case scenario 9.415 seconds and 70,6587 Gflops
Second test was performed with custom tuning. 1600MHz 6-6-6-18 1t.**
Best case scenario was 9.213 seconds and 72.2083 Gflops.
Worst case scenario 9.238 seconds 72.0148 Gflops
I also performed 3rd test with CPU frequency increased from 3800 to 3900Mhz, while keeping memory on 1600Mhz 6-6-6-18 1t
Best case scenario was 9.014 second and 73.8045 Gflops
Worst case scenario was 9.207 seconds and 72.2584 Gflops
However as you see, this test was affected by some running process, so i will take 2nd worst result instead - 9.029 seconds and 73.6841 Gflops.
Increase from memory tuning is 1.297 Gflops if comparing best results, and 1.3561 when comparing worst results.
Increase from CPU overclocking (6x Phenom II cores, from 3800 to 3900Mhz) is 1.5962 Gflops when comparing best results, and 1.6693 when comparing 2nd worst result.
The more cores the CPU will have, the higher impact on test results. Memory tuning in this case was slightly less than provided 6 Phenom II cpu cores, Overclocked by additionnal 100Mhz, it approximately gave as good results as 5 cores.
Its worth noting that shorter memory latencies do not apply only to CPU tasks (as in case of LinX) but to overall system performance as those affect all devices connected to the motherboard. CPU frequency does not have that wide impact. Typical tasks such as 3d rendering or networking usually work with much smaller data blocks than 771mb, so the impact in such cases should be bigger.
* JEDEC standard for DDR3 voltage is 1.5v. First DDR3 memory controllers on Intel board were quite bad... so they used 1.9v instead.
** There was actually some more memory tuning than just that, but it would be bit too many numbers
Edit: I missed "Memory" forum. If some mod goes around, please move it. Thx.
Also one note...
The test above demonstrates that tuning the memory latencies can get about 2% CPU performance. Many reviewers do "fair" CPU comparisons in a manner that they use same memory sticks and use "Auto" settings on two different mainboards (one for Intel, one for AMD), while screenshot from CPU-Z with current memory timings and even frequency is missing.
In many cases Intel CPU which officially supported only 1333MHz had results by specification, while AMD CPU which officially supports up to 1600MHz had worse results, just because reviewer let everything on AUTO which in case of AMD meant to run on lower RAM frequency. In such cases difference between CPU spec frequency and "Default" was 20%.
Only fair comparison between two CPUs is when RAM runs on frequency described in CPU/Memory controller specs, while timings are absolutely same. If you read review and CPU-Z info is missing or RAM does not run by specifics given by the CPU manufacturer, remind to the reviewer to keep things balanced.