CPU Undervolting & Underclocking: A Primer


August 17, 2002 by Mike Chin

Two terms that often come up in Silent PC Review are undervolting and underclocking. These terms refer to setting the CPU core voltage and clock speed under the default settings in order to lower its heat output and make it easier to cool with a quiet, low speed fan. Computing performance does take a hit, but CPU temperature can be lowered dramatically, making it much easier to obtain quiet operation.

Just a joke.Most of you are familiar with overclocking, which refers to running the CPU over its normal rated clock speed to extract the maximum possible performance. In conjunction with overclocking, the CPU core voltage (Vcore) is sometimes raised above default to stabilize a highly overclocked CPU. Both have the effect of raising temperature not only in the CPU, but as a result of the increased temperature of the air around it, the entire case. This is the exact opposite of what is desirable for a quiet or silent computer.

Clock speeds and core voltages are easy to set with modern motherboards. Recently released (~2 years) motherboards feature software control of hardware settings in the BIOS. These settings include CPU clock speed, system front side bus (FSB) speed, independent PCI & memory bus clock speeds, and Vcore adjustments. Separate bus speeds for FSB, memory and CPU allows all components other than the CPU to remain at or close to default speeds. (Many PCI peripherals are not happy with more than 5~10% deviation from the default settings.) Naturally, the degree of BIOS control varies from model to model and chipset to chipset. The most recent ones aimed at enthusiasts are the best; look for a motherboard that has a wide range of BIOS-accessed hardware settings.

CMOS Clear jumperOlder motherboards generally do not offer as much control, and some settings can be set only by physically moving jumpers on pins. With modern motherboards, the only one you absolutely need is the CMOS clear jumper. This is how you reset all the BIOS settings to their defaults when you lock up the system with settings outside the operational range of your particular set of components (CPU, motherboard, memory). Fortunately, this throwback to ancient times remains featured on every motherboard. :)

Bypass the Multiplier Lock if Possible

Almost all desktop CPUs are clock-locked. Intel started the practice soon after the first generation of Pentiums to stop overclockers and have continued since. AMD started the practice with the Athlon. The lock on the processors is specific to the multiplier. Still, even with a locked multiplier, CPU speed can be changed with the front side bus (FSB) on the motherboard. With Intel processors, adjusting FSB is the only way to change the CPU speed.

Easily unlocked AMD T-birdWith AMD Thunderbird, XP and MP, it is possible to close a set of tiny terminals etched on the CPU casing to unlock the multiplier. This is well-documented on many overclocking web sites; a google search will bring up a handful. Once unlocked, many combinations of bus speed and multiplier can be tried with the CPU. It almost guarantees good over- or under-clocking success. For T-birds, multiplier unlocking can be achieved with an ordinary sharpened pencil. With XPs, AMD made it a harder. The trick requires techniques with materials more permanent than pencil carbon, and a small slip can render the processor useless. (An expensive trick!) Still lots of people to do it successfully. Leo, for instance.

Unlocking the multiplier is also useful for underclockers to raise the bus clock speed and lower the multiplier. The performance penalty of lower overall clock speed can be offset by higher FSB speed. For silent aficionados, the prime attraction of beating the multiplier lock is access to the broadest range of possible clock speeds.

Calculating CPU Power

The utility Radiate is useful and educational to play with. It's kind of a packaged graphic interface for a handful of formulas that calculate the relationships between CPU clock speed, Vcore, and power consumption. (It also predicts CPU temperature with any given ambient and correct C/W of the heatsink used, but this is not relevant here.) The utility is accurate; its calculations were confirmed against CPU spec sheets.

Radiate interface

Extreme Underclocking / Undervolting

Here is an extreme undervolting and underclocking example with a 1 GHz T-bird unlocked over a year ago. It accepts a wider range of underclocked (and overclocked) speeds now than when it was new.** (See note at end of article)

T-Bird Test System

CPU AMD T-Bird, 1 GHz, pencil unlocked
HSF Swiftech MC462 + Panaflo 80mm L @ 7V
Motherboard ABIT KT-7A(RAID) - KT133A chipset
Video Radeon VE 32MB
PCI 1 10/100 network card
PCI 2 Soundblaster Live!
PSU Seasonic 300W
HDD Seagate 40G Barracuda IV
OS Windows 98SE
Room Ambient 25° C
AC Power 95W idle; 110W max. Measured with Kill-A-Watt.

The ABIT KT-7A(R) motherboard BIOS provides complete software control for all hardware settings. Hitting the Delete key at the beginning of the boot process as the screen is just starting to display accesses the BIOS menu . The CPU adjustment menu item is SoftMenu III. This menu screen provides options for CPU bus and multiplier speeds, CPU voltage and fine details of memory performance. It is shown below with all CPU setting at default for the 1 GHz T-bird: 10x multiplier, 100 MHz FSB with 33 MHz PCI clock, and Vcore at 1.75V.

default BIOS screentshot

Default 1 GHz T-bird settings

At default settings, the 1 GHz T-bird is rated to draw 49W typically; 54W max. Since the unit is unlocked, its clock speed is completely controlled by the motherboard. Even at the stock 1 GHz clock speed, the CPU runs fine with ~.25V less than the default voltage. But what's the coolest it can be run at?

The minimum speed allowed in the BIOS is 5X multiplier with 100 MHz FSB. This is 500 MHz, half the T-Bird's normal speed. Let's see if it boots with the minimum Vcore of 1.1V. The handy utility Radiate calculates power dissipation will be 10.2W. Cool enough to run with no fan!

Tried 1.1V, 1.2V, 1.3V, 1.4V, and finally, after a last attempt at 1.5V, tossed in the towel. This clock speed appears to be below the operating limit of the system, for whatever reasons.

OK, then how about 6X with 100 MHz FSB for 600 MHz? This is still a 40% underclock. Start with Vcore set to 1.1V.

BIOS set to 6 x 100

No problem! The system boots on the first attempt and is stable. Motherboard Monitor 5 indicates the measured Vcore to be 1.17V. What does Radiate calculate? Just 13.8W! The table below details other significant differences. Both AC power and temperature measurements dropped dramatically.

Before and After Underclocking

CPU Speed
1 GHz (10 x 100 MHz)
600 MHz (6 x 100 MHz)
Monitored Vcore
Calculated CPU power
CPU temp, idle*
48° C
36° C
CPU temp, max*
62° C
38° C
Measured AC power, idle
Measured AC power, max

* As monitored by in-socket thermistor on motherboard. Usually read 10~13° C lower than the temperature in the CPU core, which has a maximum safe temp of 90° C.

The system is using at least a third less power than before. That's about 30W less heat! The cooling headroom is great enough that the system is safe with no fan on the heatsink (at least with the powerful Swiftech MC462A heatsink and the system placed open on the testbench): during ordinary computing for an hour without a fan on the heatsink, CPU temp never exceed 52° C

What is the performance cost? Some benchmarks from SiSoft Sandra (V. 2002.1.8.59) are shown. All BIOS settings were left at defaults at both clock speeds. Tweaking memory timing and a few other BIOS settings would likely improve the scores at least 10%.

SiSoft Sandra Performance Benchmarks

1 GHz
600 MHz
Math Dhrystone ALU
2750 MIPS
1649 MIPS
Math Whetstone FPU
1372 MIPS
822 MIPS
Multimedia MMX/SSE
5474 it/s
3280 it/s
Floating Point 3DNow!
6780 it/s
4062 it/s
Mem Int. Bandwidth
715 MB/s
709 MB/s
RAM Float Bandwidth
712 MB/s
694 MB/s

The performance hit is 40% across the board, exactly the same as the change in CPU clock speed, except the the memory bandwidth, which is affected by the bus speed, not the CPU clock, per se. How do these numbers translate in terms of what I like to call the computing experience? In other words, how much impact did it have on my perception, productivity and enjoyment while working with the system?

The answer depends on what task was being done. With most business and work related applications, web browsing and e-mail, there was little change. The exception was Photoshop and some electronic publishing software, in which some operations took noticeably longer to complete. Multitasking, in general, was also somewhat affected. Multimedia did not strike me as anywhere close to 40% slower. With a few games, the system felt more sluggish at 600 MHz, but again, I could live with it. A faster video card would probably be a big equalizer. Overall, I could live with the system either at default or underclocked. Naturally, I prefer the higher speed.

Other Clock Speed Options

But it's not an either/or choice. Many CPU speeds between 600 MHz and 1.0 GHz can be reached, with a range of multiplier and bus combinations.

Various Undervolted & Underclocked Settings Achieved

Clock Setting

Vcore, measured

Math ALU
Mem. Int. Bandwidth
CPU Idle Temp
CPU Power*
600 (6x100)
1.17 V
1649 MIPS
709 MB/s
13.8 W
700 (7x100)
1.26 V
1924 MIPS
710 MB/s
18.7 W
785 (6.5x130)
1.42 V
2148 MIPS
921 MB/s
25.8 W
800 (6x133)
1.48 V
2211 MIPS
944 MB/s
29.4 W
933 (7x133)
1.48 V
2568 MIPS
944 MB/s
34.3 W
1000 (7.5x133)
1.53 V
2750 MIPS
944 MB/s
38.8 W
1000 (10x100)
1.79 V
2750 MIPS
715 MB/s
49 W

*CPU power calculated by Radiate software.

Several noteworthy points:

  • Even at the stock 1 GHz clock speed, the CPU runs with ~.25V less than the default voltage, which means it draws 35W instead of 49W.
  • Lower voltages can be reached with lower clock speeds.
  • Voltage has the greatest effect on CPU heat.
  • Higher bus speed result in better memory bandwidth, even at lower CPU clock speeds. For many applications, this translates directly to improved performance, even with lower CPU clock speed
  • These results cannot be achieved with a multiplier locked CPU because different multiplier setting are required.

Practical Advice

All of these suggest that to achieve lower temperatures with the minimum performance hit, the best approach is to:

  • Start by setting the CPU voltage to the lowest point at which the CPU will run stable without lowering clock speed.
  • If the drop in temperature achieved with the first step is not enough, then adjust the clock speed down, and try lower the Vcore further.
  • When underclocking, try to maintain the highest FSB for minimum performance degradation.
  • To be safe, ensure that the PCI and memory bus speed is not changed by more than about +/- 10%. Some would say 5%.
  • Use the drop in CPU temperature to lower the HS fan speed or replace a noisy, fast fan for a slower, quieter one.

The last point really is the crux of undervolting and underclocking: reduce CPU heat to lower cooling requirements, allowing the use of quieter fans. CPU temperature at maximum needs to be in a safe zone; it really doesn't matter how this is achieved. By reducing the heat output of the CPU, you have that much more headroom to use quieter fans. Generally, the airflow required for effective case cooling also drops, which means the case fans can be made to run more slowly, replaced with slower quieter ones, or removed altogether.

A few more points in closing:

  • Benchmark stress testing is artificial. You're better off to stress the system by using it hard and long the way you (or the user, if you;re building for someone else) would. The temperatures reached are more relevant for your system optimization. When tweaking, it is very difficult to damage your CPU this way because if the system is unstable, it will let you know by locking up and crashing long before anything can burn. In contrast, most people don't want to hang around in front of their PC while it cycles through hours of benchmarking, so the opportunity for damage is higher.
  • Use manufacturers' data or Processor Electrical Specifications about maximum safe temperatures to establish your targets, but be very aware that many CPU temperature monitoring systems are woefully inadequate. This is especially true of socket-A motherboards for AMD processors. Even the latest ones which are supposed to provide monitoring of the XP and MP embedded thermal diode appear to have issues. To be safe, 65~70° C is the maximum temp you should let your AMD socket-A CPU reach, using the built-in motherboard monitoring utilities. Current AMD processors are rated to be safe to 90° C in the core, but you don't want to get even close. Intel Celeron, P3, P4 and VIA C3 processors have embedded thermal diodes and most motherboards read them quite accurately.
  • The quietest systems usually run hotter than usual, but within safe limits. The natural ambient temperature determines the minimum noise level achievable with any system. A quiet PC that is adequately cooled in Achorage will probably not be adequately cooled in Bangkok; minimum airflow requirements will likely be higher in Bagkok, and the increased air turbulence will raise the noise level. This is as true with air-cooled systems as with water-cooled ones. The term "water-cooling", by the way, is a bit misleading, because the final cooling in a water-cooled system is not achieved by water but by forced air: a fan. Water is an intermediary that transfers the heat away from the CPU to a radiator which is then cooled by a fan. Water cooling does provide more effective cooling and lower temperatures, but NOT lower noise: not only does the fan noise remain, there is also the noise from the water pump. (Is that a challenge for water cooling enthusiasts? Hehe.) This brings us to the issue of air versus water cooling, which really belong in another article.

* * * * *

** A processor that has been used for a while, especially one has been used hard (read: overclocked and/or used for lots of demanding games), may be more accepting of non-standard speeds than one that is new. This applies to operation at both over and under normal settings. This is my experience with most processors I've used. It may be a side effect of electromigration at an early stage. All this is little more than conjecture.

Discuss this article in our Forums.

Help support this site, buy from one of our affiliate retailers!