64-Bit vs. 128-Bit Memory Controllers - What's Right For You?
For quite some time now, I have noticed that there are those inquiring as to which route to take when jumping on the 64-Bit bandwagon. You basically have two choices: Socket 754 or Socket 939. Yes, the Socket 940 is also a 64-Bit processor, but for the sake of comparing desktop systems, I have decided to overlook Socket 940 as it is more suited towards server builds.
In my review, I have collected data from the two most popular builds in the 64-Bit market. For those who know me, I'm pretty firm in my belief that a system must pass a certain number of tests prior to claiming its level of stability. Usually, I run a minimum of 50 loops of Memtest, SuperPi 32M, and over 12 hours of Prime95 Torture Test. If my processor and/or memory configurations cause any errors in any of my tests, I cut back on my frequencies and/or latency settings and test again. I feel that it is completely pointless to be able to post amazing overclocks on suicide runs if the system is only stable enough to capture a screenshot.
However, for this test, I have omitted the Prime95 Tortue Test due to the fact that it is a very time consuming process. In my experience, if a system is able to pass countless loops of Memtest along with SuperPi 32M, a cutback of anywhere between 5MHz to 10MHz is required in order to complete over 12 hours of Prime95 Torture Test.
Keep in mind that my CPUs are highly capable of attaining above-average frequencies on simple air cooling. The memory controllers of the CPUs in this test, however, are less than satisfactory. The results are, therefore, reflected in the benchmark results that I have displayed. Due to the quirky memory controllers that my processors have been cursed with, I have conducted these tests using an "average" CPU frequency as opposed to its maximum capable overclock. My point is to illustrate the variances in memory frequencies attained. Nothing more, nothing less.
Test #1
DFI LanParty UT nF3 250Gb
AMD 3200+ Socket 754 (Clawhammer)
1x256MB G.Skill TCCD
OCZ PowerStream PSU (520W)
http://webpages.charter.net/conrad-d...SKILL1x256.png
Test #2
DFI LanParty UT nF3 250Gb
AMD 3200+ Socket 754 (Clawhammer)
2x256MB G.Skill TCCD
OCZ PowerStream PSU (520W)
http://webpages.charter.net/conrad-d...SKILL2x256.png
Test #3
DFI LanParty UT nF3 250Gb
AMD 3200+ Socket 754 (Clawhammer)
1x512MB OCZ EL PC-3200 Platinum Revision 2 TCCD
OCZ PowerStream PSU (520W)
http://webpages.charter.net/conrad-d...54OCZ1x512.png
Test #4
DFI LanParty UT nF3 250Gb
AMD 3200+ Socket 754 (Clawhammer)
2x512MB OCZ EL PC-3200 Platinum Revision 2 TCCD
OCZ PowerStream PSU (520W)
http://webpages.charter.net/conrad-d...54OCZ2x512.png
Test #4 illustrates an oversight on my part: failing to press the Refresh button in order to calculate a Sandra Memory Bandwidth score. My apologies.
As most of you already know, the more dense the memory modules, the lower the clocks. We may attribute this to the strain applied to the on-die memory controllers. On the 64-Bit platform, memory performance is directly related to the level of performance outputed by the CPU. A lackluster memory controller will definitely result in lower memory frequencies. Despite this fact, some are quick to assume that the memory modules are at fault when a "low" frequency is achieved. It would be wise to test the memory in question on another system prior to passing judgement.
Take note of the 20MHz to 23MHz drop in frequencies when running memory in dual channel.
Test #5
MSI K8N Neo2 Platinum
AMD 3000+ Socket 939 (Winchester)
1x256MB G.Skill TCCD
OCZ PowerStream PSU (520W)
http://webpages.charter.net/conrad-d...SKILL1X256.png
Test #6
MSI K8N Neo2 Platinum
AMD 3000+ Socket 939 (Winchester)
2x256MB G.Skill TCCD
OCZ PowerStream PSU (520W)
http://webpages.charter.net/conrad-d...SKILL2X256.png
Test #7
MSI K8N Neo2 Platinum
AMD 3000+ Socket 939 (Winchester)
1x512MB OCZ EL PC-3200 Platinum Revision 2 TCCD
OCZ PowerStream PSU (520W)
http://webpages.charter.net/conrad-d...39OCZ1x512.png
Test #8
MSI K8N Neo2 Platinum
AMD 3000+ Socket 939 (Winchester)
2x512MB OCZ EL PC-3200 Platinum Revision 2 TCCD
OCZ PowerStream PSU (520W)
http://webpages.charter.net/conrad-d...39OCZ2x512.png
A Socket 939 solution does not guarantee memory frequencies to be similar when running in single and dual channel. This is a misconception that many subscribe to because of the assumption that a 128-Bit memory controller is as effective as two 64-Bit controllers. Close, but not close enough. In the Socket 939 configurations tested, we can appreciate a much smaller decrease in frequencies when pushing two DIMMs. The range is anywhere between 10MHz to 15MHz. This may not justify the slight increase in price between the two builds when faced with the decision of choosing a 64-Bit system. However, refer to the data comparing the same exact memory modules on the Socket 754 system. An increase of anywhere between 3MHz to 25MHz can be realized simply by opting to purchase a Socket 939 solution, regardless of whether one chooses to run in single or dual channel.
Take into consideration that the Socket 754 is a temporary solution that AMD developed to allow a niche group to enter into the 64-Bit realm. Eventually, Socket 754 will be phased out. Socket 939, on the other hand, will continue to be nourished and built upon by AMD. We can expect to see many more processors developed on the 939 platform, though I can not say the same will be held true for its lesser-pinned sibling.
A Socket 939 solution is very close in cost to a Socket 754 setup. With all of the advantages to be reaped from the former, why settle for less?
Cheers!
:toast:
