A64 Overclocking Theory 101

This is the second article in a six-part guide book.

Overclocking, although not terribly difficult to do, is sometimes difficult to understand. Far too many people overclock their systems without truly appreciating what they are doing, and how it impacts their systems. This section of the guide will take you through some of what you already know and hopefully some of what you don’t.

Power Supply and Voltage/Amperage

One aspect of a PC that is often overlooked is the power supply. The power supply is usually thought of in terms of ‘Watts’. Power supplies are much more than just their total wattage rating and there are many specifications of interest, especially to an overclocker.

Short-Media’s very own MediaMan has put together a fantastic PSU guide that any prospective overclocker should read. It is very informative, and a great read.

Without going into too much theory and staying relevant to overclocking, the following is important to understand:

• As the clock speed of your processor/memory increases, so does it’s power consumption, which is a strain on the PSU. This power consumption increases even more when you increase your CPU voltage or memory voltage.
• The +12, +5 and +3.3V rails of your PSU need to stay very steady and should exhibit very little variation from one moment to the next. Low quality or incorrectly functioning PSUs often have constantly varying voltages, which can negatively impact system stability.

Since the Athlon 64 does not operate at any of the PSUs voltages, it requires voltage regulation circuitry on the mainboard to achieve its rated 1.35 to 1.5 volts. In the past, CPUs have their core voltage derived from the 3.3V rail. Instead, many modern mainboards are now using the +12V rail to power the CPU. To make matters worse, new graphics cards (also very high power draws) are also utilizing the +12V rail. This has increased the demand for much higher output +12V rails. In the past, +12V rails were used primarily for optical and hard drive motors and 10-15A was considered sufficient for most systems.

I would recommend a power supply with a minimum of 20A on the +12V rail for today’s modern A64 systems. It is not unusual to see inexpensive ‘550 watt’ power supplies with only 15A on the +12V rail. It is VERY important to check the rail amperages and not just the total rated power of the unit. If you have a higher end graphics card, like a 6800/7800 series card or an X850/X1800 card, the greater the amperage on the +12V rail the better. When using SLI cards, you’ll want closer to 30A on the +12V rail, if not more.

Although the PSU is one of the most important ‘power’ related components, when it comes to overclocking, the mainboard is also very important. As mentioned earlier the mainboard has ‘power regulation circuitry’ to produce the correct core and memory voltages your system requires. In essence, you can think of your mainboard as a second power supply in your system. Some lower-end or malfunctioning mainboards exhibit fluctuating vcore and vDIMM. This problem can worsen when using a PSU with similar problems. When trying to find your maximum overclock, you can be severely short-changed if your vcore keeps dipping or fluctuating. For example, if your best CPU overclock is stable at 1.55V of vcore and it dips to 1.5V, your system stability could be instantly compromised. Most ‘enthusiast’ mainboards use 3-phase regulation circuitry to produce stable and accurate vcore and vDIMM.

There are many high quality PSUs on the market today. Some brands to consider include OCZ, Super Flower, Antec, Enermax and of course, PC Power and Cooling. Most higher quality mainboards from reputable manufacturers should have steady voltage regulation.

Heat

Heat is one of the biggest problems that overclockers face. As clock speed and voltage increases overclockers must find ways to dissipate the increased heat output.

Without getting too technical, there are some general rules relevant to overclocking:

• As CPU temperature increases, the maximum attainable overclock decreases.
• As CPU temperature increases, the possibility of instability also increases.

It is clear that high temperatures are an enemy of the overclocker.

It is also important to recognize that every CPU has its maximum rated ‘die’ or ‘casing’ temperature from AMD. Older processors without a heat spreader were rated for a maximum ‘die’ or core temperature. A64’s are rated for a maximum ‘case’ temperature which is not the temperature within your PC, but rather the temperature underneath the protective metal heat spreader.

AMD has done a fantastic job in producing lower-power consuming A64 CPUs. This is especially true with their latest batches of 90nm ‘D’ and ‘E’ revision based cores. With most models operating at only 60-80W and with a vcore of only 1.35-1.4V, they use very little power and run very cool.

Below is a table of maximum A64 case temperatures (sometimes simply referred to as ‘tcase’). If you are unsure of your core type you can use CPU-Z to find out. The new ‘E’ revision processors (San Diego, Venice, Manchester and Toledo) seem to have varying maximum ‘tcase’ values. They are usually in the range of 57-67 degrees. You can use the ‘MaxTcase’ utility in the ‘Tools’ section for more information. Anything exceeding this threshold will not immediately damage the CPU, but prolonged exposure to these high temperatures may cause permanent damage and render your beloved chip a key-chain ornament .

Athlon 64 MaxTcase Values by Core

AMD Product Name	Core Type	Default Voltage	Maximum Temp.
Athlon 64/FX	Clawhammer, Newcastle	1.5V	70’C
Athlon 64	Winchester/Palermo	1.4V	65’C
Athlon 64 X2	Manchester/Toledo	1.35/1.4V	Varies (usually 57-67’C)
Athlon 64/Opteron	San Diego/Venice	1.35/1.4V	Varies (usually 57-67’C)

At default clock speeds your A64 should be able to remain stable at temperatures close to its maximum tcase value. When the temperature exceeds that threshold, instability may occur. When running your A64 with a higher clock speed, the stability limit at the maximum tcase may no longer apply and the chip may begin exhibiting instability at much lower temperatures. Every chip is different, but, generally speaking, once you start pushing the clock speed higher, lower temperatures will be required to maintain stability.

Cooling

Keeping your CPU running cool is important to maintain stability, especially when running your CPU outside of AMD’s specifications.

So just how cool should my A64 be running if I want to overclock?

The simple answer to that question is ‘the cooler the better’. There is a generally accepted law that for every 10C decrease in temperature; a microprocessor should be able to increase its clock speed by about 3%. This has driven hardcore enthusiasts to cool their CPUs with just about everything from air to liquid nitrogen.

Note: Some 90nm A64 processors have what is referred to as the ‘cold bug’. With these ‘bugged’ processors, very low temperatures below zero degrees can cause odd system instabilities. Air/water cooling can not provide the sub-ambient temperatures needed for a ‘cold bugged’ chip to fail, so the ‘cooler the better’ rule still applies to traditional cooling methods.

For the average user, A64 temperatures below 50 degrees at load are considered ‘acceptable’ when using traditional heatsink/fan combos. At temperatures in this range you should have a pretty good amount of overclocking headroom available. Newcastle/Clawhammer chips will likely run a bit hotter, which is normal since they use a larger manufacturing process and a higher default voltage. Most mild CPU overclocks will actually impact CPU temperature very little, so long as CPU voltage is left at its default value.

All retail box A64 processors include what is commonly referred to as the OEM heatsink and fan (OEM HSF). AMD’s A64 stock coolers are actually quite good compared to past standards and employ an 80mm fan and more densely packed aluminium fins. The FX series of processors and the new dual core x2’s come with much nicer copper HSFs. They even have several heat pipes for better thermal transfer away from the CPU.

This is one of several different AMD OEM HSFs. This particular HSF came with a Sempron 2600+ and is 100% aluminium. An 80mm fan is included.

“Do I need to replace my OEM HSF with an aftermarket one if I want to overclock?”

It depends. In the socket-A days, upgrading away from the OEM HSF to an aftermarket unit with a higher flowing fan could drastically reduce CPU temperatures. Today, the A64 OEM HSF is already quite decent and the chips run so cool that most entry-mid price range ‘performance’ HSFs provide only a minor improvement. I’d encourage anyone new to overclocking to try out the OEM cooler before running out to purchase an expensive replacement. For mild overclocks and small vcore increases, it should do just fine.

The OEM HSF includes something called a ‘thermal pad’, which is essentially interface material to better conduct heat between the CPU and the heatsink bottom. This material is not the best available and is really only chosen because of its simplicity. Most aftermarket heatsinks do not include any sort of thermal pad but they do still require some type of interface material. This is required to fill the microscopic pits and other imperfections on the CPU and heatsink surfaces. There are several available materials including some of the more popular ‘Arctic Silver’ varieties. Only a ‘paper-thin’ layer should be used, as too much can negatively impact the thermal transfer.

Once the power consumption of the processor increases as a result of overclocking, the OEM HSF will likely prove to be insufficient. A higher-end cooler will be much more important to keep the core temperature down in this situation. Thermalright makes a great line of ‘air coolers’, including the popular XP90 and XP90C. Watercooling also does a great job in removing the large thermal output of those high-voltage overclocks, but is clearly not for the faint of heart. There are also many more exotic cooling methods available that will actively cool the CPU to below ambient temperatures. Phase change units and TECs (Thermo electric coolers) are the most popular ways to obtain ‘below zero’ temperatures. People with this calibre of cooling usually try to take advantage of very high CPU voltages and the 3% per 10C rule.

A CPU cooler on its own is not enough. Ensuring that you have adequate airflow in your case is also essential. Mainboard components, hard drives and other add-in cards do require some air circulation to keep cool and run reliably. If you do not have at least one exhaust fan blowing hot air out from the top or rear of the case, I’d suggest that you look into installing a fan or two. A frontal intake fan can also be very useful if your case supports one.

MediaMan did a great article on airflow and heat that would be a beneficial read if you question your case cooling.

How to Overclock a 754/939 Processor (in a nutshell)

As most ‘old school’ overclockers have probably already noted, the base HTT or ‘reference clock’ is the key to overclocking the A64 platform.

• Increase the reference clock and increase the CPU frequency.
• Increase the reference clock and increase the memory frequency.
• Increase the reference clock and increase the HTT bus frequency.

This general theory is not something new. When the CPU multiplier cannot be increased, adjusting the bus or ‘reference clock’ is the only way to increase the CPU’s operating frequency.

If you are lucky enough to own one of AMD’s flagship ‘FX’ series chips, you have been graced by a ‘completely unlocked multiplier’. With an FX chip you can simply increase the CPU multiplier to obtain a higher clock speed without changing the reference clock. The rest of us peons must increase the reference clock frequency.

Overclocking the HTT bus: Is it beneficial?

In short, overclocking the HTT bus is not beneficial. Unlike the FSB in older platforms, the HTT is in no way a bottleneck that limits system performance. There has been some benchmarking done in the past and any performance gains were often within the margin of error. The system is simply not utilizing the HTT bus to its full potential. Increasing it does next to nothing for performance and can begin to cause system instability if it is increased too much.

The goal when overclocking the 754/939 platforms is to keep the HTT as close to the default frequency as possible. On the 754 Platform, the default frequency is 800MHz/1600MHz DDR (derived from a 4x multiplier) and on the 939 Platform, it is 1000MHz/2000MHz DDR (derived from a 5x multiplier).

Once you start increasing the reference clock frequency to overclock your CPU/Memory, your overall HTT will begin to exceed its rated operating frequency. To compensate for this unwanted increase we simply decrease the HTT (or LDT) multiplier. There is a safe range that you should try to stay within:

Socket 754: 1300MHz- 1800MHz

Socket 939: 1700MHz- 2200MHz

See the next section for more information on how to calculate your HTT clock speed.

A64 Mathematics

One of the most useful tools in my overclocking arsenal is the nGEAR NG-942 electronic calculator. As with most other complicated things in life, overclocking requires some mathematical calculations. The next few sections will take you through all of the important overclocking related calculations used in the future sections.

NCIX.com threw in the above calculator along with my DFI NF4 Ultra-D mainboard. They must have thought people buying these boards were overclockers, and would need it. They were right in my case.

Calculating CPU Clock Speed

Much like how the CPUs of previous platforms ran at a multiple of the FSB frequency, the Athlon 64 overall CPU clock is determined by using a multiple of the reference clock (default of 200MHz). For example, the 1.8GHz Sempron 3100+ uses a clock multiplier of 9 to obtain its default 1800MHz clock speed.

CPU Clock = Reference Clock * CPU Multiplier

So if we were to increase the reference clock of this same Sempron to 220MHz for example, we’d have the following:

CPU Clock = Reference Clock * CPU Multiplier

CPU Clock = 220MHz * 9

CPU Clock = 1980MHz

So really, there are only two variables that will increase your CPU clock speed. Your CPU multiplier, and your reference clock frequency.

Calculating HTT Bus Speed

Much like the CPU clock frequency, the overall HTT frequency is also calculated using a multiple of the reference clock frequency. Since the HTT is double pumped, we multiply the entire result by 2.

Overall HTT Bus Clock = Reference Clock * LDT Multiplier * 2

For example, the socket 939 A64 chips have a default LDT multiplier of 5, and the default reference clock for all A64’s is 200MHz, so we can determine the default HTT as follows:

Overall HTT Bus Clock = Reference Clock * LDT Multiplier * 2

Overall HTT Bus Clock = 200MHz * 5 * 2

Overall HTT Bus Clock = 2000MHz.

As an additional example, let’s assume that we increased the reference clock to 230MHz, and decreased the LDT multiplier to 4.

Overall HTT Bus Clock = Reference Clock * LDT Multiplier * 2

Overall HTT Bus Clock = 230MHz * 4 * 2

Overall HTT Bus Clock = 1840MHz.

It is important to calculate your HTT bus speed when overclocking. Failure to do so will inevitably put a cap on your system’s potential.

Calculating Memory Clock Speed (1:1)

When a memory divider is not being utilized, the memory frequency simply operates at the same frequency as the reference clock.

Calculating your effective memory clock speed in this situation is very simple:

Memory Clock = Reference Clock

Note: This formula only holds true when using ‘whole’ CPU multipliers and no memory dividers. See the below sections for more information.

Calculating Memory Clock Speed (Using Dividers)

When using memory dividers, calculating the overall memory clock is not as simple as you may think. People are sometimes surprised to see that their memory is not operating at an expected frequency. It is a common misconception that the memory clock speed is simply a fraction of the reference clock speed. For example, a 5/6 memory divider implies that the memory should operate at 5/6 th the reference clock speed. Although this simple formula does work in some situations, there are also many situations in which it will not.

The underlying reason is simple: The memory must operate at a clean fraction (integer divisor) of the overall CPU clock speed.

For example: Assume you have increased the reference frequency to 300MHz, and that you are using an 8x CPU multiplier for a CPU clock speed of 2400MHz. When using a 5/6-memory divider, it seems logical that the memory clock might be calculated as follows:

Memory Clock = Reference Clock * Divider

Memory Clock = 300MHz * 5 / 6

Memory Clock = 250MHz

However, you may be surprised to find that your memory is only operating at 240MHz not 250MHz. This is clearly not 5/6 ths of the reference clock speed.

Since we know that the memory must operate at an integer divisor of the CPU clock speed, we need to determine which divisor the memory controller will select. We can do that by using the following formula:

Divisor = CPU Clock Speed / (BIOS Divider * Reference Clock)

Divisor = 2400MHz / (5/6 * 300MHz)

Divisor = 2400MHz / 250MHz

Divisor = 9.6 <-- not possible to use. Must be a positive integer, so round up to the next integer, which is 10

Integer Divisor = 10

So as you can see, to actually achieve 5/6 th of the reference clock a CPU clock speed divisor of 9.6 would be required. This is not a positive integer and therefore cannot be used. The memory controller simply selected the next highest integer. Using a value of 10 allows the memory to run as close as possible to 250MHz without exceeding it.

Once we know the divisor we can simply calculate the memory clock speed using the below formula:

Memory Clock = CPU Clock Speed / Integer Divisor

Memory Clock = 2400MHz /10

Memory Clock = 240MHz

Once you wrap your head around them, dividers are not so daunting. Unfortunately, we are often forced to use them especially with chips that have low CPU multipliers.

TIP: A64MemFreq 1.1, outlined in the Tools section of this article, can calculate your memory frequency for you when using memory dividers. If you are not mathematically inclined, give it a try.

Calculating the Memory Clock Speed (When Using ‘half’ CPU Multipliers)

‘Half multipliers’ (such as 8.5, 9.5, 10.5 etc) are not ‘true’ CPU multipliers. You will find yourself getting some bizarre memory behaviour when using half multipliers. Your memory will run at a slightly slower speed than expected. Many people find this puzzling, however when you look at it mathematically it is actually quite simple to understand.

Your A64 does this simply because your memory must run at a clean fraction of your CPU frequency (just like when using memory dividers). This can only be achieved when ‘whole’ (integer) multipliers are used. So when using a half multiplier your memory frequency is still running at a fraction of your CPU speed, however that fraction is the next highest whole multiplier value .

So if you choose a 9.5 multiplier, your memory runs at CPU Clock Speed / 10 , not CPU Clock Speed / 9.5 as you would expect. To calculate your memory frequency when using half multipliers you can use the following formula: (Only applicable when running memory 1:1, i.e. not on a divider)

Memory Frequency = CPU Frequency / Next Highest ‘Whole’ CPU Multiplier

Take a 3800+ processor for example, running at a 2400MHz default clock and a multiplier of 12x. The memory simply runs at ‘1/12 th ‘ the CPU frequency as can be seen below:

Memory Frequency = CPU Frequency / CPU Multiplier

Memory Frequency = 2400MHz / 12

Memory Frequency = 200MHz

If we were to use an 11.5x multiplier on this 3800+ processor, the CPU frequency would obviously decrease. The memory frequency would still be calculated based on the next highest ‘whole’ multiplier value, which is 12. Therefore, you have a lower CPU clock divided by the same CPU multiplier value, resulting in a lower overall memory clock.

Memory Frequency = CPU Frequency / Next Highest ‘Whole’ CPU Multiplier

Memory Frequency = 2300MHz / 12

Memory Frequency = 191MHz

The above theory is good to know, however my simple advice is to avoid half multipliers all together. They unnecessarily complicate things and are usually not required to obtain the best performance out of your A64.

TIP: A64MemFreq 1.1 outlined in the tools section can calculate your memory frequency when using half multipliers.

Multipliers

Unlike some Socket A chips, that are completely ‘unlocked’, AMD’s A64 platforms (with the exception of the FX series chips) do not allow free adjustment of the CPU Multiplier. Although these chips are not completely unlocked most of them are not completely locked either.

Thanks to AMD’s implementation of Cool ‘n’ Quiet technology, the CPU multiplier must be able to decrease in order to reduce the CPU’s operating frequency. So any CPU multiplier below the CPU’s default multiplier can be selected. This can be very useful when trying to utilize high reference frequencies to push memory beyond the DDR500 mark. Unfortunately, AMD Sempron processors operating at 1.6GHz and lower are 100% locked and cannot be adjusted.

Below is a listing of CPU multipliers and the chips that utilize them. Most A64 processors fall in the 1.8GHz to 2.4GHz range, with 9x to 12x multipliers.

Examples of AMD Processors and their Multipliers.

CPU Multiplier	Clock Frequency	Some Retail Examples
7	1.4GHz	Sempron 2500+ (locked)
8	1.6GHz	Sempron 2600+ (locked), A64 2800+
9	1.8GHz	Sempron 3000+, Sempron 3100+, A64 2800+, A64 3000+, Opteron 144, Opteron 165
10	2.0GHz	3200+, 3000+,X2 3800+, Opteron 146, Opteron 170
11	2.2GHz	3500+,3700+,X2 4400+, Opteron 148, Opteron 175
12	2.4GHz	FX-53, 3700+, 3800+, 4000+, X2 4600+, X2 4800+, Opteron 150, Opteron 180
13	2.6GHz	FX-55, Opteron 152
14	2.8GHz	FX-57, Opteron 154

You can get a more complete listing of A64 processors and their multipliers (including mobile chips and more).

Clearing the CMOS

It is inevitable. Sooner or later you’ll push your system too far over the stability line, resulting in what is commonly referred to as a ‘no post’. If your system can not post, you will be unable to return to the BIOS to reverse whatever settings that have caused the extreme instability.

Thankfully, all hope is not lost. Your BIOS parameters are stored in ‘volatile memory’, which is powered by a small 3V battery on your mainboard. If, for whatever reason, that power is lost, your BIOS settings will be lost and the optimal defaults will be loaded upon your next boot. You’ll have to reset all of your parameters including the time/date.

There are two ways to clear your CMOS:

1. Remove and re-insert the CMOS battery
2. Short the ‘Clear CMOS’ jumper on the mainboard (recommended)

When the Clear CMOS jumper is shorted, it has the same effect as cutting the battery power to the CMOS. Refer to your mainboard manual to find the location of your clear CMOS jumper.

In my experience, Athlon 64 mainboards are much more inherently ‘post’ stable than their socket A predecessors. The DFI NF4/NF3 series of mainboards are especially resilient in this area. In the hundreds of setting changes I have made on those boards, I have had to clear the CMOS only once.

Bus Locking

As mentioned in the A64 Mathematics section, three things are increased proportionally with the reference clock frequency: The Overall HTT bus speed, the CPU clock speed, and the memory clock speed.

These three factors are very easy to control with memory dividers, the LDT multiplier, etc. Some mainboards have what is known as ‘unlocked’ AGP or PCI buses. This means that the AGP and PCI bus frequencies also increase proportionally with the reference clock frequency. This is not desirable because increasing the AGP or PCI bus frequency does nothing for performance and simply compromises system stability. Having these bus speeds increase with the reference clock severely limits the potential system overclock. Many on-board components reside on the PCI bus including some raid controllers, serial ATA controllers, sound devices, USB controllers and more. Some people have reported data corruption on their hard drives from these out of spec bus speeds, so it is important to know your hardware.

Most high-end and enthusiast mainboards keep the AGP and PCI frequencies locked at 66MHz and 33MHz respectively (100MHz for PCI-Express). As the reference clock increases, the AGP and PCI frequencies do not change. This allows the reference clock to be increased without impacting your serial ATA drives, USB devices, sound devices and many other components.

Some mainboards have only certain components ‘unlocked’, like the DFI NF3 250GB mainboard. This mainboard requires you to use SATA ports 3 and 4 if you plan to overclock. Using ports 1 and 2 will limit your reference clock to about 240MHz before your hard drives start playing some funny games with you.

Mainboard manufacturers do not always advertise bus-locking capabilities, so I would encourage you to do your research prior to buying. Seek out reviews from enthusiast sites and look for boards that other overclockers have had good luck with when selecting your hardware.

Memory Modules

Contrary to popular belief, you do not need superior memory to overclock an Athlon 64. With memory dividers you can simply reduce the frequency of your memory as you increase the reference clock frequency. This way your CPU clock increases, but your memory frequency still stays within acceptable limits.

Having large amounts of RAM and numerous sticks does often pose much more of a problem than the quality of the RAM. These issues are much more common with older revision A64s. With three or more DIMM slots populated, your CPU’s memory controller will become stressed, even at default clock speeds. This may very well limit both your CPU and memory overclock. The 754 platforms generally reduce memory speed to DDR333 (PC2700) as soon as three DIMMs are used. With older 939 chips, DDR333 is also the default with 3/4 DIMMs. The newest Venice/San Diego chips provide better handling of >1GB of memory and allow 4 DIMMs at DDR400, however 2T command timing is defaulted. Thankfully, the latest core revisions are maturing and we are starting to see much greater overclocking success rates with 2GB of RAM. Some of the latest batches of Socket 939 and Venice/San Diego chips are actually handling 2x1GB DIMMs to almost the same degree as older platforms did with 2x512MB. New games like Battlefield 2 chew up quite a lot of resources and we’ll inevitably see more games and applications follow suit.

Memory tips for high overclocks:

• Use no more than 2 dimms for the best overclocks. One DIMM is ideal for 754.
• Stick with matched memory if using in pairs. Memory should be identical, regardless of platform.
• Check your mainboard’s manual to determine which slots to use with fewer than 4 DIMMs.
• Stick with 1GB or less, unless you need more.

For this article, we’ll be using 1x512MB, 2x256MB and 2x512MB configurations in our testing.

Memory Timings

There is more to memory than just its operating frequency. Low memory timings can make the modules more efficient, completing more work per unit of time. It would be beneficial to do some ‘pre-reading’ up on how memory actually works and what all of the timings actually mean. There is a great ‘Memory Technology’ overview at Anandtech that can help to shed some light on the subject:

http://www.anandtech.com/memory/showdoc.aspx?i=2223&p=1

For the purposes of this guide, I’ll simply outline how timing adjustments can impact stability, overclocking headroom and performance. There are literally dozens of memory timings that can be adjusted using DFI boards or A64 Tweaker. We’ll be focusing on the ‘primary’ or most common timings, which can be adjusted on just about any system and have a relatively large impact on stability and performance.

Low numerical timing values are often referred to as ‘tight timings’. Many memory modules have their timings specified in a format such as: 2.5-4-3-8 or 2-2-2-5 or some other combination of four numbers. Some types of generic memory simply state the CAS value in the form CAS2.5 or CAS3. Not all memory manufacturers use the same ordering of the timings in the specifications. The most popular order is as follows:

CAS – tRCD – tRP – tRAS

Generally speaking, the first three timings have a large impact on performance. The lower the rated timing values are, the faster the memory is.

CAS: CAS is probably one of the most well known timings and can greatly reduce or improve system stability and performance when adjusted. Tighter CAS values are best and most high-performance DIMMs use a CAS value of 2. Commonly found values are 2, 2.5 and 3. It is not advisable to use a CAS value higher than 3.

RAS to CAS (tRCD): tRCD is a sensitive timing and can greatly reduce or improve system stability and performance when adjusted. Most PC3200 DIMMs have a tRCD of 2 or 3. This is generally the first timing I loosen when trying to achieve higher memory clock speeds. It can greatly increase the overclocking headroom available when loosened. For the best memory performance, tRCD should be set as low as possible. I consider ‘4’ to be the highest tRCD value that should be used.

Row Precharge (tRP): Much like tRCD, tRP is also a sensitive timing and can greatly reduce or improve system stability and performance when adjusted. Most PC3200 DIMMs have a tRP of 2 of 3. For the best memory performance, tRP should be as low as possible. tRP values of 4 or higher should be avoided if possible.

tRAS: tRAS is a bit of an ‘odd-ball’ when it comes to performance. In the Socket A days, it was discovered that 2-2-2-11 timings (tRAS=11) were actually better than 2-2-2-5 in many synthetic benchmarks, proving that lower is not always better. tRAS is generally not too sensitive and can be adjusted a bit in either direction to look for the best performance. In the world of Athlon 64, tRAS of 6-7 seems to be ideal. Anything lower than 5 or higher than 8 seems to worsen performance. Many people with high performance 2-2-2-5 memory increase their tRAS to 6 or 7 to extract a bit of extra performance. There has always been some debate as to whether these ‘optimized’ tRAS values are actually better for gaming and real world use. Some have speculated that they are simply useful to improve synthetic benchmark scores and that lower is always better.

Command Timing/Command per Clock (CPC): CPC is not a memory timing in the same sense as CAS/tRCD etc. CPC is actually a function of the memory controller, which allows it to handle numerous banks/DIMMs. Some manufacturers specify 1T at the end of the primary timing list, for example: 2-2-2-5-1T. This is an illogical specification because any DDR module is capable of 1T timing. Command timing has nothing to do with the memory module itself, but rather the platform that it is being used in. When you combine too many modules, the memory controller will use a 2T or ‘CPC Disabled’ timing to reduce stress on the controller. CPC has a large impact on performance, and typically reduces memory performance by 15-20% when set to 2T (according to synthetic benchmarks). In gaming, CPC Disabled usually results in a 3-5% performance decrease. CPC Enabled (1T) should always be used unless your memory controller cannot support your number of modules or the amount of RAM in your system. I recommend against using CPC Disabled as a means to obtain a higher memory overclock. Although your memory may clock higher, it is rarely enough to offset the performance degradation caused by 2T timing. In some situations, 2T timing is required with 2x1GB of memory to maintain stability.

When trying to obtain higher memory clock speeds, I usually loosen timings in the following order:

CAS-tRCD-tRP

• 2-2-2
• 2-3-2
• 2-3-3
• 2.5-3-2
• 2.5-3-3
• 2.5-4-3
• 3-4-3
• 2.5-4-4
• 3-4-4

tRAS is not listed above because it can usually be kept constant without greatly impacting stability. Please note that not all memory types operate well at loose timings. Some of the old BH-5 based modules simply refuse to work at anything looser than CAS 2.

CPU Stepping Codes and CPU Selection

There is a lot more to an A64 than just it’s PR rating (2800+, 3800+ etc). Over the years AMD has released many different core revisions. These revisions covered off on different socket types, new manufacturing processes and other changes which may be relevant to a CPU’s overclockability.

Take the AMD Athlon 64 3200+ for example; there were literally six different varieties of it over the years:

Athlon 64 3200+ Core Variations

PR Rating	L2 Cache Size	Clock Frequency	CPU Core	Socket Type
3200+	512KB	2.2GHz	Newcastle	754
3200+	1MB	2.0GHz	Clawhammer C0	754
3200+	1MB	2.0GHz	Clawhammer CG	754 (Mobile)
3200+	1MB	2.0GHz	Clawhammer C0	754 (Mobile)
3200+	512KB	2.0GHz	Winchester	939
3200+	512KB	2.0GHz	Venice	939

Aside from a ‘PR’ number, seemingly identical AMD processors have many properties that are printed on the processor itself. This data is often referred to as an ‘OPN’ or ‘Ordering Part Number’. Understanding AMD’s naming conventions and core types can assist when you are looking to buy your next hot overclocker. You have surely heard of a good ‘Stepping Code’ or ‘Core Revision’ on a forum at one time or another.

Check out the following site for some great information on how to decipher OPNs:

http://fab51.com/cpu/guide/opn-64-e.html

AMD was kind enough to keep the heat spreader visible in the retail box package so that you can inspect its OPN prior to buying. Let’s take my 3500+ for example:

These are the first two lines printed on the chip:

ADA3500DIK4BI

CBBID 0450XPMW

ADA = Athlon 64 Desktop Processor

3500 = Model Number of Processor (3500+)

D = Socket 939 w/Heat Spreader

I = 1.40V default vcore

K = Maximum case (inside of heat spreader) temperature of 65 degrees

4 = 512KB Cache Memory

BI = Winchester Core, Revision D0

This first line of information tells me just about everything I need to identify the processor in AMD’s product list. The most important piece of information to an overclocker is likely the core type. In my case, I actually went into my hardware store and inspected the processor to ensure that I got a ‘ BI ‘ 3500+. At that time there were quite a few Newcastle ‘AX’ 3500’s on the market, which I did not want.

On the second line there is some additional information that can vary from chip to chip.

CBBID = Stepping Code

0450 = Production Week, 50th week of 2004

XPMW = Additional Production Information

On many forums you’ll see overclockers praise a certain stepping or week of processor. When I was looking to purchase my Winchester, I saw quite a few 0450’s hitting 2.6+ GHz and I jumped for that specific production week. This is not always a perfect indication of overclockability as chips can vary greatly in their quality. As many veterans have stated over and over again: “getting a good overclocker is often just the luck of the draw”.

We can however, make some educated assumptions about a chip’s overclockability based on the first line of text. Generally speaking, newer core revisions overclock better than older revisions and often require less voltage. Below is a table that I have produced which outlines my personal views on overclockability by core. This information is not hard fact, however most of the results I have seen fall within the Min/Max clock frequencies I have listed below. This is assuming normal ‘air cooling’ is used of course. Extreme cooling methods and high vcore will often equate to higher clocks.

Overclockability by Core Revision

Socket	Core	Rev	Vcore	Last 2 characters of OPN	Min OC Avg	Max OC Avg	Comments
754	Clawhammer	C0	1.5V	AP	2.1GHz	2.3GHz	Least desirable chips for overclocking. These were the original A64 models. 2T command memory timing not possible.
754/939	Clawhammer/ Newcastle/ Paris	CG	1.5V	AR, AS, AW, AX	2.2GHz	2.6GHz	Large improvement over the C0 revision. AMD used this revision for quite a while. Newcastle (AX) chips are essentially 512K CG revision processors, and the remainder are 1MB cache revisions.
939	Winchester/ Palermo	D0	1.4V	BI, BA	2.2GHz	2.7GHz	Overclocks can vary greatly depending on production week. Some steppings have very weak memory controllers. Many do not react well to very low sub-zero temperatures. Lower 1.40V vcore, and low power consumption compared to Newcastle/Clawhammer chips. The Sempron BA’s are usually made of lower grade silicon and do not overclock as well.
939	Venice/San Diego	E3/ E4/ E6	1.35/ 1.4V	E4 = BN E3 = BP, BO	2.4GHz	2.9GHz	More modern core revisions, which generally have a lot of headroom for overclocking. Better memory controllers and better support for >1GB memory. Operates at 1.35 or 1.40 vcore.
939	Manchester/ Toledo	E4/ E6	1.35/ 1.4V	E4 = BV E6 = CD	2.3GHz	2.7GHz	X2 Processors are similar to the single core ‘E’ revisions, however usually have less headroom due to increased heat, and other challenges that come along with two cores.

So as you can see above, I have seen a great deal of variation from one core to another, however we can generalize that CO < CG < D0 < E3 < E4 in most situations. Another important factor to take into account is the actual PR rating. A 3800+ Newcastle will most certainly be produced using higher-grade silicon than a 3200+ Newcastle, and that will be reflected in the overclocking results. Sometimes the opposite will occur, however for the most part, you get what you pay for (when comparing the same core revision, with different PR ratings). When you look at different cores however, there is a very good possibility that a 3000+ Venice will push further than a 3500+ Newcastle. This is simply because of a much newer, fine-tuned manufacturing process.

Every core revision can have a lot of variation and what I have depicted above is merely my personal experience as an enthusiast. As always, your mileage may vary. Remember, each revision also brings new features to the table not just greater overclocking headroom. SSE3 instructions and better memory controllers are probably the ‘E’ family’s greatest benefits.

Mobile Athlon 64 Processors

Just about any veteran overclocker remembers the overclocking success stories of the Mobile Barton XP processors a few years ago. Mobile processors must be built to withstand the high temperatures present in a laptop and consume less power than their desktop counterparts. AMD often uses higher-quality silicon in these mobile processors to ensure that they are literally running at a fraction of their clock speed potential. This will help the processor to run trouble free at high temperatures and lower vcore voltages. It did not take overclockers long to realize that they could harness this unused headroom, through the use of higher vcore, and the nice ‘comfortable cooling’ of a large desktop class HSF.

The mobile Athlon 64 processors are single channel, socket 754 based parts. Most of them are CG revision clawhammers with 1MB of L2 cache. AMD has also recently released the ‘Turion’ series of 90nm mobile processors. These chips are definitely of higher-grade construction than their desktop counterparts; however their overclocking headroom is not quite as stretched out as the old mobile Barton XP processors.

One of the most obvious physical differences between the Mobiles and the Desktops is their lack of a protective heat spreader. This allows the CPU core to be exposed directly to the laptop heatsink for better cooling and for lower weight. Unfortunately, this physical difference also makes the CPU a millimetre or so shorter than a ‘hooded’ desktop chip. There are only a few HSFs on the market that can actually provide proper contact on the core. The SOI based A64 core is very fragile so it is important to use extreme care when mounting your cooling device.

Mobile Athlon 64 processors also require some kind of compatibility from the mainboard. There are quite a few mainboards that will not support mobile processors. Many others do not ‘officially’ support them but they do indeed operate correctly. This will require some research on your part.

There is a somewhat old but very informative thread on the Anandtech forums that has more information if you are interested.

‘Other’ Hardware Selection

Aside from the A64 processor itself, the most important component with respect to overclocking is the mainboard. Just about every overclocking adjustment that can be made depends on support from the mainboard. Many lower end or non-enthusiast mainboards have few overclocking-friendly options and have difficulty maintaining suitable component temperatures, and voltage regulation.

As mentioned earlier, the reference clock is the key to increasing the CPU’s operating frequency. In order to obtain high overclocks, (especially with 3200+ and slower processors) you’ll need a mainboard capable of reference clock frequencies above 250MHz. Most high quality mainboards should be able to maintain stability at reference clock speeds exceeding 300MHz.

NOTE: The examples in this article assume the use of a high quality mainboard, capable of reference clock speeds exceeding 250MHz. If you encounter odd refusals to ‘Post’ after a certain reference clock is exceeded, your mainboard has likely reached it’s limits. Also, refer to the ‘A Quick Note on Bus Locking’ section for some more information on mainboard limitations.

By far, the most important thing to do is research your potential purchases. There are many enthusiast/overclocking websites online that review mainboards on a regular basis. There are dozens of boards out there that will do the job very well and simply far too many for me to mention within the scope of this guide.

DFI currently holds the crown of A64 overclocking, with their NF3/NF4 series of mainboards. Other manufacturers also have great offerings, including EpoX, Abit, MSI, Asus and others.

Generally speaking, you should look for the following in an enthusiast mainboard:

• Modern chipset platform, such as the Nforce4 or Nforce3 250GB.
• Locked AGP/PCI/PCI-E bus (see bus locking section)
• Stability at high reference clock frequencies above 300MHz.
• Wide range of vcore selections
• Wide range of vDIMM selections
• Voltage/Temperature probes on board (for use with monitoring software)
• Quality chipset cooler, and some form of power regulation cooling is also beneficial
• ‘Heavy Duty’ power regulation circuitry

To extract the best performance out of your system, quality PC3200 or better memory is recommended. Low end memory will not usually limit your CPU’s overclock, but will affect your overall system performance. High clocking memory such as TCCD is a great match for the Athlon 64, as well as tight timing memory like BH5 and UTT. Again, thoroughly research your selection prior to buying. There are countless online reviews of overclocker friendly memory around. We’ll be using two examples of OC friendly RAM later on in this guide.

Companies such as OCZ, Corsair, G.Skill, Mushkin and Crucial are all very reputable manufacturers who cater to the hardware enthusiast.

Research is important. You can never read up too much before you buy. Be sure not to base your buying decision on just one article. It is always a good idea to get a second perspective on the same product.

If you ever have any doubt about a product, or if you are unsure, feel free to ask one of Short Media’s veteran overclockers in the ‘Overclocking Forum‘.

Next section: Overclocking Tools

Previous section: Introduction