Page 3 of 5
Test System and Overclocking
Keep in mind that during overclocking, the latencies are as follows: CAS-TRCD-TRP-TRAS
Because we have no support from Intel and we're not made of money, I chose to go with the LGA775 520 2.8GHZ CPU for testing. The CPU offers the lowest multiplier for consumers who don't have the $1000 to buy the Extreme Edition. Even the 3.6GHZ comes with the lowest multiplier of 14x. With the testbed setup, the overclocking process begins. First, I shot for the lowest latencies possible when running at 200FSB with the DDR533 divider, making the memory run at 266MHZ. There is no point of going even lower to 200MHZ since it'll render using DDR2 practically useless. The latencies are already high, so high frequencies are needed to balance things out. At 266MHZ, the lowest timings attained were 3-3-2-6. With the lowest timings, I was only able to get another 7MHZ out of the sticks before they started to error. The results are certainly not bad, though. Next stop was to see how low the latencies would run at the specified speed. At 337MHZ, which is the rated DDR675, the latencies went as low as 4-4-3-8. That was pretty much the end point for lower latencies. An increase in 1MHZ resulted in errors during Memtest. I let loose the timings to the specified 4-4-4-10 and the memory just took off. It hit 354MHZ, which is DDR708, without a problem. This is with a 265FSB, resulting in a CPU clock of 3.71GHZ. Unfortunately going beyond 265FSB would cause major instability. I believe that the Abit AA8 motherboard put a ceiling on me. I know this for a fact because I swapped in my 3.6GHZ and manually lowered the multiplier to 14x, the same 265FSB was reached. This is certainly not bad considering that it overclocked easily at stock voltage. There was no need to increase the voltage from the rated 1.8v. Note the memory is rated to run DDR675 4-4-4-10 with 1.8v, the lowest rated voltage compared to competitors. With the memory running at 354MHZ or DDR708, everything ran smooth and fast. This just means that I need to find a better motherboard that's capable of higher FSB.
A series of memory intensive benchmarks will be run to test the performance of our sample. The following tests will be based on common benchmark programs such as 3DMark and PCMark. Real life performance differences will be tested through current games such as Far Cry, UT2003, UT2004, and Halo. Memory bandwidth benchmarks will be based off of SiSoft Sandra and Aida32. Super PI and PiFast will also be used to test memory performance, as the benchmark benefits largely from FSB increases.
SiSoft Sandra is the most widely used memory bandwidth test. You can see that at the same rated speed, the lower latencies didn't really help memory bandwidth all that much. SiSoft Sandra seems to favor the higher frequencies when it comes to DDR2 memory.
Much like Sandra, Aida32 tests the memory's ability to read and write data in large chunks. The results show that the lower latencies seem to help just a little with performance.
3DMark is probably the overclocker's favorite benchmark. Unlike 3DMark2003, which stresses mainly the video card, 3DMark2001SE tests all of the system's main components. Increasing the CPU speed, memory speed, and video card speeds will result in higher results in the final score. The lower latencies definitely helped to squeeze out more points and frames.
PCMark is another benchmark created by Futuremark. It tests the memory compression and decompression speeds in small to large chunks of data.
Although Doom 3 is known to be more graphic dependent than system dependent, the game responds positively when latencies are lowered and system bus is increased.