Limiting Factors to Obtaining High Memory Frequencies

While scaling memory frequency on a Pentium 4 platform, the main limitation to FSB and consequently memory frequency scaling, assuming the CPU isn't limiting factor, is the quality of the motherboard, which can vary inside of a particular motherboard model lines even. For example, two Abit IC7 motherboards may reach markedly different levels of FSB and memory frequency overclocking given identical components with which to work. This then points to quality of the motherboard and electrical components, with a special emphasis on the Northbridge chipset and MCH (memory controller hub) contained therein.

In my experiences, Athlon 64 Socket 754 systems just don't scale very well with 2x512mb of memory. This can be seen as a limitation of the CPU and platform as running 1GB of ram off one 64bit memory channel is quite intensive. While there are some that have achieved high frequencies with 1GB of ram on Socket 754 platforms, I would say that the majority of people have difficulty trying to reach high memory MHz with this amount of memory at acceptable timings and 1T command rate.


Socket 754 CPU

With Athlon 64 Socket 939 we are dealing with several other issues competing and conspiring to limit our fun. To elaborate, there are 4 main factors that I have found to play key roles in memory frequency scaling on Socket 939 platforms, the CPU core quality/memory controller, Motherboard/Chipset, bios and voltage, and memory modules themselves. For the purposes of this article we will be discussing and then reviewing memory frequency scaling or overclockability on Socket 939 platforms specifically.


Socket 939 CPU

As stated, the Athlon 64 Socket 939 CPU enjoys a dual channel memory controller that is on die to the CPU. With the placement of the memory controller onto the CPU itself, AMD had brought about significant performance advantages with the Athlon 64 in comparison to previous CPUs. However, there is also the inherent ability for each CPU to be somewhat different in quality and overclockability, this quality logically extending to the on die memory controller as well. When overclocking, there is an element of luck that must be present in order to get a good core that will allow high memory MHz operation, just as there is some luck present in getting a good overclocking CPU.

Whereas the Socket 754 CPU uses one 64-bit memory controller in single channel, the Socket 939 CPU uses 2x64bit controllers for dual channel operation. This is both a benefit in added memory bandwidth, but also eases the "load" on the controller, and in effect allows somewhat easier memory frequency scaling to a point with 2x512mb or 1GB of system memory. There will come a point however where the CPU and specifically that memory controller in question will become a limiting factor. My experiences with about 12 socket 939 CPUs has been mixed with a few of the CPUs maxing at or around 250MHz maximum memory frequency at acceptable timings and 1T command. The majority of CPUs I have worked with top out in the 260-280MHz range as far as usable memory frequency is concerned, again at acceptable timings and 1T command rate. While potentially, we could loosen timings and set command rate to 2T, the performance losses almost always overshadow the often small increases in achievable memory frequency, thereby making that option a losing proposition.

Motherboard and chipset implementation can also play a factor in how high we can scale our memory on Socket 939 platforms. Via chipsets while decent performers tend to overclock less than their nForce3/4 counterparts. Hopefully this will change in the future, but for now I steer people towards nForce platforms for socket 939. Assuming you have made a chipset decision, different manufacturers approach the enthusiast community differently, and tune their BIOS’s accordingly. MSI has made big waves with their K8N Neo2 939 Motherboard and generally the ability to scale the memory inline with the HTT is as good or better than any board currently out now. That said, I have heard future NF4 boards will be even better, but without working samples of NF4 Ultra boards I cannot confirm. What I can say that I have used about 20 different Socket 939 motherboards, some of which were within the same model line, and each behaved differently. The potential exists for our board to be a limiting factor to our overclocking or scaling of memory frequencies.


MSI K8N Neo2 – popular nForce3 overclocking motherboard

In regards to motherboard bios and voltage selections, I have never known memory in all of my past experiences to be so sensitive to memory voltage. I can say that each set I have tested for this review behaved somewhat differently than the next set. Even similar sets from the same manufacturer such as G.Skill required separate testing to delineate what voltage setting provided the best results etc.. No longer can we be confident that poking our noses into the BIOS and setting our VDIMM to max available will be most beneficial. This can actually have a detrimental effect as the maximum voltage tolerated out of any set in my possession was 2.8v, with most sets preferring between 2.65-2.75v. Sets that liked 2.65v would actually "lose" overclocking potential with even a slight voltage increase to 2.7v. This is where accommodation by manufacturers and their bios options come into play. MSI has seen fit to include increases in VDIMM in 0.05v steps, whereas other makers have us make due with 0.1v jumps. This could ultimately prohibit us from finding our max potential frequency, as either too much or too little voltage might be used and not allow us to hit a "sweet spot". I would be nice to have even finer voltage options in BIOS, possibly something we could petition motherboard makers for would be 0.025v or even 0.0125v steps.

Finally, when discussing potential limiting factors to how far we can push each set of memory would of course be the quality and design of the memory modules themselves. While one might think that since each set uses the now famous Samsung TCCD memory chips, the ram modules themselves would perform and overclock in a similar manner. This is far from the truth, and while certainly this is true to some extent, the PCB and culling or speed binning that manufacturers perform have an impact as well. As has been discussed in Legit Review’s coverage of “Memory PCB Selection Matters” by Nathan Kirsch, the selection of which PCB the ram is attached to can have a direct and very important impact on how the ram behaves during high overclocks. That said, even sets within same line, using same week of production modules and same PCB will exhibit different tendencies when pushed to the limits. We go back to luck of the draw when purchasing PC3200 memory with TCCD chips, or rely on the memory manufacturers to remove a certain element of risk as they design, qualify, and speed bin certain sets for higher operating frequencies.

Whatever your cup of tea, there is sure to be your flavor among the many samples we are reviewing here in the lab. So in our manic frenzy we bring you the main course---Prawn memory ala uberclock. Have a gander will you........