A Newbie's Guide to Overclocking Memory

Why Write A Newbie Guide
A few months ago I had no idea how to begin the process of overclocking my machine. It was painful even trying to figure out what questions to ask. I still consider myself a newbie.

Therefore, this guide is far from complete and will contain errors, but hopefully it helps those just beginning. If it proves to be of any value maybe we can continue to correct, revise and improve it in the future. If not then it'll scroll off the bottom of this forum and fall into a bit-bucket somewhere someday.




A Newbie's Guide to Overclocking Memory
Last Updated Sept. 14, 2002

To understand overclocking, you need to realize that there is a single clock which controls all operations made by the computer. The various operations (for example graphics) are slaved to the clock. The clock measures time in what is typically called a tick, or cycle, or the clock refresh rate and reflects a discrete measure of time. The cycle is also commonly converted to an operational frequency, (1.0/tick), in many discussions. For example, a clock cycle of 7.5 nanoseconds converts to a 133 Mhz clock speed. (The clock cycle is a sinusoidal wave for those familiar with such representations. The cycle is thus made of a positive half and a negative half.)

A Simple Perspective of Overclocking, a Balance of Various Component Demands
Because the various computer operations are slaved to the single clock, a balance of the various device speeds must be achieved so that each device is stable.

The clock controls the Front-Side Bus (FSB) and the different components use that bus speed to determine their own speed. With a single speed available from the FSB, this creates the need for "multipliers" a term you often hear. Some multipliers speed the FSB reference speed up for use by their component, for example the CPU multiplier is typically greater than one and for an AMD XP 2100+ is set at 13 when the CPU is locked. Others slow it down; the PCI bus for example typically uses a multiplier of ¼ on many of today's motherboards. If the multiplier slows the speed down, it may also be called a divider. For this discussion multipliers and dividers are viewed as performing the same operation, i.e. changing the FSB clock speed for use by their particular device.

One definition of overclocking then is the process of balancing the various component speeds within the machine to produce stable operation at different FSB speeds.

A Simple Overview of Memory
There are far too many characteristics associated with memory and its operation to include in this simple guide. Thus, I will try (with the help of other contributors) to provide a brief perspective that will hopefully help you better understand memory itself. This perspective will look at predominate Dynamic Random Access Memory (DRAM) types in use today and a few of their characteristics that directly influence our initial efforts in overclocking.

. Predominant RAM Types
A reference discussing much more completely the various memory types can be found at http://www.pctechguide.com/03memory.htm. Here I will provide some detail related to the two main types of memory in use today, SDRAM, DDR SDRAM and RDRAM.

SDRAM – Synchronous DRAM (SDRAM) employs the fact that most PC memory accesses are sequential, i.e. the information need has been stored sequentially. Thus, all the bits can be retrieved in a burst. The SDRAM uses a memory controller to determine the location and size of the block of memory required and then supplies the bits as fast as the CPU can take them, using a clock to synchronize the timing of the memory chip to the CPU's system clock. This feature of SDRAM allows it to deliver data at burst rates of up to 100Mhz while using a FSB of 66Mhz.

PC133 SDRAM is an advancement of SDRAM allowing a doubling of the burst data rate and compatible with a 133Mhz FSB.

(This information related to SDRAM was taken from http://www.pctechguide.com/03memory.htm#SDRAM.)

DDR SDRAM - Often just called DDR (Double Date Rate). DDR works on both "sides" of the clock cycle (positive portion and negative portion), so its rate (frequently noted as bandwidth) is doubled, hence the name.

This type memory is the most common, cost effective, and arguably best RAM option currently available. Also, it is the most overclockable RAM, with some high-end sticks allowing FSB speeds in the 200s. It is rated by both speed and bandwidth. For example,
PC 1600 = DDR200 = 1.6 GB/s
PC 2100 = DDR266 = 2.1 GB/s
PC 2700 = DDR333 = 2.7 GB/s
PC 3200 = DDR400 = 3.2 GB/s

Note however that to achieve these ratings, the memory needs to be driven at its rated speed. If you put some PC 3200 in an Athlon stock mobo with a 266 Mhz FSB; it would just run at PC 2100. Doesn't hurt it but you have wasted your money because you are not running it at its rated speed. You would need to overclock the FSB to 200 Mhz to fully use the PC 3200 at its rated capacity.

RDRAM - A slightly more expensive, yet top performing memory, exclusive to the Intel P4. RDRAM uses a clock generator of its own to run at fast speeds, which achieves the high rate (bandwidth). However, it can get quite hot, and is not a good OR easy overclocker. If the RDRAM multiplier is lowered, decent FSB speeds might be reached. For example, stock PC800 is 100 * 4, but thats equivalent to 133 * 3. Speed is given in Mhz * 2, since it is on a dual-channel architecture.

PC800 = 4*100 = 3*133 = 3.2 GB/s
PC1066 = 4*133 = 4.2 GB/s

Note that conventional RDRAM boards don't give options beyond low 150's for FSB.

. Some RAM Characteristics
Today's memory may operate as either synchronous or asynchronous. Synchronous memory is synchronous with the FSB. DDR generally operates asynchronously but many mobos (motherboards) can switch memory operation to a synchronous mode either manually in the bios or automatically based on a particular FSB speed being reached. Asynchronous memory responds to input signals whenever they occur as long as the signals are applied in the proper sequence, with signal durations and delays between signals that meet the specified limits.

The multipliers used by memory also frequently change at a particular FSB speed. For an Epox (the board I currently have) the FSB speed at which the multiplier behavior changes is 166 Mhz. This change occurs as follows:
- Below 166 Mhz, the multiplier is 1.25. Another factor of 2 results from DDR. Thus for a FSB of 133 Mhz, the memory speed will be 1.25*2*133Mhz which produces the 333 Mhz memory speed you commonly hear today. And the memory is operating in an asynchronous mode.
- At and above a FSB of 166 Mhz, the multiplier is 1.0. Thus for a FSB of 166 Mhz, the memory speed would be 1.0*2*133Mhz which also produces the 333 Mhz memory speed. In this case, the memory is operating in a synchronous mode.

You should now have noted that a FSB of 133 Mhz and one of 166 Mhz both produce the same memory speed of 333 Mhz for DDR memory as a result of the multiplier change in value at 166 Mhz. The difference between the two cases being the first is in an asynchronous mode of operation while the second is in a synchronous mode.

I summarize this behavior in the following table:
~

If others will provide information related to other boards, I will add it to the table. Corrections too please.


The Mechanics of Overclocking the Motherboard FSB and Memory
1) Go into your bios and set the most aggressive memory timings available to you. I'm running 2-2-5-2 on my Corsair 3000 cas2. (See http://www.lostcircuits.com/memory/esdram/2.shtml or ~ http://www.itcp.net/~awsh/Speed_Dem...Guide/Index.htm or http://www.radiativenz.com/guides/m...ide/index.shtml (my favorite) should you need help with the meaning of such things as 2-2-5-2.) Start at a stock FSB speed (133 should be stock on new mobos) and also make sure Vcore and Vdimm are set to default voltages.

2) While in the bios also set the multiplier on your CPU down (when the CPU is locked the multiplier is fixed and cannot be changed), maybe 8. Yes, this is low but you want the CPU to be stable as you increase the FSB so you make sure it is running slow. (By the way, CPU speed is FSB*multiplier. You probably know that, but just making sure.) You'll reset the CPU multiplier higher later.

3) Run an application that really exercises the memory at each FSB speed to see if you're stable (see Testing below). Sandra doesn't cut it for this test. After you know this initial setup is stable, start increasing the FSB speed and retest each time. Yes, this takes a lot of time!

You may wish to jump to 166 FSB once you've tested stock. (I believe that that is the point at which one or more multipliers change and you'll want them at the higher levels of the FSB anyway.)

4) When the test results are unstable, increase the Vdimm and try again. Continue increasing the Vdimm till stable or you don't want to go any higher in voltage.

5) Using this process of increasing the FSB speed, testing and increasing the Vdimm as needed, work your way up to the highest FSB you can reach for the max Vdimm voltage you're willing to run.

Be careful to keep the CPU near stock speed, or lower, by adjusting the CPU multiplier down should you FSB speeds reach high levels. This will keep the CPU stable without having to increase its voltage {Vcore), an operation that we wish to delay and keep independent from the memory overclock.

There are various strategies about how to minimize the time you spend reaching a maximum FSB but I'll not take up space to suggest those --- I'm sure you can figure several others out.

6) You can reduce your aggressive memory setting (2-2-5-2) to get higher FSB speeds but it has been noted several times by others that "In general, you can get better memory performance out of a lower clock speed and better timings than out of a faster clock speed but worse timings."

Often, your memory performance is more important to overall performance than your overclock, so a good approach is to overclock as far as your memory can operate at its most aggressive settings.

7) You can have problems (unstable behavior) with video, networking and other cards, as the FSB gets high. The FSB speed at which the multipliers change is critical to these problems. Surf those forums/threads for help.

8) You would work next on the CPU speed but that isn't part of this guide, at least not for now. Basically, you start increasing the CPU multiplier and increase the Vcore as needed for stability in this case. And test, test, test, …. always watching your CPU temperatures!!! The CPU you save maybe your own.

Testing
A number of the steps above involve testing of your memory overclock. So we should try to better understand this part of the process a little better.

You can find a very large number of tests at http://www.benchmarkhq.ru/english.html?/b_e.html available to download but which ones should you run and what do they produce? In using any test it is extremely important to understand what the test is measuring and how to interpret the results. Does it test a component alone or does it involve the overall system?

For memory, a good stress test is Memtest86 (http://www.memtest86.com/ ). It runs an array of tests of memory reads, writes, copies, and moves in different patterns and block sizes, and it runs in the first small segment of memory and tests the whole range independently of the OS and HDD, PCI devices, etc.

Prime95 ( http://www.mersenne.org/freesoft.htm) is an excellent application to do ongoing testing of your system stability. My experience thus far says if Prime95 will run stable your system is stable.

Results – An Initial Overclock
So what do you accomplish with the overclock of the memory? Here are 2 results from my machine after I unlocked my CPU. Results were obtained with a Vdimm=2.9volts.
~

Three results are shown from Sandra SiSoftware: CPU Arithmetric, CPU MultiMedia, and Memory Bandwidth. Higher is better. Within these results, which is better? The results indicate the CPU performs better at a FSB of 143Mhz using a multiplier of 13. It should because the CPU speed is 143Mhz*13=1859Mhz compared to 168Mhz*9=1512Mhz. Yet, the memory did perform better at 168Mhz. Does this matter?

As noted earlier, Prime95 does stress (work) the complete computing part of your machine. Most importantly it did show my system to be stable at the FSB speed of 168. In addition, a look at the results from running Prime95 show the time to run a cycle of Prime95 has decreased by about 10% for the higher FSB. Thus while the CPU performance has decreased at the higher FSB speed due to the multiplier I used, the overall performance of the system has improved!

What might a newbie learn from this example? First, understand what your test is telling you. Sandra is reporting exactly what it said it would, the performance of the CPU. It didn't claim that was a true measure of my machine's ability to compute even though I may have assume it to be the case. Prime95 did run stable at the higher FSB and provided me with an indication of the better performance of my machine even with the CPU multiplier set at 9. Memory performance is very important to the overall performance of your overclock.

Will your machine behave in this manner? Maybe, may not. Pentiums will behave differently than my AMD. Various motherboards may also change the results. However, hopefully with these tools you can begin to experiment.



Caution
. Monitor the CPU temperatures. A program like MBM 5 (http://mbm.livewiredev.com/ ) will give you readouts on your temperatures, voltages and CPU speed.
. Test, test, test.



Contributors
macklin01 –> testing
---X--- –> memory types

Credit for this info goes to deeppow @ Overclockers.com Forums