Beginners Guide To Overclocking: Part 2

by: FunkZ

In Part 1 of this series we introduced the idea of overclocking and went over the basics of how to overclock. We showed how to calculate CPU speed by multiplying the FSB (front side bus) times the CPU Multiplier, and how to maximize the abilities of your memory and processor by increasing these. We went over various BIOS settings important to overclocking, discussed the issue of voltages and heat, and reviewed various testing utilities such as Memtest86, Prime95 and SiSoft Sandra 2005.

In Part 2 we will discuss additional utilities available, delve into memory timings and different types of modules, and go in depth on overclocking a slightly different breed of AMD processor, the modern Athlon 64.

As always, we ask that you please apply knowledge gained from this article with caution and a bit of common sense. There is always a risk in overclocking, just like everything else in life, so if you're not sure about something, please ask. We have many experienced members who can answer questions in our forums and will be happy to assist you.

Test Utilities

Memtest86
Windows Memory Diagnostic

These utilities are designed to put your memory through its paces. If you've got a faulty module or an unstable overclock, these programs will find it. Either one can be loaded onto a floppy diskette and used to boot the computer from. They can also be a real life-saver when testing the limits of your hardware. Spare yourself the chance of corrupting a hard drive file system, figure out what works with these first.

To use, simply put the program on a floppy diskette and boot the computer. The utility will automatically load (assuming you have BIOS configured to boot from floppy) and begin running the tests. You may find that a CPU overclock that runs either Memtest or WMD successfully without error may not be completely stable in Windows. In these cases, typically a slight bump to vCore (increase in CPU voltage) will usually resolve the problem.

CPU-Z
WCPUID

CPU-Z is probably the most popular program to verify and display your system overclock. With the latest version there's even a way to submit your overclock online for verification and to get a comparison link, similar to many graphics benchmarking programs. WCPUID is a similar program, however it has not been updated in some time, and may not recognize all the latest processors and chipsets.

Prime95
eCalc
Super Pi
OCCT

Here are a few Windows-based programs that can help you verify you've got a stable overclock before you actually start using your computer for other tasks. In Part 1 it was mentioned that Folding@Home can be used to test stability, however a failure often results in losing the work unit, which is why I don't like to use F@H for this purpose.

Memory Timings

Memory timings or latency refers to how quickly the system can get data in and out of the RAM. This is different from Memory speed, or the frequency that the memory runs at in relation to the processor and system bus. Think of it in terms of a mass-transit system. The memory speed is the rate at which the Metro train moves from station to station. The latency measures how quickly the people can move on and off the train at each stop. Generally, the lower the memory timing value, the less latency there is, and the faster the memory responds.

Most BIOS are configured by default to Auto detect timings from the memory module by SPD or Serial Presence Detect, however many have the option to change this to manual so that the user can adjust the settings individually. SPD values are programmed into the memory by the manufacturer, and are typically printed on a label on the side of the module. Timings are usually referred to in this order, along with some available settings in the BIOS.



CAS is sometimes referred to as CL or Cycle Length. Some motherboards have an option as low as 1.5 for this setting. But the effect of CAS on memory latency is much less than tRCD, tRP or CMD. CMD or Command Rate has the most effect on memory performance. Not all memory and/or motherboards are capable of running a 1T CMD however.

Memory manufacturers and overclockers usually refer to memory timings in the same order as listed above. For example, some low-latency memory might indicate CL2 2-2-5 right on a sticker on the module itself. Some memory (such as TCCD) may be rated differently at different speeds such as low timings of 2-2-2-5 at PC3200 (200 MHz DDR400) and higher timings of 3-4-4-8 at PC4400 (275 MHz DDR550). Many memory modules do not advertise CMD so you should check reviews before purchasing to get an idea if it will run at 1T.

There are many manufacturers of individual memory chips (such as Samsung, Winbond, Hynix) and also manufacturers of memory modules (such as Corsair, Kingston, OCZ) who use other companies' chips to make their modules. Memory chips are tested and "binned" by the manufacturer following production and then sold to other companies to make the modules. Some chip manufacturers (such as Samsung, Geil) also make their own modules. Memory chips come in many different flavors so there are a few things to watch for.

One buzz word I'm sure you've heard about in the last 2 years is BH5, or more specifically, Winbond BH-5 chips. These have become almost legendary in the overclocking enthusiast world for their ability to run at low latency timings, even at high speeds, albeit when supplied with an extreme amount of voltage.



More recently, companies have taken to using BH5-based UTT chips to satisfy overclockers' needs. Some people have had good luck with modules made using these chips, however be aware that the UTT designation means that the chips came untested from the manufacturer. When memory manufacturers have a wafer come off the line that for whatever reason doesn't meet specification, rather than scrap the entire piece they often (depending on market demand) sell off the chips as UTT and it's up to the module manufacturer then to test the chips and determine if they're any good. Since these come out of at least a partially defective wafer, it can't be said with any certainty that the chips can take all the extra voltage and speeds people throw at them.

In any case, both UTT and BH5 based modules are typically only good up to ~225 MHz at the voltages available on most motherboards, i.e.. 2.85 to 2.9 volts. Many DFI motherboards are capable of supplying more than 3 volts to the memory, up to and even including 4 volts! If you don't have a DFI board, you can check out OCZ's DDR Booster to see if it's compatible with your motherboard. For many boards the Booster will give you from 3.4 to 3.8 volts available.

The Samsung TCCD is another type of chip that has caught on lately, and may just surpass the BH-5 for "King of the Memory Hill" because it can run at tight timings at default speeds, loose timings at much higher frequencies, and doesn't require much more than stock voltage to keep it running.

Most system memory made today is of the TSOP variety, or Thin Small Outline Packages, rather than BGA (more commonly found on video cards) or Ball Grid Array. The names have to do with the way the chips are made and how they attach to the circuit board of the memory module.

Athlon 64 Overclocking

Although Part 1 of this guide was not processor-specific, the procedures detailed apply more to Socket A overclocking than the latest A64 chips. There are some significant differences which are worth mentioning to help you get the most out of your Socket 754 or 939 processor.

First off, the A64 does not really have a FSB or front side bus speed per say. The term FSB refers to the frequency of the connection between the CPU and Memory Controller. On an Athlon XP chip this could be 133, 166 or 200 (effective 266, 333 or 400 DDR) depending on the model. But the Memory Controller is integrated into the processor on an A64 chip and therefore runs at the same speed as the CPU. There is a connection to the Northbridge on the motherboard however, called the Hypertransport Link, which can be either 800 MHz (effective 1600) on Socket 754 or 1000 MHz (effective 2000) on Socket 939.



Now the Hypertransport Link speed is determined from the base HTT speed of 200 (referred to as CPU Frequency in this BIOS above) times the HT Multiplier (shown as HT Frequency below) which is by default, 4x on S754 and 5x on S939. It is very important to remember to lower the HT Multi as you increase HTT. Ideally you want to try to keep the overall link speed close to the default 800 or 1000 as going much above these will result in instability. More than a few times I have seen someone complain they can't get more than 220-230 HTT on their overclock and think they've topped out the memory or CPU. Had they reduced the HT Multiplier by one step more they likely would have found they could keep going higher on the HTT.



The principle behind the CPU Multiplier is the same for A64, only now they refer to it as the FID, or Frequency ID. If you take the base HTT frequency and multiply it by the FID you end up with the speed that the CPU runs at. Unfortunately with A64 processors, only the default multi and lower is unlocked and available to use. Some BIOS will allow half-steps on the FID, however these have been shown to either cause instability or not even work at all, so it's best to just stick with the full multi's. FX chips have all multipliers unlocked, so these can be adjusted both higher or lower than the factory default.



Unlike AXP systems, with A64 it is not as important to make sure the FSB remains synchronous with the memory speed. While benchmarks will show a slight increase staying with a 1:1 ratio, going asynchronous is not the detriment to performance it once was. Considering the high speeds available to modern S754 and S939 processors and motherboards, it's a good thing that memory dividers can be implemented.

Speaking of memory dividers, this is another setting that sometimes confuses people. While the idea of memory ratios or dividers have existed for a while, AMD users were always told not to use them. I think it was drilled into our heads. Now that we can use them we need to understand that the exact ratio changes slightly depending on the CPU multiplier you use. The reason for this is with the memory controller built into the processor, any divider used takes into account the CPU Multiplier when calculating the ratio. Below is a chart that shows what the different settings for memory divider in BIOS will result in.



The numbers in the top row correspond to the memory speed setting in the BIOS. Some motherboards will only have standard JDEC speeds available such as 200, 166, 133 and 100 whereas others may have the listed "in-between" speeds. The number in parenthesis beside the memory speed indicates the hypothetical ratio for that particular setting. For example, to run memory at 166 we start by taking the base frequency of 200 and multiply that by the ratio of 5/6 and we get 166.66 exactly. However, as mentioned above, the ratio has to be a factor of the CPU Multiplier, so we need to look at the row indicated by the multiplier being used. For example, a 3000+ "Venice" stock multi is 9x, so if we come down to that row, then move across the row to the 166 memory column we find that the ratio used for this setting will actually be 9/11 rather than the 5/6 as indicated at the top. The 9/11 ratio yields a memory speed of 163.63 which is close, but not quite the same as what it should be for a true 166 speed. This is not a problem but just something to be aware of.

Now let's put our knowledge to the test and go through some sample overclocks...

Athlon 64 Overclocking continued

We're going to overclock two hypothetical systems using the latest "Venice" core, a 3000+ with some PC3200 (DDR400) memory, and a 3200+ with some PC4000 (DDR500) memory.



Take note that as we raise the HTT frequency we lower the HT Multiplier to keep the HTT Link speed as close to default as possible. Sometimes this may fall slightly more or less than default, within 5 or 10% should be okay. Less than that and you may actually hurt performance, more than that and you can experience instability.

Also note that as we raise the HTT frequency we must also lower the memory setting, first to 166 and then to 133 in the PC3200 example. This is done to keep the memory running close to its rated speed. Obviously if you have memory that will perform better than its rated speed you could take advantage of this by using a higher memory setting or HTT frequency. It may take a piece of scratch paper and a calculator, but you can easily figure out what speeds you should be running at with any particular BIOS setting by doing the math.

Since the HTT base frequency is not a FSB speed, there is no benefit to trying to maximize this value. Concentrate instead on maximizing the CPU and Memory speeds, and use the HTT and Memory dividers to achieve this.

Now we'll put our knowledge to the test and try out our overclocking on a real system. We're going to be using a 3000+ Venice in an nForce4 motherboard with some PC4400 memory.



In the first screen shot of CPUz we've taken the memory as far as it can go. The voltage to the DIMMs was increased to the board maximum of 2.9 and the HTT speed was increased from 250 to 290 in 5 MHz increments. At each increase the system was booted from a WMD diskette and allowed to complete at least one test pass to verify stability. As we went higher in frequency we also had to increase the memory timings to keep things stable. In this way we've determined that the memory can run at a maximum of 290 MHz at CAS 3-5-5-10.



In the second screen shot we've applied the 166 Memory setting in BIOS, which yields a 9/11 ratio since we are still on the default 9x CPU Multiplier. Then we continued to raise the HTT from 290 all the way up to 320, again in 5 MHz increments, also using the WMD boot diskette to test stability. If the system hard-locked during a test pass, the voltage to the CPU was increased by one step and the test attempted again. By using the board maximum CPU voltage setting of 1.7 we can reach nearly 2900 MHz stable. With the memory divider however we're running its speed under spec, therefore we can get away with lowering the timings to CAS 2.5-4-3-7.

At this point if we had a 10x CPU Multiplier available to use we would be finished. With a 290 HTT and a 10x Multi the CPU would be at 2.9 GHz and the memory could run 1:1 at 290 MHz. But with the 3000+ the default multiplier (and thus highest available) is 9x. So now it is time to break out the calculator and determine what the best possible speed to run at is.



By reducing the CPU Multiplier to 8x and leaving the memory divider at 166 we can increase the HTT to 360 which places both the CPU and memory at or near their maximum stable operating frequencies. Notice that the memory ratio has changed to 8/10 with the switch to the 8x CPU Multiplier.

Conclusion

I hope this article has been helpful to you and given you more understanding into the intricacies of overclocking. If you have questions not answered here, please feel free to post in our Forums and our members should be able to help you.