How To Put Together a 172W Peltier

Add Your Comments

How did this start?

At the beginning, I just to make my homebrew PC as fast as possible without breaking the bank. Athlon seemed to be a logical selection to do that.

Then I realized the beast is at melting point all the time, so I ended up with an Alpha 8045 to cut the temps to more reasonable levels, but that and a few other 80 mm screamers to the case made the whole thing hard to live with.

Then I saw Petar’s article
on a homemade waterblock, and I couldn’t resist trying water-cooling. That was the real breakthrough; finally my PC was decently silent with Athlon breathing at 35°C; an XP1700+ jumped from 1700 to 1760 MHz.

Soltek Waterblock

Finally, overclocking craze came over me. Now I wanted to reach a magical 2 GHz, or as would AMD say, 2500+ !

You don’t need an M.Sc. degree to realize that keeping a CPU to room temperature isn’t enough to go sky-high with MHz. You need to use active means to get it well below room temps.

Unfortunately, there are only two types available.

One is evaporative phase change “refrigerator type” such as Vapochill or Kryotech, which is too complex and bulky for my taste. The other, much more convenient alternative for home PC builders is Peltier.

Simple calculations showed that reaching 2GHz+ would squeeze more than 100W out of Athlon, so to be on the safe side, one should use at a least 200W Peltier. 220W units had just been announced at the time, still were not available, and there was nothing else more powerful available than 172W model made by TE Distributing.

Since I already had the rest of the hardware together, I decided to give it a try.
I won’t elaborate here the theory and principles of both Peltier and water-cooling, as all of that can be found on this and other OC sites. My intention is to present solutions and techniques used in this project that differ from others shown elsewhere.

Here is what I had:


Athlon XP 2100+ AROIA


Epox 8K3A+


2 x Samsung 512 Mb 333 MHz

Video card

MSI Ti4400


JNC big tower

Power supply

Maxpower 400W

Hard disks

Maxtor ATA133 80 Gb, IBM ATA100 45 Gb



And this is what I did with them . . . .

Miljenko Devcic

The Athlon was unlocked, with L10 “switched over” to lower multipliers. As this switching requires very steady hand and steel nerves, for the faint-hearted, I would suggest using XP2000+ or lower instead (they are wired for lower set of multipliers). The processor’s underside was covered with silicone compound to prevent condensation.

Athlon Bottom

The first motherboard change I made was to locate the processor thermosensing diode pins and a pair of thin wires was soldered to them. This connection allows me to add circuitry for exact core temperature measuring in the future.

Then I began insulating the motherboard.

The backside of the motherboard around the processor region was covered with silicone, while the processor socket was filled with vaseline. A 5mm Depron plate was inserted into the central socket cavity, with a thermistor hole made prior to insertion. This hole was filled with white thermal silicone paste in order to get good thermal contact with the processor.

Siliconed Socket

Two other measuring mods have being made as well.

The “system” thermistor was unsoldered from the motherboard and replaced with pair of connectors and 40 centimeters of wires ending with the same thermistor. This makes a very useful portable probe for measuring temps all over the PC.


Instead of original passive Northbridge cooler, a 486 fan-equipped cooler was used.

Around the processor socket, cold plate and Peltier, thermal insulation was applied. These consisted of various Depron and styro plate thicknesses selected according to element thickness. All sheets have 80 x 80 mm outer dimensions, while cutouts were made to fit.

For instance, the processor socket is surrounded by one 5 mm sheet, the CPU by one 2 mm styro sheet, the cold plate by two 5 mm sheets, etc. Depron and thin styro sheets are sold as thermo insulation under wallpaper.


The case is, understandingly, the most modified component in this project. JNC case was chosen because it has ample room to accommodate an additional power supply at the back, a big fan at the front and a two-liter water tank in the upper chamber. Two 80 mm fans could be also be fitted on the upper back.

Some cutting was necessary to fit the Peltier power supply and front fan.. The original power supply case cutout pattern was transferred vertically up 90 mm and cut out.


The lower front baffle was cut to accept a 150 mm major fan produced by Comair Rotron.

Big Fan

The original power supply that came with the case was inadequate, so the Maxpower 400W PS was used instead. However, I needed to make a couple changes to it.

5V rail voltage on the Manpower measured incorrect values, ranging from 4.7 to 4.9V. Divider resistors within PS were detected and trimmed for correct voltage.

Another modification made to the Manpower was a relay inserted between mains input and mains output, powered by either 12V or 5V rail, depending on relay rating. Its purpose is to power up the Peltier power supply (connected with short mains male/ female cable) and water pump at the same time as the rest of the system.

One more power supply to go . . . .


Miljenko Devcic

I needed another power supply to make the cooling cool. According to FerroTec application notes.
, a Peltier doesn’t require stabilized voltage, so I decided to design and build my own “classic” transformer-and-rectifier: 300W toroidal mains transformer with 22V output (value found by experiment) was used, connected to 25A diode bridge and 33000uF capacitor. All components could fit into empty standard ATX power supply casing that was mounted above standard case PS.

The heart of water cooling is the waterblock. My waterblock is designed according to Petar Lazarevski’s article (which I thoroughly recommend studying), with a few modifications to adjust it to my needs and somewhat simplify the construction.


Since the total thermal dissipation expected was much higher with a Peltier than without, the waterblock capacity had to be maximized. All the “flute” diameters were enlarged; input and output ones to 12 mm and four interconnecting to 8 mm..

Therefore, the total block thickness was changed to 16 mm, and the length was enlarged to 80 mm to use 4 motherboard cooler mounting holes.

On the lower waterblock surface, four threaded holes were made for mounting Peltier and cold plate.

Connecting pipes were the biggest mod. To provide much more effective area (and happily, much less work), I used copper central heating pipes.

In Europe, 6 meter pipes are cut to length and hard soldered to various pre-bended pieces. One side is broadened to accept standard diameters of 12mm, 15 mm and wider.

In this project, both 12 and 15 mm dia(meter) pieces were used in different bending angles and curvatures. They were soft soldered using gas torch and special soldering paste containing zinc. A gas torch doesn’t work for soldering pipes to waterblock, so instead, the waterblock was put onto an electric cooker and when soldering paste melted, pipe assemblies were inserted into input and output holes.

For the waterblock, 12 mm 45° and 90° bended pipes were used. For the water tank and air/water heat exchanger (Toyota Supra heater), I used the 15 mm variety.

The trick is in getting different connecting style between copper pipes and plastic interconnecting tubes together.


They have all 12 mm internal and 16 mm external diameter. While internal dia fits nicely onto 12 mm pipes, external dia fits snugly into widened 15 mm pipe end (“female” side made for inserting 15 mm pipe).

When connected, both plastic and copper tubes are squeezed with 40 mm length of thermo shrinking sleeve. This makes perfectly leak-proof and secure fitting.


Water tank is a 1.8-liter food container with tight fitting lid. I couldn’t get Eheim 1048 water pump (probably the best solution), so I used an Eden 114 submerged aquarium pump declared for 600 l/h. The total system flow was measured and very good value of 6.1 l/min was achieved.

Generally, submerged water pumps are not that good for this purpose as their working temperature exceeds allowed maximum, so the service life could be compromised.

When soldering new copper in/out pipes to the radiator (original ones rarely suit angles and diameters needed), a tiny brass tube on the bottom was added to allow for emptying the whole system for servicing. 30 cm of silicon tube with end plug was added to this pipe.

When I fitted the front fan and radiator, I inserted a tailored Depron ring to avoid air leaks. It is made of two 5 mm sheets with different cutouts (circular and square).


Some Airflow Math

Besides high performance, water-cooling is much less noisy since there are high-speed fans revving.

However, due to many high dissipation devices inside the case (water “radiator” being the main heat source) decent air flow must still be provided.

To maintain low noise levels, I did a few simple math equations dealing with fan flow and noise. They showed that cutting the fan speed in half lowers the flow in half as well, but cuts noise by 15dB!

This fact lead to fan configuration as shown: four 42 cfm (71 cubic meters per hour for us with ten fingers :), Sunon fans at the back driven at 7V (connected between 5V and 12V source) provide four times 29cfm, or 116 cfm flow.

Since there is a prominent airflow resistance inside flow path (a water heat exchanger), an additional fan is usually recommended, so that’s where the front fan play its role. It must be adjusted to exactly the same airflow as (all) back fans produce to keep internal relative air pressure at zero.

As the Major fan provides 235 cfm (400 cmh) at 3500 rpm (at 24V), 12 V was found to be perfectly adequate to keep it spinning at calculated 1700 rpm in order to provide 116 cfm. The noise was cut from nominal 54 dB to very acceptable 38 dB, so when added to back fans it sums as 40 dB.

Now to put this all together . . . . {mospagebreak}

Miljenko Devcic


Now it was time to put it all together.

Before that, though, just a word on copper pipes entering the waterblock, radiator and water tank.

The selection of their shape, angle and length is made by temporarily mounting those three elements, and “dry-mounting” various pipe shapes to achieve straight and level connection between them. Pipe joints are then marked with pencil and soldered. That way, only 40 cm total length of plastic tubing was required.

I mounted the Peltier between cold plate and waterblock using plastic M3 bolts and spring washers. Thermal paste was spread on both sides. Depron sheets were inserted at this point.

The previously-prepared motherboard (with all the soldering and silicon covering described earlier) was fitted with four 65 mm long M3 bolts around processor socket.

The processor was inserted, thermal paste applied, and waterblock assembly put on the top. Protruding bolts were fitted with springs, flat washers and nuts.
While screwing nuts, spring length is checked in order to get the proper force.

Before mounting, one of those springs was loaded on the digital scale and its length measured at the 20 N (cca 2 kg) point. So the total force applied to processor should be 80 N.

The motherboard was placed in position, as well as the water tank and radiator. All other components were mounted inside the case, but not connected.

Plastic tubing was cut to length and connected with all three “water components” using shrinking tubes and metal ring clamps. Paper towels were wrapped around all water tube joints and pump connected to mains directly, without powering the rest of electrics.

The system is checked for leaks and left for at least 12 hours in order to get rid of tiny air bubbles that lower the thermal conductivity. This won’t be necessary if you use WaterWetter. That is, unfortunately not available in this part of the world, so 15% of antifreeze fluid was added to distilled water instead to prevent algae buildup.

The water pump was then connected to the rest of system. All the components were connected as well and the whole system switched on.

Startup process was held at BIOS setup screen (system health page) in order to monitor the temperatures and be able to react quickly to abnormal values. After 30 minutes or so, as everything was OK, a time consuming tuning process begun.

Let’s see how I did . . .


Miljenko Devcic

The Results

For speed freaks, a processor clock of 2066 MHz was reached (172 MHz FSB x 12), so Athlon XP2600+ could be announced!

On the dark side, there are few serious shortcomings at this speed. Core voltage had to be raised to 2.1 V, so temperature jumped all the way to 30°C and not all tests could be finished (specially demanding graphical ones, such as 3D Mark 2001).

Obviously, processor heat went to levels that the Peltier could not handle, and moreover, the AGP clock was too much for the GF4 (a common occurence with many GeForce4 cards). Unfortunately, the Epox motherboard has no provision for raising AGP voltage.

At the same time, memory clock was able to reach 180 MHz with no problem, proving that Samsung memory is excellent overclockers’ choice, although 512 Mb variety is double sided and not able to reach 200 MHz as 256Mb sticks can.

The next step was lowering the clock to more stable levels. I finally ended up with both FSB and memory clock of 166 MHz, for a still fantastic 2.0 GHz, or as would AMD say, Athlon XP2500+.

For perfect stability on all tests and applications, only 1.80V core voltage was needed. At that speed and voltage, processor backside temperature ranged from 10°C to 16°C. Water temperatures did not exceed 40°C even on hot summer days with room temp reaching 30°C.

Memory voltage was raised to 3V, as this seemed to improve overall stability. The video card was overclocked too, with 320/660 MHz found as the perfectly stable point. The original fancy MSI GeForce4 cooler was replaced with Thermaltake Cristal Orb, which got me 10°C additional cooling.

All the following test results were measured with mentioned settings, BIOS v.2815 was set at Fastest preset, vertical sync to off, with Windows 2000 SP2 and Detonator 4041 used.

Test Results

Sandra 2002 Std.

CPU Arithmetic


CPU Multimedia




AMD N-Bench

10513 (328.55 fps)

AMD N-Bench2

2665 (58.95 fps)

Wintune 98

6114 MIPS


PC Mark


3D Mark 2001SE


3D Mark 2000


XS Mark



78.6 fps

Commanche4 BM

42.97 fps

Code Creatures

32.7 fps