Searching for the Memory Holy Grail - Part 2

Does Memory Speed really matter in the Real World?

There have been skeptics in the computer industry who have boldly stated that Dual-Channel 266 memory will give performance as fast as you can get — in the real world — with a Canterwood or Springdale motherboard. Another popular variation of this is that memory over DDR400 (PC3200) will make no difference in real-world performance of an 875/865 system. To determine whether these statements held any validity, we decided to first take a look at the influence of memory speed alone on our chosen benchmarks.

We wanted to isolate performance variation as much as possible to Memory. The very design of the Intel 875/865 chipsets makes this difficult, since we have very limited settings for memory ratio. We finally decided to test stock 800FSB settings with different memory speeds, and to also test the highest standard CPU FSB setting we could run at different memory timings. To keep variables at a minimum, we looked at our benchmark results for memory that would perform at both 533 and 400 at the same timings. We settled on two 512 MB DS OCZ 4000 DIMMs. We ran the memory at stock 2.4GHz at 2.5-3-4-6-1 and at a high overclock of 1066FSB at the same 2.5-3-4-6-1 timings. Since we required 2.75V for stable operation at 1066FSB at these timings, we decided for consistency to set the vDimm to 2.75V for all tests. At each setting, we varied only the memory speed at the available 1:1, 5:4, and 3:2 ratios.

Double-Sided Memory

Since there is much confusion about when and whether Intel’s PAT is activated, and its effect on performance, each setup was checked with CPU-Z 1.18C. Under the “Memory”’ tab, with this version of CPU-Z, there is a box for “Performance Mode”, which will indicate “enabled” or “disabled”.

In all configurations, except one, with the ASUS P4C800-E with 800FSB or higher selected, both Synchronous and Asynchronous, CPU-Z indicated Performance Mode “enabled”. We will talk more about the exception later.

As you can clearly see from the tables above, gaming performance continues to improve as memory gets faster on the Intel 875 motherboard. As we move from an 800FSB CPU running 266 memory to the CPU running the same speed with DDR400 memory, we see Quake3 frame rate increase a bit over 9%, while UNBuffered Sandra increases about 40% in memory bandwidth. At 1066 constant CPU speed with memory increasing from 355 to 533, we see a larger increase of 11% in Q3 frame rate, while Sandra UNBuffered increases about 46%. The increase in Quake3 frame rate from 800/266 to 1066/533 — which admittedly includes a large boost in CPU speed — is significant, at 41%.

UT2003 also shows a similar pattern of increases over the same range, with increases over the whole range of 32%-39%. Sandra UNBuffered measurement of memory bandwidth shows a 90% increase over the same range from 800/266 to 1066/533. As we would expect, both UT2003 and Quake3 appear to respond more to a CPU speed increase than a memory increase, but alone, the increases in game benches from memory speed increases are real and significant.

Our pure number-crunching benchmark, Super PI, is shown to be more sensitive to memory speed than you might expect. We saw increases of 7% to 9% in the “memory only” increases at a fixed CPU speed. These increases are only a little smaller than those found in our Quake3 tests.

It is conclusive from these benchmarks that Memory Speed does matter in real-world performance on Intel 875/865 motherboards. Game benchmarks, Super PI, and Sandra Memory Tests all benefit from increases in memory speed. This was true up to the highest memory speed that we tested — DDR533. The differences, when we looked at just the effect of memory alone, varied from 9% to 11% at a given CPU speed, with the limited memory ratios Intel has provided us on the 875/865 chipsets. You will have to decide if the increases in performance from using faster memory are worth the cost of that speedier memory. For some, these increases will matter a great deal, while for others, they will not be worth the cost.