Page 3 of 5
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 3DMark2001. Real life performance differences will be tested through current games such as Far Cry and Half-Life 2. Memory bandwidth benchmarks will be based off of SiSoft Sandra and Aida32. Super PI will also be used to test memory performance, as the benchmark benefits largely from FSB increases.
I included all of the timings and speeds that I achieved through overclocking, but I will also add 2 new tests to show the differences of running an Athlon64 with or without a divider with low latency timings. I will also add the change in memory bandwidth with a change in CPU speed. Hopefully, these new tests will shed some light on how complicated memory benchmarks can be on an Athlon64.
I'll be including the different settings for each test as one color on the bar graphs side by side. The first test should show the differences in performance while running a divider on an Athlon64. Sometimes your Athlon64 may have a bad memory controller on it and make a high 1:1 overclock difficult. Or perhaps your brand new motherboard cannot run high HTT speeds but can run high CPU clockspeeds by increasing the multiplier (if you are even lucky enough to own an FX processor). This test will illustrate how low latency timings might be able to make up for the lack of high memory speed overclocks. The second test will illustrate how the Athlon64's memory bandwidth greatly increases with CPU speed, despite memory speed being the same. Although this may seem rudimentary to most experienced Athlon64 overclockers, this should help Pentium 4 and AthlonXP users realize why their memory bandwidith is diminutive in comparison.
*The "D" after the "215HTT(2-2-2-5)D" in the legend denotes "Divider", which means the 166Mhz memory divider was used in this particular test. Despite HTT being at 263Mhz, memory was only at 215Mhz at the latencies of 2-2-2-5.
3dMark2001 is an excellent benchmark for determining overall system speed. It uses older DirectX technology, which allows the benchmark not be GPU dependent. A small increase in memory or CPU performance will result in a significant iincrease in scoring. The 4 columns display an expected linear scoring result, but notice that 285Mhz (2.5-4-4-8) is still not enough to beat the lower latencies of 275Mhz (2.5-3-3-7). Also notice the small increase in speed from running high memory frequencies over low latencies, as indicated by the orange columns.
SiSoft Sandra 2005 Memory Bandwidth
The Sandra Memory Bandwidth results are nearly identical. For every increase in 1:1HTT with memory speed, there is an increase in overall memory bandwidth. The increase comes regardless of latency, although at the same memory speed lower latencies should bring higher bandwidth. The first four columns reveal the increases in memory MHZ scale as well in memory bandwidth. The orange columns show the loss of performance when not running 1:1, despite the lower latencies. However, notice while there is a loss in bandwidth in Sandra Memory, there isn't much loss in 3dMarks. This illustrates how an Athlon64 is much more CPU speed dependent than memory. Also notice how there is an increase in memory bandwidth with increased CPU speed despite memory speed remaining the same (indicated by the yellow columns). This again illustrates how CPU dependent the Athlon64 is on memory because of its integrated memory controller. A Pentium4 would show the almost the same results with the same speed in memory regardless of a change in CPU speed.
Everest Home Edition (Memory Benchmarks)
Everest Home Edition's Memory Benchmarks show the same results as Sandra Memory Bandwidth, although the performance gaps are much smaller.
Everest Home Edition also includes a latency benchmark. The Athlon64's performance with various latencies is also dependent upon its increase in CPU speed, despite the earlier columns having much lower latency. Notice how at 285 MHZ at 2.5-4-4-8 is 1ns higher than 275Mhz at 2.5-3-3-7. This indicates that at similar CPU speeds, latency does help increase system performance. Athlon64 speed can greatly effect latencies as well. The best latency configuration would be a highly clocked A64 with a divider on the memory (so that the latencies can remain at 2-2-2-5).