You are here

Ultimate Underclock & Undervolt Project

July 17, 2002 by Leo Velikovich

In this article I share my experiences building a near-silent PC with an Athlon XP and GeForce4 Ti4200. The entire system including power supply is cooled by two under-volted 80mm Panaflo fans, which combined produce around 18 dBA No exotic or expensive cooling methods are needed. My goal is to show you that quietness CAN be accomplished even with hot-running, high-performance CPU's or graphics cards. All it takes is determination and a little ingenuity. (Editor's note: A key aspect of Leo's project is the successful modification of a GeForce4 Ti4200 video card with underclocking, undervolting and heatsink replacement for fanless cooling.)

GeForce 4 Ti4200: Now cool and silent

Why did I do this? The cost of noise.

Before we begin, I have a confession to make: I was once on the Dark Side, an overclocking speed addict. I would run a Swiftech Air Peltier just to eke out an extra 10%, and its horrific hair-drier fan noise didn't hold me back. But one fine day, God condensed water on my CPU. I was forced to replace the Peltier with a normal heatsink/fan, and funny how my hearing suddenly improved! Just because I was "used" to the noise didn't mean it wasn't costing me. My costs included a less comfortable environment, difficulty of concentration, and reduced sound quality. I am convinced that many PC users who tolerate noise would be much happier without it. Imagine reading a book with a whiny fan strapped to it; that's the sad present state of computing.

What is underclocking and why do it?

Many PC enthusiasts harbor a vague hostility towards the idea of underclocking (Who in their right mind would SLOW DOWN a fast processor?!) It's a prejudice, which I fight with information:

Yes, underclocking means you run your CPU or video card at a slower frequency. The goal of underclocking is to reduce heat output, resulting in less need for noisy fans. But the beauty of underclocking is that only a TINY reduction in actual performance gives you a LARGE reduction in noise. There are many reasons for this:

1) Massive heat reduction from lower voltage. Any processor requires a certain core voltage; for most Athlons it is 1.75V, for newer Pentium 4s it is 1.5V, and for a GeForce4 chip it's around 1.65V. If not given sufficient voltage, the processor may become unstable at its default MHz. However, if you underclock you'll be fine at a lower voltage. The reason this helps is that total heat output (in Watts) is proportional to voltage SQUARED. This means if you halve the voltage, heat output should be cut by a factor of 4! Simply because I run my 1.75V Athlon at 1.6V means it's only producing
(1.6/1.75)squared = 84% of the original heat.

2) Heat reduction from lower speed. Lower MHz automatically means lower wattage. For example, the approximate formula for an Athlon XP/Palomino's maximum wattage is: 20 +
.03 x MHz. Try it yourself: a 1600 MHz model produces 68 watts max, an identical model at 1400 MHz only produces 62 watts max per AMD's specs. By underclocking my 1466 MHz Athlon XP to 1400 MHz, I'm decreasing heat production from 64 to 62 watts = 3% less heat (and it adds up). (Editor's note: The relationship between CPU clock speed and power dissipation is linear - Leo's underclocking-related CPU heat drop is actually closer to 4.5%.)

3) Higher temperature tolerance at lower speed/voltage. Consider the Intel Coppermine Pentium III: as they pushed the P3 all the way to 1 GHz its temperature tolerance fell drastically. The 600 MHz version should endure around 80 Celsius, the 1 GHz could fail at 60° C (Intel spec). This is why overclocked processors need so much extra cooling. Underclocked processors may work fine even past their ceiling temperature. This means you have a greater safety margin. At 57-67° C temps, my AthlonXP is well below its 90° C spec, and further below its true limit @ 1400 MHz.

4) Better clock-for-clock performance at lower speed. This is the law of diminishing returns, only working in your favor. The processor may run slower, but the memory and bus still run at the same speed! So the real performance loss isn't as great as the loss in frequency. In reality, Athlon performance loss equals roughly 80% of the frequency loss (average calculated from Anandtech benchmarks). My AthlonXP is running 4.5% slower, but I should only be losing about 3.6% in performance. This phenomenon is more profound in memory bandwidth-limited video cards; frame rate losses might be less than half the clock speed losses.

Now, for a concrete illustration of all this:

EXAMPLE #1: My Athlon XP 1700+ allowed the following possibilities:

Athlon XP 1700+
Default
MY SETUP
Further
Extreme
Clock speed
1466 MHz
1400 MHz
1266 MHz
1000 MHz
Core voltage
1.75V
1.60V
1.45V
1.15V
Maximum heat(*)
64.0W
51.8W
39.8W
21.6W
% performance(**)
100%
96%
89%
74%
% original heat
100%
81%
62%
34%

(*) Editor's Note: Leo arrived at the wattage figures using calculations based on his comments in 1) and 2) above:

  • voltage_ratio = (1.60V/1.75V)squared = 0.836
  • wattage_at_default_voltage = (20 + 0.03*1400 MHz) = 62W
  • actual_wattage = 0.836 x 62W = 51.8 Watts

They correspond closely to results obtained via the utility Radiate, which has the AMD MP in its database. MP is identical to XP in its electrical characteristics.

(**) Using the formula: Performance loss = 0.8 x Clock speed loss (average from Anandtech benchmarks for AthlonXP)

At speeds within 1 GHz, the "hot" Athlon XP could run fanless altogether! And with respectable performance, likely far better than VIA's low-power C3. How hot your Athlon runs is completely up to you. (Editor's note: Well, maybe not quite fanless, but close, with a massive powerful heatsink. The C3s run 10W max, typically way less, and even they can't be run fanless in all cases. )

EXAMPLE #2: My GeForce4 Ti4200, which I modified for lower voltage and fanless operation -- I'll explain later. I don't know its true wattage, but here are the remaining numbers:

GeForce4 Ti4200
Default
MY SETUP
Clock speed
250 MHz
225 MHz
Core voltage
1.64V
1.40V
% performance(*)
100%
93%
% original heat
100%
67%

(*)Using Direct3D 8.1 SDK GPU-stressing benchmarks at 800x600x32 resolution with Quincunx FSAA.

Is underclocking starting to make sense? At a price of 4-7% performance I gained 20-33% cooler components; in combination with highly efficient cooling, this permitted the near-silent system I now have.

On to the mod details...

Contributor Leo Velikovich is an undergraduate computer science student at the University of Maryland, College Park.

System components

  • Enlight 7237 case
  • QuietPC 300W PSU
  • Asus A7V133 mainboard
  • AMD Athlon XP 1700+
  • 384MB PC133 SDRAM
  • Gainward GeForce4 Ti4200 64MB, AGP
  • 3COM 10/100 NIC, PCI
  • AverTV Studio TV tuner, PCI
  • Midiman Audiophile2496 soundcard, PCI
  • IBM 75GXP 7200rpm 30GB HDD
  • Plextor 12/10/32A burner

Fanless northbridge

Some things just don't need fans, but companies attach one anyway (to woo overclockers?) My Asus A7V133 motherboard came with an obscenely noisy fan atop the VIA KT133 Northbridge chip (4W chip power consumption). This was easily solved. I removed the fan, then detached the heatsink. I lapped the heatsink until it was flat and shiny, and lapped the Northbridge as well. Then I put some Arctic Silver II onto the Northbridge and reattached the heatsink. When I checked the temperatures I laughed: VIA's spec allows up to 110° C operating temps, above boiling point! I recall my sensor readings were in the 60's!

Under-volted Panaflo fans mod

Yes, fans get under-volted as well. This makes them spin slower, producing less noise. Guides on this are available on many websites. I am using two 80mm Panaflo L1A low-speed 1900RPM fans at 7V instead of default 12V. I did not choose 5V because the fans seemed to generate much less airflow than @7V. Even at 7V, the two fans are practically inaudible inside the case.

Athlon XP cooling mod

Alpha PEP66H heatsink with Panaflo blowing up towards PSU

When using quiet or passive cooling, it is crucial to maximize heat transfer by using high-quality heatsinks and thermal compound. I used a "new" Alpha PEP66H heatsink -- it has a copper base for drawing heat away from the CPU and aluminum fins to dissipate it into the air. I lapped the CPU and heatsink surfaces to shining point; then used a careful application of Arctic Silver II between them.

Then I used good ol' tape and attached one of the Panaflos directly underneath the Alpha, blowing upwards. The odd choice of position is no accident: the fan's purpose is to move all hot air upwards towards the PSU, where it can be sucked out by the other Panaflo acting as exhaust fan. This is the other golden rule of quiet cooling: fans must be placed strategically for maximum usefulness.

I unlocked the AthlonXP 1700+ by connecting L1 bridge contacts with conductive ink. Unlocking the CPU is crucial here, since you only want to underclock the CPU, not the FSB. Once I decided to go with 1.4GHz, I had to find the lowest voltage (read: lowest heat) at which the CPU remained 100% stable.

Power supply mod

Photo shows coverless PSY and both Panaflo fans

The PSU's maximum safe temperature is around 60° C, much lower than a CPU or video card. I started with a QuietPC 300-watt AMD-approved PSU; it has heatsinks for better heat dissipation. I removed the cover and replaced the default fan with my other 80mm Panaflo fan @7V. The cover would only obstruct airflow, and I needed this fan to remove exhaust air from the system as efficiently as possible, so the cover stayed off.

Shroud/hood

Hood geometry

To prevent a pocket of hot air from being trapped inside my case, I constructed a hood from thick poster carton. It forces air to flow from CPU heatsink to PSU and out of the case. I must note that it doesn't help much: CPU/motherboard temperatures went down 0.0-1.0° C when using it. However, I believe it keeps the front of the case cooler, i.e. the hard drive.

Hard drive

I am using an IBM 75GXP two-platter 30GB model inside a QuietPC SilentDrive enclosure; spinning noise is inaudible from a couple feet away. I also turned IBM's acoustic management to 75% quiet operation, at which point seeks are inaudible as well.

Fanless GeForce4: Fanless Mod

Here comes the fun part. The GeForce4 Titanium is NVIDIA's fastest videocard yet -- and unfortunately the hottest-running, which is why all of them ship with whiny high-speed fans. I got the Gainward Ti4200 64MB, because 4200 is the cheapest and coolest-running GeForce4 GPU. I also bought a "new" Alpha PAL6035 copper/aluminum heatsink along with Arctic Silver thermal adhesive.

I removed the stock HSF. Then I used sandpaper on a flat piece of glass to thoroughly lap the Alpha surface. Once it was perfectly flat, I put the sandpaper on the back of the Alpha and used it as a sanding block to lap the GPU. Once both mating surfaces were flawless, I proceeded to glue them with a thin layer of Arctic epoxy.

Fanless GeForce4: Voltage Mod

Unfortunately, the above mod was insufficient: my GeForce4 ran stable, but it reached a scalding 84° Celsius under load! Then I found an excellent article by Xbitlabs on GeForce3/4 voltage modification. While they talked about increasing voltage for better overclocking, I needed just the opposite -- to lower the voltage.

I noticed that my Gainward Ti4200 uses SC1102 voltage regulator chips (found on the backside of the board); the GPU regulator is the one near the metal mounting bracket. Xbitlabs posts its schematic here.

The voltage passed to the GPU equals 1.265 x (1 + R8 / R7) volts, so Xbitlabs added another resistor to R7 in parallel to reduce total resistance, thus increasing the R8/R7 ratio and increasing voltage. I needed to do the opposite, so I had to find R8 and add another resistor in parallel to it.

First I realized (from the diagram) that Pin 11 was directly connected to one side of R8; I just had to find where its other side was. I bought myself a multimeter and used it to measure the resistors on the board near the SC1102 chip. Ultimately, I found R8. See the pictures for the SC1102 chip and R8's two contacts A and B. I measured, by connecting to SC1102's Ground pin and point B on R8, that the default GPU voltage is 1.64V.

So here is the bottom line: to undervolt the GeForce4 you solder a resistor between points A and B in the diagram. You can use resistors from 22 to 100 Ohms for practical results. Here are the results I measured for those resistors:

Ohms Vcore Comments
0 1.64V -
100 1.43V -
47 1.40V, R8 removed my current setup
68 1.39V -
47 1.36V could still work at 250 MHz; but, occasional T&L errors
22 1.32V -

I'm rather new to soldering, and I eventually made a mess of my GeForce4 after trying several external resistors. I eventually broke off the original onboard R8 resistor and used simply my own 47-Ohm resistor instead of it. So now I'm at 1.40Vcore. Frankly, I'm amazed my videocard still works after all my clumsy experimentation :) (Editor's note: Those tiny surface mounted resistors are challenging even with a micro-tip soldering iron.)

Fanless GeForce4: BIOS Mod

Since I'm underclocking, I made my card believe 225 MHz (not 250 MHz) is its true speed. This eliminates the nuisance of using overclocking programs or CoolBits. I used the NVIDIA BIOS Editor and NVFLASH to extract and modify my card's BIOS. While at it, I added a cool boot message to signify my victory over heat.

Temperatures, Stability and Conclusions

Here are the results, with case cover on.

Stability is obviously not a problem, because I picked speed/voltage combinations that are 100% stable and error-free.

CPU temps were measured by the Asus socket thermal sensor. Please note that Asus readings are much higher (10-15° C higher) than other manufacturers' socket readings. This way, the readings better approximate true CPU temperatures. My CPU temp readings range from 57° C (idle) to 67° C (heavy load). The max allowable by AMD is 90° C, and I'm not anywhere close to that. Also, if I used VCool, my idle temps would go down but the worst-case temps wouldn't be affected.

GeForce4 temps were measured by a thermal sensor stuck between the heatsink and card, touching the GPU edgewise. The temps range from 64° (idle) to 68° C (heavy load). NVIDIA doesn't disclose the max allowable, but Anandtech's comparison of similar GeForce4 Ti4200 cards reported temperatures (under load) from 56° to 70° C -- using default speed and fan cooling! In other words, the fanless GeForce4 is entirely within spec, just 6-7% slower.

All of this proves that you don't have to give up high performance for quietness. Much better heatsinks, fans, and thermal compounds are available today than a few years back. Manufacturers are finally starting to realize that quietness matters; I purchased an AOpen H300A Slim Case for my server and the power supply uses a thermally-adjusted 80mm Yate Loon fan which runs at around 5V! Furthermore, this fan is positioned to suck air off the CPU heatsink, so the CPU gets cooled as well; I won't even need any modding.

The future is looking promising, but no need to wait: you can do any of the mods in this article right now.

* * * * *

Discuss this article in our Forums.

POSTSCRIPT July 28, 2002

After completing this article, I managed to destroy my GeForce4 accidentally while using the multimeter to measure GPU voltage. I shorted two contacts with the multimeter's pins while the card was powered up. Please note -- this was NOT caused by the modification itself, which proved perfectly safe/stable for something like 100 hours of 3D apps/games. However, if you are doing a similar project, my advice is not to take ANY electrical measurements on the card while it's running inside the PC. Lower temperatures will SAFELY reveal whether you succeeded lowering the voltage. The videocard died nobly for the cause of science - a moment of silence, please.

Google

www SPCR