DDR3 Memory Overclocking Charts Based On 200.000+ Results

Memory by massman @ 2010-08-22

Everyone knows that the HWBOT database is filled with benchmark results, overclocking data and so much more. One of the perks of having a database like this is not only to amuse (and annoy) overclockers globally, but mainly to analyze the results afterwards and have a better understanding of how the overclocking game really works. Processor and video card charts are scattered around the website, but memory overclocking charts is something we have not touched before.

  • next

Introduction & Methodology

Introduction

Everyone knows that the HWBOT database is filled with benchmark results, overclocking data and so much more. One of the perks of having a database like this is not only to amuse (and annoy) overclockers globally, but mainly to analyze the results afterwards and have a better understanding of how the overclocking game really works. Processor and video card charts are scattered around the website, but memory overclocking charts is something we haven’t touched before.

Madshrimps (c)


Although that last sentence was one of the reasons why I wanted to dig into the data we have on memory overclocking, it’s certainly not the most important one. I know, I know, some people see HWBOT as yet-another marketing website, but those who really understand what we do and how we look at marketing will know that the opposite is true. So, the main reason why I wanted to make memory recommendation charts is because I’m getting more and more fed up by the latest trend to use abnormally high timings and enabling B2B-Cas Delay for high frequency CPU-Z suicide shots and using those shots as some kind of proof that this memory kit is in fact the best you can get. No, people, no … DDR3-3000 CL10-15-10 B2B=31 is not the best you can get.

Methodology

The biggest issue with building memory recommendation charts based on overclocking results is that the relevance of those charts goes hand-in-hand with the purpose of the memory. In addition, it’s not just a matter of calculating the highest average frequency that a product can reach. It’s about memory performance, which includes the memory timings.

The current generation of DDR3-based memory seems to be all about three different variables: frequency, tCL and tRCD. In fact, the tRCD value seems to be giving memory kits a lot more overclocking room. So, to make our charts more relevant and useful, we have to take these three variables into account:

  • Frequency
  • tCL
  • tRCD

    The second issue to solve is to find a way to get numbers that actually make sense and can be used for further calculation. To give an example: it’s useful to rank processors by average overclocking capability, but not useful to rank them by highest frequency ever reached. The latter is a problem because then the chart is biased because we use just a single record to represent an entire group.

    So, in its essence, memory overclocking is always about two things: frequency and timings. In general, the performance of your memory kit increases if you increase the memory frequency or lower the memory timings (latencies). So, to calculate the true performance/overclocking quality score of a product we can use the following formula:
    True Score = MHz x ( 1 / timings )

    understanding that:

  • MHz, higher = better
  • timings, lower = better

    Since the group of timings consists of two variables, we have to re-write the formula to:
    True Score = MHz / (tCL + tRCD)

    This formula is almost useful.

    The third aspect of building a memory overclocking chart is deciding what kind of data you will use from your database to calculate the score. In my case, I decided to make use of the following principle. (Note: good understanding of this principle is key to understanding the charts!)

    The principle:
    “The quality of a DRAM product is determined by how people configure the memory in a safe-clock situation. The easier it is to reach frequency X, the more likely people will use it as safe-clock. The more difficult it is to reach frequency X, the less likely people will use it as safe-clock”.


    Following this principle, we can make good use of the data in our HWBOT database. We just have to alter the formula as follows:

    HWBOT Memory Index = Avg(MHz) / [ Avg(tCL) + Avg(tRCD)]

    So, the more people set a high frequency to memory kit A, the better the quality of this memory kit is. Before we can make use of the formula, however, we need to adjust something very small but very important. Although this formula would give you a decent and useful chart, we overclockers know that a different platform also means different scaling.

    For instance, back in days of the legendary BH-5, running timings very low (CL2-2-2-5) was very important. The same goes with all current systems: some platforms prefer tight timings; other platforms want only raw memory bandwidth. So, we adjust our formula to be able to generate platform-specific memory charts:

    HWBOT Memory Index = a x Avg(MHz) / [ (b x Avg(tCL)) + (c x Avg(tRCD))]

    Setting the variables as followed:

  • a = importance of frequency
  • b = importance of tCL
  • c = importance of tRCD

    Continued on the next page ->
    • next
    Comment from thorgal @ 2010/08/23
    This is a very nice effort, however, as you state yourself, Everest might not be the best tool to make your equations with.

    We found out ourselves that the real world differences between fast/slow memory is only a couple of % (1-3%), however, everest makes the differences much bigger. There's also the question whether a better everest bandwidth score automatically means a better real world performance : it often does, but not always.

    My biggest gripe is with the charts though, and is not "your fault" in any way : there are plenty of types of OCZ blades for example, some good, some bad, with different chips on them and different specs. Still : they're all thrown into one category "OCZ blade". Contrary to that : Corsair GTX is always elpida, so naturally sticks out (but should devided into several types itself). What this leads to is over generalization : "If I buy Corsair GTX I'll be ok", whereas a blade series @ eg 2ghz cas 7 is still better than a 1866 cas 8 corsair.

    All in all, forgive me PJ, I find the graphs presented this way much too general to provide any real use.

     

    reply