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

Step 1: Getting Started

Now that you have a better understanding of the A64 platform, A64 overclocking theory and the tools you'll be using, we will begin the overclocking process.

TIP: One thing that I'd like to stress right away is that taking the 'proper and logical' steps can be a very time consuming process. Patience is very important. If you begin rushing through the steps, you'll simply waste more time in the long run and frustrate yourself in the process. Following the logical steps that I outline will give you a very thorough picture of what your system is capable of.

Qualifying Your System for Overclocking

It is always a good idea to get some temperature/voltage readings and verify system stability before you begin. We want to be sure that the system is 100% stable and capable of withstanding the increased electrical and thermal stresses that overclocking will bring.

You may think that this step is a waste of time, but you'd be surprised just how many systems have bad sticks of RAM in them or exhibit all sorts of thermal and electrical problems from the get-go.

The following should be checked:

CPU Idle Temperature : Idle CPU temperature is simply 'when at practically 0% utilization'. To check this, simply boot up your system, allow it to sit for 10-15 minutes and open up your favourite Monitoring application (see the 'Tools' section) to check.

• Good: Low 40s or lower
• Caution: Mid to High 40s.
• Unacceptable: Above 50 degrees.

CPU Load Temperature: CPU Load temperature is your CPU's temperature when at 100% utilization. To check this, start up Prime95 or OCCT and let it run for approximately 15 minutes. Open your favourite monitoring tool while the stress test is running to check your CPU temperature.

• Good: Mid to high 40s or lower
• Caution: Low to mid 50s
• Unacceptable: Higher than mid 50s.

Power Supply Voltages: Boot into your operating system as you normally would and use your system monitoring application (MBM5, Smart Guardian, etc) to monitor the system voltages. The ATX specification states that a PSU should supply voltage with +/-5% of the rated rail voltage. At the minimum, you should ensure that your voltage rails are within the below ranges. If they are not within spec, I would strongly recommend an immediate PSU replacement.

ATX Voltage Specifications

Output Rail	Minimum	Maximum
+3.3V	3.14V	3.47V
+5V	4.75V	5.25V
+12V	11.4V	12.6V

Be sure to verify your voltages using a digital multi-meter before tossing your PSU in the trash. Mainboard voltage sensors can often be off. Also, check the voltages reported in the BIOS and compare them to your monitoring application. Sometimes there can be discrepancies between software applications.

Check for voltage fluctuations as well. If you see any of the voltage rails constantly dipping and spiking, this could be an indication of a malfunctioning power supply. Minor fluctuations are normal (0.01 to 0.05 or so is normal in either direction).

TIP: Set your monitoring application's refesh/sampling interval to a shorter period of time (i.e. 1 second instead of 10 seconds). This will make it much easier to spot fluctuations.

TIP: Voltages can often differ between CPU idle and CPU load. Be sure to check them before and during CPU stress with Prime95/OCCT. If your voltages are close to any of the above thresholds, you should consider a new PSU. See the 'power' section for more information.

Memory Stability at OEM specs: Run a few loops of Memtest86+ (guide here) to verify that your DIMMs are stable at stock clocks. At the very minimum, run a complete loop. This should only take about 15 to 20 minutes, and will help you sleep a little easier.

Any failures observed at OEM specified speeds likely indicate bad RAM, or power issues. Be sure to check each stick individually in your system if you do encounter any errors. Often only one stick is bad.

Test for CPU Stability: Use Prime95 or OCCT to test for any CPU calculation errors at default clock speeds. Although it is unlikely to find any problems, it is good practice. There have been bad batches of processors to hit the market in the past; including a few of the early Winchester based processors.

Testing Configuration

For my examples, I'll be using two different systems. One is a more budget-minded 754-based system and the other is a mid-to-high end 939-based system.

Budget 754 Overclocking

Let's face it, not everyone can afford high-end Athlon 64 processors. This does not mean that you can't be as much of a hardcore overclocker because you bought a Sempron 2600+. It could be worse: you could have bought a Celeron. At 1.6GHz and with only 128KB of L2 cache, this chip is one of the slowest in AMD's socket 754 arsenal. Let's see if we can change that.

System Configuration:

• AMD Sempron 2600+ (Socket 754, 1.6GHz, 128K L2. Palermo Core)
• DFI Lanparty NF3 250GB Mainboard
• 80GB Seagate 7200.7 IDE Hard Drive
• ATI Radeon 9250 Graphics Card
• Antec Smartpower 350 (21A +12V rail)

Cooling:

• OEM AMD Heatsink and Fan
• 120MM Intake, 120MM Exhaust in an Antec mid-tower.

NOTE: The CPU temperature is not reported correctly on my DFI NF3 250GB. It is reported about 6 or 7 degrees too cold at all times. Please keep this in mind when looking at my results with the OEM AMD HSF. I will reflect this in all of my recorded temperatures, but the screen shots will not match up.

Mid-High Range 939 Overclocking

The Athlon 64 3500+ is not the fastest chip in AMD's arsenal, but it is a good performing mid-range processor. Let's see how well it does when paired up with higher end hardware, like the PC Power and Cooling 510, and DFI NF4 mainboard.

System Configuration:

• AMD Athlon 64 3500+ (Socket 939, 2.2GHz, 512K L2. Winchester Core)
• DFI Lanparty NF4 Ultra-D Mainboard
• 2x36GB Western Digital Raptor drives in Raid-0 on nvraid
• 160GB Seagate 7200.7 Storage SATA drive
• ATI X850XT Graphics card
• PC Power and Cooling 510 Express (34A +12V rail)

Cooling:

• Thermalright XP90 (Alu) w/Retail 92mm 55cfm fan.
• 2x80mm intake, 2x80mm exhaust fans in a Lian Li PC65B case

Memory Testing

I will be using several different types of memory for this guide. Each of the below memory types are fairly unique, and behave differently when overclocking.

• 2x512MB OCZ Platinum EL Rev.2 PC3200 (Samsung TCCD)
• 2x256MB Kingston HyperX PC3500 (Winbond BH-5)
• 1x512MB Kingston Value RAM (Samsung TCCC)

Go to table of contents

Step 2: Find the Maximum Stable CPU Clock

Since there is such a close relationship between memory and the Athlon 64 processor (thanks to it's on-die memory controller), our first goal is to determine the maximum stable CPU clock attainable. To do this, we'll take memory out of the picture by running it on a divider and using the 2T (CPC disabled) command timing. This will greatly reduce the stress on the CPU's memory controller and will allow us to see just how high his CPU can clock. These memory settings will significantly reduce the system's performance, but we are not concerned about sheer 'overall' speed at this point, just the maximum CPU clock attainable.

Getting Started:

Since we want to basically take memory 'out of the picture', we'll do the following right off the bat:

• Impose a memory divider : We'll force the memory to operate at a fraction of its rated frequency. This will be much less demanding on the memory controller. Using the 2/3 divider (DDR266) is my preference. This essentially causes your memory to behave like PC2100 rather than PC3200.
• Set 'Command Rate' (sometimes called 'CPC') to 2T or 'CPC Disabled': This usually results in a 15-20% memory performance decrease on the A64 platform.
• NOTE: If you have an older C0 revision 'Clawhammer' processor, 2T command rate is not supported. You'll have no choice but to use 1T.
• Manually set CAS, tRCD, tRP and tRAS timings in the BIOS : This may sound unnecessary, but we want to ensure that we have predictable memory timings. Some memory modules have different SPD timings depending on the memory's clock speed, so we don't want any surprises. Some more experienced overclockers may want to 'loosen' their timings here (optional), which can further take the stress off the memory controller. Since not all memory operates well at loose timings, I'd simply recommend the usage of default PC3200 timings. If you are unsure of your default memory timings, simply boot into Windows, launch CPU-Z, and check the 'Memory' tab.
• OPTIONAL: Increase the chipset voltage just a touch: The NF3 250GB chipset operates at 1.6V by default, so I'm going to increase it just a touch to 1.7V. This helps to enhance the 'board stability' a bit when pushing the reference clock. Most enthusiast mainboards will operate just fine at very high reference clocks and default vchip, but we're just trying to eliminate variables from the get-go. If your mainboard does not support this type of voltage increase, don't worry because it is not critical.

Once memory is 'out of the picture' we'll begin the overclocking process by doing the following:

• Slowly (incrementally) increase the reference clock frequency : Increase the reference frequency incrementally.
• Adjust the HTT (LDT) Multiplier as necessary : As we increase the reference clock frequency, we'll do some calculations, and ensure the overall HTT bus frequency stays within acceptable limits. LDT multiplier will be adjusted as necessary.
• Stability Testing after each increase : After each reference clock increase, we'll boot into windows to do some stability testing, as well as recording our temperatures, etc.
• The reference clock is continually increased until we see our first stability failure : Once we see our first stability failure, we'll bring vcore into the picture and adjust accordingly.

NOTE: Do not increase your vcore (CPU voltage) right off the bat. This is recommended so that we can see how well our CPU scales with voltage. It will be increased later on. We'll start with default voltage.

NOTE: To use as an example, I'll be going through the steps using the socket 754 hardware listed in the 'Testing Configuration ' section. The memory being used is irrelevant for this section, however I will be using Kingston HyperX PC3500 in this machine.

Go to table of contents

Finding Your Maximum CPU Clock Speed

First things first: enter the BIOS.

Once you are in the BIOS, you'll usually be greeted with several options at a main menu. Refer to your mainboard manual to determine where your HTT and specific memory and CPU parameters are located.

This Phoenix AwardBIOS menu is a popular one that many people will find familiar. With DFI mainboards, the overclocking parameters can be found in the 'Genie BIOS Setting' submenu.

NOTE: DFI, along with many other manufacturers, incorrectly call the reference frequency 'FSB'. So whenever you see 'FSB' in the BIOS screenshots, please interpret that as 'reference clock'. LDT is synonymous with HTT, as mentioned earlier. DFI uses 'LDT' instead of HTT all over the BIOS.

The DFI NF3/4 boards have their memory (DRAM) parameters under the 'DRAM Configuration' submenu. Again, refer to your mainboard manual for the DRAM/memory settings location in your BIOS.

The memory timings available for modification in the DFI BIOS can be quite intimidating at first glance. We'll be leaving most of them at 'Auto' for this step. Only the more common timings need to be forced at this point including CAS (tCL), RAS to CAS (tRCD), Row Precharge (tRP) and Minimum RAS active time (tRAS).

2-2-2-5 timings were used for my BH5, as I know it is stable at 200MHz with these timings.

As you can see above, the following was set:

• DRAM Frequency Set (MHz): 133 : This is a 2/3 divider, which will force our 200MHz PC3200 to operate at roughly 2/3 its rated frequency (See 'A64 Mathematics' section for more information). This is exactly what your system would default to if you used PC2100 (DDR 266) RAM in your A64 system.
• Command Per Clock (CPC): Disable : This forces the command timing of 2T, and greatly reduces stress to our memory controller. If you use more than two DIMMs in your system, this will be your default value already.
• tCL, tRCD, tRP and tRAS: Set to 2-2-2-5 respectively . These are known stable timings for the BH5 in my test system. You should use your RAM's default timings that you looked up in CPU-Z. Experienced overclockers can use loose timings here that are known to be stable at 200MHz.

NOTE: Increasing the chipset voltage (sometimes called vchip) can sometimes help when greatly increasing the reference clock frequency. If your mainboard supports this adjustment, increase it slightly. Any more than a 0.1V increase is not necessary, and just creates extra heat. This step should be considered optional. If your mainboard does not support this adjustment, simply disregard this step. Vchip increases were much more beneficial with the older NF2 platforms than with the latest batch of A64 chipsets.

Once you have made these changes, save and exit the BIOS. Boot into your operating system as you normally would. Once in Windows, launch CPU-Z and click the 'Memory Tab'. You should see that your memory is now running at about 133MHz as opposed to 200MHz (refer to the 'A64 Mathematics' section for more information on memory dividers).

Now comes the fun part. Let's start cranking up the reference clock.

Since our test subject is a lowly 2600+ Sempron with a rather low 8x multi, I'll have to make some rather large reference clock jumps to greatly impact the CPU clock. For the first increase, I usually take a fairly large step up. In this case I'll increase it by about 15MHz. The reason for this larger increase is that just about every chip has a fairly good amount of initial headroom. No sense increasing 2MHz at a time, it would simply take too long in the early steps. If your chip has a higher multi (11 or 12 for example), I'd select a smaller 10MHz initial increase, as any changes you make will have a larger impact on the overall CPU clock speed.

Let's take a look at exactly what I did with this simple reference clock change before I leave the BIOS.

TIP : I'm no math whiz, but this is not difficult to work out if you read the 'A64 Mathematics' section.

As mentioned in the previous sections, increasing the reference clock (in our case to 215MHz) also increases the following: CPU clock frequency, Memory clock frequency, and the overall HTT bus frequency. Lets calculate:

CPU Clock = Reference Clock * CPU Multiplier

CPU Clock = 215MHz * 8

CPU Clock = 1720MHz

The clock speed has been increased from 1.6GHz to 1.72GHz. What about the Memory?

Since I am using a memory divider, I first need to find the memory divisor that will be used by the memory controller.

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

Divisor = CPU Clock Speed / (2/3 * 215MHz)

Divisor = 1720MHz /143.3

Divisor = 12

Twelve is indeed an integer divisor (positive whole number), so no rounding up is necessary. The memory clock speed can now be calculated as follows.

Memory Frequency = CPU Clock Speed / 12

Memory Frequency = 1720MHz / 12

Memory Frequency = 143.3MHz

TIP: if you are not good with math, you can use the A64MemFreq tool outlined in the tools section. As you can see, our Memory is still running at a fraction of its rated speed, but it also increased as a result of our reference clock increase.

What about the HTT bus?

HTT Bus = Reference Clock * LDT Multiplier * 2

HTT Bus = 215MHz * 4 * 2

HTT Bus = 1720MHz

As expected, the overall HTT bus speed is now overclocked by about 120MHz. I did not decrease the HTT multiplier to 3x yet, simply because the HTT would be a bit too low at only 1280MHz. Remember to keep your overall HTT within the 'safe range'. Refer to the 'Overclocking the HTT bus' section for more information.

When finished, save your changes, exit the BIOS and boot into Windows. With any luck, you'll see the below screenshot and not a blue screen of death.

Once you see your Windows desktop, you can breathe a sigh of relief. You have passed the first stability test: booting into windows. If your system were terribly unstable, you wouldn't have much luck getting even this far.

The first thing I like to do when in Windows is to verify my settings within CPU-Z.

As you can see above, it is exactly what we expected. The Sempron testing machine is now indeed running at 1720MHz. The memory is also indeed running at 143MHz as a result of the changes.

Now I'm going to do some basic stability testing using the tools mentioned earlier. 1720MHz is pretty useless if it can't be used for anything other than CPU-Z screen shots. For my examples, I'll be using Prime95, but you can use OCCT if you like. Since we'll be doing quite a few more increases, we'll be stability testing for a relatively short period of time. I usually let Prime95 finish the 1024K self-test, using the 'Small FFT' torture test. This process takes about 15-20 minutes, and will give you a good indication of 'initial stability'. While running the test, I'd strongly recommend keeping an eye on your CPU temperature and system voltages. This small increase will likely have a minimal effect, but it is good practice to keep an eye out.

I let it run for about 15-20 minutes and all appears well. The temperatures seem unchanged from stock clock speeds, as do the system voltages. Clearly 15 minutes of Prime95 is not an indication of 100% stability, but it provides me with some reassurance that we can continue to push the machine harder, through further reference clock increases.

I'm always pretty particular about recording all of my results as I progress. Recording metrics like load temperatures, memory clock, and HTT clock can help later on. Below is a chart of my results so far. I'd suggest you put something similar together to keep track of your progress. Refer to the 'A64 Mathematics' section if you are unsure how to calculate any of the values.

CPU Overclocking: Progress Chart

REF	MEM CLK	CPU CLK	HTT BUS	LOAD TEMP	VCORE	CPU GAIN	15 min P95/OCCT?
200MHz	133MHz (2/3)	1600MHz (8x)	1600MHz (4x)	44'C	DEFAULT	0%	PASS
215MHz	143MHz (2/3)	1720MHz (8x)	1720MHz (4x)	44'C	DEFAULT	7.5%	PASS

So far, things appear to be going well. Obviously, we're not going to just settle for a 7.5% CPU clock gain. Once the 15 minute Prime95/OCCT testing has been done, we'll return to the BIOS for another increase.

I'm going to take one last 15MHz reference clock step. After this, I'll take smaller 10MHz increase steps instead. Remember that if you are using a 3500+ or higher A64, you should use smaller 5-10MHz increments to make up for your higher CPU multiplier. I'm simply going to increase the reference clock to 230MHz, and leave everything else in the BIOS as-is.

Now that the HTT bus speed has increased even more, it would be a good time to reduce the HTT multiplier. The HTT bus speed would now be 1840MHz if we kept the LDT multiplier at 4x, which I would consider too high. At this point, I'm going to decrease it one step downwards. Don't forget to calculate your HTT bus frequency every time you increase the reference clock.

As can be seen above, I decreased the LDT multiplier to 3x. The Sempron's default LDT multiplier is 4x. If you have a 939 processor, you would decrease it one step downwards to 4x.

After applying my changes, I was able to boot into Windows just fine. I was able to run Prime95 for 15 minutes without any issues. Here is an updated chart with my results:

CPU Overclocking: Progress Chart

REF	MEM CLK	CPU CLK	HTT BUS	LOAD TEMP	VCORE	CPU GAIN	15 min P95/OCCT?
200MHz	133MHz (2/3)	1600MHz (8x)	1600MHz (4x)	44'C	DEFAULT	0%	PASS
215MHz	143MHz (2/3)	1720MHz (8x)	1720MHz (4x)	44'C	DEFAULT	7.5%	PASS
230MHz	153MHz (2/3)	1840MHz (8x)	1380MHz (3x)	45'C	DEFAULT	15%	PASS

Now we are starting to see some larger clock speed increases. A 230MHz reference clock has given us a 15% increase in CPU clock speed. Our Sempron has now exceeded the clock speed of the 'Sempron 3000+'. This processor has 128K of cache like the 2600+ and a 1.8GHz default clock speed.

For the next reference clock increases, I'm going to stick with smaller 10MHz (5MHz if you have a 3500+ or higher chip) increments until I get a Prime95 failure. Simply continue the process described above, and update your table with updated data. You'll want to keep increasing until you see a prime95 error. Here's how far I got:

CPU Overclocking: Progress Chart

REF	MEM CLK	CPU CLK	HTT BUS	LOAD TEMP	VCORE	CPU GAIN	15 min P95/OCCT?
200MHz	133MHz (2/3)	1600MHz (8x)	1600MHz (4x)	44'C	DEFAULT	0%	PASS
215MHz	143MHz (2/3)	1720MHz (8x)	1720MHz (4x)	44'C	DEFAULT	7.5%	PASS
230MHz	153MHz (2/3)	1840MHz (8x)	1380MHz (3x)	45'C	DEFAULT	15%	PASS
240MHz	160MHz (2/3)	1920MHz (8x)	1440MHz (3x)	45'C	DEFAULT	20%	PASS
250MHz	167MHz (2/3)	2000MHz (8x)	1500MHz (3x)	45'C	DEFAULT	25%	PASS
260MHz	172MHz (2/3)	2080MHz (8x)	1560MHz (3x)	45'C	DEFAULT	30%	PASS
270MHz	180MHz (2/3)	2160MHz (8x)	1620MHz (3x)	46'C	DEFAULT	35%	PASS
280MHz	187MHz (2/3)	2240MHz (8x)	1680MHz (3x)	46'C	DEFAULT	40%	FAIL

So at a reference clock of 280MHz, I had a Prime95 error after less than a minute of stability testing. The machine was stable enough to boot up, and appeared stable, however once again, Prime95 had the final word in stability. It made it nowhere near the 15 minute threshold that it did at 270MHz. Often times with A64 chips, you'll reach a limit very suddenly, and only a few MHz of reference clock can make all the difference in stability. Clearly, for some reason, the chip is not stable at this speed. (See below screenshot).

Why did it fail? If we refer to my OC chart above, the memory was still significantly below its default clock speed, so that is an unlikely candidate. The HTT is also still within reasonable limits, albeit, overclocked by 80MHz. Setting the LDT multiplier to 2x only made things worse, and the system did not even get into Windows. The CPU temperature is only two degrees hotter than stock clock speeds, and still well within reasonable tolerances.

Really, at this point the CPU clock speed has simply been stretched too far on too little vcore. If we want to continue onwards we'll have to begin increasing the CPU core voltage. Let's see what happens when we increase the stock vcore of 1.4V to 1.45V.

CAUTION: Increasing your CPU voltage will increase the heat output and power consumption of your CPU. Please proceed with caution, and be very observant of your CPU temperature and system voltages. Remember, you don't HAVE to go beyond this point, many people would be very happy with the 35% clock speed increase we have achieved on default vcore. If you want to see just how much more headroom your CPU has, then let's keep going.

At this point, I'll enter the BIOS, and increase the vcore (or CPU VID as per DFI) by 0.05V. In my example, the Sempron 2600+ default vcore was 1.4V. If you have a 130nm core, your default vcore will be 1.5V and you should increase it to 1.55. If you have a Venice/San Diego or newer dual-core processor, your default voltage may be as low as 1.35V, in which case you should increase it to 1.4V.

I'm going leave everything else as-is and retry the 280MHz reference clock. Hopefully with the touch of extra 'juice' it'll pass the 15 minute initial stability test.

It appears that 2241MHz was too much to ask for with only a 0.05V increase. But, if you look at the Prime95 window above, you'll see that the test ran a full 13 minutes as opposed to several seconds. Clearly, stability has improved as a result of the vcore increase. If you take a close look at the above screenshot, you'll also notice that the vcore is only 1.42V. The vcore seemed to fluctuate a bit between 1.45 and 1.42 volts. Prime95 likely failed on one of those low dips to 1.42 volts. It looks like we'll have to increase it a bit higher. This time I'm going to try a full 0.1V increase in the BIOS and set the vcore to 1.5V. We'll likely see a larger temperature increase as a result of this vcore escalation. I'll be keeping a close eye on temperatures at this point.

NOTE: This situation is a perfect example of why a high quality power supply is important. The Antec SmartPower 350 is a decent PSU, but not intended to be enthusiast grade, with rock solid voltages to two decimal places.

Once I ran the stability test again at 1.5V, the system had no problem tearing apart the 1024K self test in Prime95. There was still a bit of vcore fluctuation (between about 1.48 and 1.52 volts), but even the low dips were not enough to cause instability. Obviously, 15 minutes of Prime95 is not an indication of absolute stability, but it's enough for us to move a little bit further ahead.

Moving forward, use much smaller reference clock increases. For the 2600+ Sempron, I'll use 5MHz increments from this point forward. If you have a 3500+ or higher processor, you should use smaller 3MHz increments.

285MHz seemed to do well at 1.50V as well, but 290MHz failed. Here is an updated chart:

CPU Overclocking: Progress Chart

REF	MEM CLK	CPU CLK	HTT BUS	LOAD TEMP	VCORE	CPU GAIN	15 min P95/OCCT?
200MHz	133MHz (2/3)	1600MHz (8x)	1600MHz (4x)	44'C	DEFAULT	0%	PASS
215MHz	143MHz (2/3)	1720MHz (8x)	1720MHz (4x)	44'C	DEFAULT	7.5%	PASS
230MHz	153MHz (2/3)	1840MHz (8x)	1380MHz (3x)	45'C	DEFAULT	15%	PASS
240MHz	160MHz (2/3)	1920MHz (8x)	1440MHz (3x)	45'C	DEFAULT	20%	PASS
250MHz	167MHz (2/3)	2000MHz (8x)	1500MHz (3x)	45'C	DEFAULT	25%	PASS
260MHz	172MHz (2/3)	2080MHz (8x)	1560MHz (3x)	45'C	DEFAULT	30%	PASS
270MHz	180MHz (2/3)	2160MHz (8x)	1620MHz (3x)	46'C	DEFAULT	35%	PASS
280MHz	187MHz (2/3)	2240MHz (8x)	1680MHz (3x)	46'C	DEFAULT	40%	FAIL
280MHz	187MHz (2/3)	2240MHz (8x)	1680MHz (3x)	48'C	1.45V	40%	FAIL
280MHz	187MHz (2/3)	2240MHz (8x)	1680MHz (3x)	51'C	1.50V	40%	PASS
285MHz	190MHz (2/3)	2280MHz (8x)	1710MHz (3x)	51'C	1.50V	42.5%	PASS
290MHz	193MHz (2/3)	2320MHz (8x)	1740MHz (3x)	51'C	1.50V	45%	FAIL

At this point, the memory frequency is beginning to get closer to the 200MHz default value once again. The overall HTT is beginning to get a little on the high side, at 140MHz overclocked. Unfortunately, the 2x LDT multiplier would produce only 1160MHz HTT, which is still too low to retain stability. I am forced to stick with the 3x LDT multiplier at this point. 1740MHz is still well within acceptable limits, so my guess is that the CPU simply needs more voltage to scale any higher. With a load temperature of 51 degrees and a vcore of 1.50volts, it is getting pretty close to maximum acceptable limits. I'm going to push onwards a little harder to 1.55V, but I think this will be the last vcore increase before temperatures become unreasonable.

Before further increasing the clock speed, I want to further ease the stress to the memory controller. I did this by further reducing the memory divider to 100MHz (1/2) rather than 133MHz (2/3).

TIP: 1.55V is the maximum CPU voltage that the DFI boards support without using the 'VID Special Control' function. To exceed 1.55V, you have to use one of the 'Special' voltage boost percentages. 4% to 36% can be selected to allow the combination of almost hundreds of vcore values.

So after my best efforts, this was the best I could do with 1.55V, and the OEM HSF.

CPU Overclocking: Progress Chart

REF	MEM CLK	CPU CLK	HTT BUS	LOAD TEMP	VCORE	CPU GAIN	15 min P95/OCCT?
200MHz	133MHz (2/3)	1600MHz (8x)	1600MHz (4x)	44'C	DEFAULT	0%	PASS
215MHz	143MHz (2/3)	1720MHz (8x)	1720MHz (4x)	44'C	DEFAULT	7.5%	PASS
230MHz	153MHz (2/3)	1840MHz (8x)	1380MHz (3x)	45'C	DEFAULT	15%	PASS
240MHz	160MHz (2/3)	1920MHz (8x)	1440MHz (3x)	45'C	DEFAULT	20%	PASS
250MHz	167MHz (2/3)	2000MHz (8x)	1500MHz (3x)	45'C	DEFAULT	25%	PASS
260MHz	172MHz (2/3)	2080MHz (8x)	1560MHz (3x)	45'C	DEFAULT	30%	PASS
270MHz	180MHz (2/3)	2160MHz (8x)	1620MHz (3x)	46'C	DEFAULT	35%	PASS
280MHz	187MHz (2/3)	2240MHz (8x)	1680MHz (3x)	46'C	DEFAULT	40%	FAIL
280MHz	187MHz (2/3)	2240MHz (8x)	1680MHz (3x)	48'C	1.45V	40%	FAIL
280MHz	187MHz (2/3)	2240MHz (8x)	1680MHz (3x)	51'C	1.50V	40%	PASS
285MHz	190MHz (2/3)	2280MHz (8x)	1710MHz (3x)	51'C	1.50V	42.5%	PASS
290MHz	193MHz (2/3)	2320MHz (8x)	1740MHz (3x)	51'C	1.50V	45%	FAIL
290MHz	145MHz (1/2)	2320MHz (8x)	1740MHz (3x)	52'C	1.55V	45%	PASS
295MHz	148MHz (1/2)	2360MHz (8x)	1770MHz (3x)	53'C	1.55V	48%	PASS
300MHz	150MHz (1/2)	2400MHz (8x)	1800MHz (3x)	53'C	1.55V	50%	FAIL

It looks like that 2.4GHz mark is not quite obtainable when limited to 1.55V. With load temperatures approaching 53 degrees, that is about as far as I am prepared to go with this CPU cooler.

When I plot out the results graphically, you can clearly see the relationship between Vcore and Temperature. The two graphs are very similar. CPU core speed increases alone actually impacted temperature very little. It was the Vcore increases that had the most profound impact to temperature.

So at this point, we have a pretty good idea of what this CPU can do and what kind of voltage it expects to remain reasonably stable at specific clock speeds.

How did the 3500+ System do?

The process is nearly identical, so I won't waste your time displaying step-by-step procedures for the 939 platforms. Here is my overclocking chart for the 3500+ Winchester . I used 2-3-2-5 2T timings, and TCCD memory for the testing. Remember: The 939 Platform uses a default 5x HTT multiplier.

CPU Overclocking: Progress Chart

REF	MEM CLK	CPU CLK	HTT BUS	LOAD TEMP	VCORE	CPU GAIN	15 min P95/OCCT?
200MHz	133MHz (2/3)	2211MHz (11x)	2000MHz (5x)	39'C	DEFAULT	0%	PASS
210MHz	140MHz (2/3)	2310MHz (11x)	2100MHz (5x)	39'C	DEFAULT	4%	PASS
220MHz	147MHz (2/3)	2420MHz (11x)	1760MHz (4x)	39'C	DEFAULT	9%	PASS
225MHz	150MHz (2/3)	2475MHz (11x)	1800MHz (4x)	39'C	DEFAULT	12%	PASS
230MHz	153MHz (2/3)	2530MHz (11x)	1840MHz (4x)	39'C	DEFAULT	14%	PASS
235MHz	157MHz (2/3)	2585MHz (11x)	1880MHz (4x)	40'C	DEFAULT	17%	FAIL
235MHz	157MHz (2/3)	2585MHz (11x)	1880MHz (4x)	43'C	1.45V	17%	PASS
238MHz	159MHz (2/3)	2618MHz (11x)	1904MHz (4x)	44'C	1.45V	18%	PASS
241MHz	161MHz (2/3)	2651MHz (11x)	1928MHz (4x)	44'C	1.45V	20%	FAIL
241MHz	161MHz (2/3)	2651MHz (11x)	1928MHz (4x)	47'C	1.50V	20%	FAIL
241MHz	161MHz (2/3)	2651MHz (11x)	1928MHz (4x)	53'C	1.55V	20%	PASS
244MHz	163MHz (2/3)	2684MHz (11x)	1952MHz (4x)	53'C	1.55V	21%	FAIL
244MHz	163MHz (2/3)	2684MHz (11x)	1952MHz (4x)	56'C	1.60V	21%	PASS

So as you can see above, I may have been able to get further on this 3500+, but temperature was clearly a problem. Even with a Thermalright XP90, 1.6V is simply getting outside of the air-cooling realm. Judging by the vcore increase patterns, I'd venture a guess that 2.7GHz may be attainable, but at a very high vcore, of over 1.6V. Looking further back on the chart, the 2.6GHz mark looks very attainable on air-cooling and appears to require only 1.44V for this level of stability.

As with most A64 chips, this chip's limit was like a brick wall. In this case, 2.6GHz seemed to be the sweet spot before heaps of vcore was required for a meagre 84MHz increase.

Longer Term Stability Testing

As I mentioned earlier, 15 minutes of Prime95 is not a very good indication of stability. At this point of the process, we are not too concerned about long-term 24/7 stability testing simply because we need to put memory back into the picture. There is still a fair bit of work to be done. I will, however, select key points to test based on when vcore was increased. I want to do this, simply because if a result turns out to be stable, I can be reasonably confident that all results behind it will also be stable at the same vcore. For example, if 2.16GHz turns out to be stable for two hours with default vcore, I can be sure that all other slower results will also be stable with the same amount of vcore.

That being said, I will run a longer-term stability test (About 2 hours or so) on the following results:

CPU Overclocking: Longer-Term Stability Testing

REF	MEM CLK	CPU CLK	HTT BUS	LOAD TEMP	VCORE	2 Hours Prime95
295MHz	148MHz	2360MHz	1770MHz	53'C	1.55V	?
285MHz	190MHz	2280MHz	1710MHz	51'C	1.50V	?
270MHz	180MHz	2160MHz	1620MHz	46'C	DEFAULT	?

I started from the bottom of the list, and I was pleased to discover that 2.16GHz was stable for two hours at default vcore.

I can be assured that just about anything below 2.16GHz will also be stable at default vcore. I followed suit with the second result, and also had success completing two hours of Prime95 testing. Here is an updated chart:

CPU Overclocking: Longer-Term Stability Testing

REF	MEM CLK	CPU CLK	HTT BUS	LOAD TEMP	VCORE	2 Hours Prime95
295MHz	148MHz	2360MHz	1770MHz	53'C	1.55V	?
285MHz	190MHz	2280MHz	1710MHz	51'C	1.50V	PASS
270MHz	180MHz	2160MHz	1620MHz	46'C	DEFAULT	PASS

Our highest successful overclock, which passed 15 minutes of Prime95 failed after about 30 minutes in this longer test.

Since I am already in Windows, I will take advantage of 'Clockgen' to reduce the reference clock by about 3MHz (See the 'Tools' section for more information).

After reducing the clock by about 3MHz, Prime95 had no problem crunching away for about two hours.

So for our Sempron 2600+, our updated chart looks like this:

CPU Overclocking: Longer-Term Stability Testing

REF	MEM CLK	CPU CLK	HTT BUS	LOAD TEMP	VCORE	2 Hours Prime95
295MHz	148MHz	2360MHz	1770MHz	53'C	1.55V	FAIL
292MHz	146MHz	2336MHz	1752MHz	53'C	1.55V	PASS
285MHz	190MHz	2280MHz	1710MHz	51'C	1.50V	PASS
270MHz	180MHz	2160MHz	1620MHz	46'C	DEFAULT	PASS

So what about our 3500+ system? Surprisingly, all of our tests passed without issue.

CPU Overclocking: Longer-Term Stability Testing

REF	MEM CLK	CPU CLK	HTT BUS	LOAD TEMP	VCORE	2 Hours Prime95
244MHz	163MHz	2684MHz	1952MHz	56'C	1.60V	PASS
241MHz	161MHz	2651MHz	1928MHz	53'C	1.55V	PASS
238MHz	159MHz	2618MHz	1904MHz	44'C	1.45V	PASS
230MHz	153MHz	2530MHz	1840MHz	39'C	DEFAULT	PASS

So now we have a better idea of system stability, and what our CPU is capable of with memory out of the picture. Only the highest 295MHz reference clock on the Sempron 2600+ proved unstable. Although two hours of Prime95 provides us with some reassurance, it is still not enough to consider this overclock stable. We'll be doing a 24-hour test once we finish. For now, let's move on to the next section!

Next section: Memory Clock and Timings

Previous section: Overclocking Tools