WRDOWY17 heads in water

Edit Jan 2007: Since November three of my wetted sensor tripplets have failed. 3 years in constant use for some of them, the occasinal dry day included. Subjected to a wide variety of antifreezes, surficants, antifungals ...

Excuse the sloppyness here -- my first post.

10/23/2003 03:25p 54,126 mill_melter.JPG

Using a brass rod the same diameter as the WRDOWY17 head, I milled a flat on one side to match the flat of the sensor head. The same thing counld be done carefully with a hand file. Then I heated the rod and melted holes in 8x32 nylon locknuts, which I had previously turned round and small enough to slide in one side of my nylon tubing wyes.

10/23/2003 03:34p 36,232 pre_seal1.JPG

Components cleaned of flashing ready for assembly. The larger fuzzy looking white doughnut, triangulated by the knife, wye, and probe head is the turned, melted, and trimmed up by hand nylon nut.

10/23/2003 03:35p 31,510 pre_seal2.JPG

I`m using 3M Marine Adhesive Fast Cure 5200 here, a much higher grade of sealant than what my local auto and hardware stores sell -- it cures under water, more quickly, more densely ... ( If there are no boating stores near you, check online )

Here are 4 probes with the mounting rings on the probe heads only.

10/25/2003 10:25a 45,474 probes_drying.JPG

Slide the probe head and ring into one end of the wye and backfill it with the silicone. I used an automotive vacum hand pump connected to the flow ends of the wye to draw silicone in -- that picture did not turn out well.

10/24/2003 11:51a 32,533 ptex_hot.JPG

A sealing method that destroyed one probe straight off, and has made two others return frequent "CRC errors" to my 633 was to drip molten p-tex ( used to fill gouges in the base of a ski or snowboard ) behind the probes. One probe-wye put together that way is still functioning properly, most likely from getting no p-tex soot in the wye and from dripping the p-tex in in shallow layers with pauses, so the probe and wires never got too hot. ( The fuzzy black lumpy thing in the upper right corner is my water cooled and sound dampened hard drive enclosure. )

11/07/2003 10:01a 70,510 probe6_qfx_in place.JPG

One completed probe-wye in place. Kind of a sloppy picure there, `scusme. Center backround is a double sided GPU / Video RAM cooler on an NV30GL. Top left slantwise ( drain plug at edge of pic. ) is a now fanless Antec Tru-430 in an oil bath -> copper -> water heat exchanger.

{edit by admin: linked pictures}
Looking for additional LCD resources? Check out our LCD blog for the latest developments in LCD technology.


Last edited:
Having run this setup for several months, I thought I'd post some results.


As the "Rube Goldberg Mark 4" sits on my desk day to day. I made the following thermal management charts in the MK4's native state, air flow blocking clutter and all.

Working at the limits of the Dallas One Wire sensors' accuracy: having brought it up to a stable thermal plateau by running prime 95, I put it in suspend mode for one minute to record a baseline from which to compare measurements under load. This chart is of the 633's six sensors in the flowing coolant under system standby. The comb-like vertical bars off each of the horizontal traces suggest the potential sensor error / innaccuracy of each sensor, as per. Figure 17., page 19, of Dallas Semiconductor's tech sheet ds18b20.pdf. Any errors in math and methodology and logic of course being my own.

Water Flow: from radiators -> sensor "in" -> motherboard waterblock -> sensor "mb-hdd" -> water cooled hard drive mounting -> sensor "hdd-cpu" -> cpu waterblock -> sensor "cpu-gpu" -> graphics card cooler -> sensor "gpu-pwr" -> power supply cooler -> "sensor out" -> radiators.

The same six 633 sensors recording the thermal load of alternating between two minutes at idle and one minute of Prime95 stress-test. Speculative corrections were applied to the raw data; visual buisiness of the error indicating bars removed.

So for each second of recording, one can show how much warmer the water exiting the cpu cooler (sensor "cpu-gpu") is than the water entering (sensor "hdd-cpu"), and plot this as a function of time and system activity. The more stable and darker blue line that drops to zero on system shutdown indicates the coolant circulation rate.

A calorie is by definition the amount of heat energy that can raise one gram of water (= one ml) one degree celcius. There are 4.2 joules in a calorie. To connect to what is familiar to system cooling enthusiasts: a watt is by definition one joule per second. The cpu chart above shows a maximum heat absorption of roughly 34.6 watts. The Barton XP 3000+ is rated ( by silentpcreview, from memory here ) as putting out an abominable 74 watts maximum. Due to the construction of my sensors, about half ( varries between probes ) the probe heads are bound in silicone and nylon and are in a turbulated pocket off the main flow-- they are neither wholly immersed in the coolant nor directly in the coolant flow. The two sensors yeilding the cpu chart above are also separated by approximately 15 to 20 cm of tubing. I speculate that these factors explain most of the lack of precision in measuring thermal flux on the order of 0.0625 degree celcius / second.

The charts and data here should be taken as a light weight "proof of concept" demonstration. They contain no thesis, much freethinking speculation, and no evangelicized conclusions. Other than that I enjoy puttering with my 633.

ZIP file with raw 633_log.csv; MS Excel.xls with charts; charts in bitmap format for those without full MS Excel; notes.txt with more comments, explanations, speculations:

Last edited:
Next time I upgrade my motherboard / cpu / RAM & etc I'll see about retooling my thermal monitoring as well. It's all been stable with the covers on for a couple of months -- wow, did it turn out to be a bigger DIY than I'd initially imagined!

As far as a general sense of what one might hope to get out of the ds18b20 sensors here, here's one more .xls and chart.


The x axis of this chart ( 10 o'clock to 4, left to right ) is the flow rate running through an hypothetical water cooling rig. My flow has been stable around 1.3 liters / minute and this version of the chart details that range. An overclocker ( as apposed to a fan remover / silencer like me ) would prefer a much higher rate, say 8 liters ( ~ 2 gallons ) per. minute. The xls sheet on which the chart is based has "flow start" and "flow increment" cells prominently at the top, and if you change those values the chart will redraw itself accordingly. Ditto for the y-axis ( 7 o'clock to one o'clock ), watts of heat absorbed by a waterblock. The idea is, a particular combination of flow rate and wattage must be above the floor-of-the-chart clear white wireframe band in order to be detected by an hypothetical pair of ideally mounted and calibrated ds18b20's bracketing the heatsource. For my rig, at 1.3 l / m, a pair of ideally installed sensors should detect around 5 watts, and following the 1.3 l / m line up the surface, 72 watts would cause between a 0.75 and a 0.8125 ( 12 -13 sensor increments ) degrees celcius rise accross the pair.


5/24/04 Appology & Erratta:
I have been having trouble with my flowmeter: it freezes in place.
I get to say 'Doh!' and note that while my flowmeter specs clearly state that the turbine wants water filtered to "< 50 microns" and that I have been saying to myself "50 microns" I have in fact been thinking and using "50 thousands of an inch," a 25 fold error in the real world. Slippage in my brain for oh four or five months now... My little 5 watt Eheim I doubt could push a liter a minute through a 50 micron filter.
Lightly here now, I'd suggest http://www.mcmaster.com/, search for part number 9687K11 for an inexpensive possible alternative that I have not used or tested or broken myself.

7/27/04 addendum -- that flowmeter from mcmaster-carr pulses at between 25 and 200 hz -- too low a freqency range for most sensing fan headers.

And I would point out here that in regards to the accuracy of the sensors, a conservative approach would be to demand twelve 0.0625 steps of difference beteen sonsor readings on the last 3D temperature by fow by heat detected chart. ( Hey!, hey! would someone out there with a big budget, tenacity, and a buisness education please kindly steal my ideas and sell me the results once its all been figured out and shrink-wrapped? )

On the plus side I get to note that none of my wetted probes have leaked or failed at this point.
Last edited:

CF Tech

What an incredible amount of thermal detail and analysis you have dug into there.

It positively makes me beam with pride that you are using the CFA-633 (ok, the DS18B20s are the real heroes, but it is the CFA-633 that makes them accessible) in a serious data gathering project.
Not being too satisfied with the quality of data that I was logging and feeling the need to fiddle and twiddle with something, I re- plumbed my rig -- pulling all the wetted sensor heads. I melted apart the nylon tees and wyes, primarily with an eye for anything detracting from accuracy.

From left to right, the three most interesting on takedown:

Probe #7, between hard drive cooler and cpu. This was from my first batch, put together with hot p-tex. It looked the messiest, with p-tex soot and big air spaces. The molten p-tex sure did an ugly job on the wires. The probe head looked well exposed to the flow however, with no spurious blobs of p-tex or silicone. The wires had not actually failed and there was no sign of water past the seal.

Probe #13, water temp out of computer before radiators. From the second batch, all silicone. The little nylon spacer / bushing had rotated nearly 90 degrees on insertion, and silicone spread past the bushing covering most of the sensor head and leaving it in a small pocket off the main flow -- more so than with any of the other 6 wetted sensors. Despite the rotation, there was no sign of moisture on the wires.

Probe #2, between the motherboard cooler and the hard drive. This was the only Tee, with a 90 degree turn in the flow. Another all of silicone. The sensor head was nicely centered, the mount and silicone job all looked clean, maybe a little less silicone waterside of the nylon bushing would hove been pleasing. But on further dissasembly, I saw green copper / solder corrosion seeping out of the seam between the sensor head and the soft wiring harness boot -- exactly what the nylon spacers / bushings were conceived to prevent. One other sensor head (total 2 of 6 here), all silicone and otherwise looking cleanly done and well set showed signs of moisture creeping around to the rear of the sensor head.

Observations of the moment:

The pinkish tinge to the silicone and some of the nylon is from Redline water-wetter (approx one quart / 6 liters distilled water). While the silicone took color, on poking, scratching, and cutting, the water-wetter didn't obviously and markedly degrade the silicone over this short three months.

Any moisture creeping to the back of the heads would corrode the wires in the long run. Tiny little finicky glue job where a spot of finger oil or dust or plastic flashing certaily could ruin a sensor over the years.

P-tex: no.

The little black plastic wiring harness boots on each sensor head can and probably should be cut away carefully prior to this sort of sealing.

The sensors seem fairly robust ay least in the short term -- further as yet unpictured experiments show that they can be sanded, filed, cleaned with citrus-based bio-degreaser from the bike shop and isopropyl alchohol for a better gluing surface, and with a smallish (12 watt) soldering pen stand to have their wires replaced.

Three DS18B20 heads soldered onto one wire set. I was careful to select the triplets in accordance with the order that they report under software -- trying to guess which is which once they're installed would be difficult.

The first coat of Arctic Alumina thermal epoxy. I cleaned the soldering residue off first, and with a slice of 1500 grit garnet sanded around the base for better adhesion. Arctic Alumina in electrically non-conductive, dries quite rigid, and is engineered to stick to chips. Whether it holds up to water and anticorrosive / surfactant solutions in the long run I cannot say for sure: it is a materials testing question beyond by memory of high-school chemistry. I've had the final assemblies wet for a month now with no sign of wire corrosion or leakage. I promise to swallow my pride and post on any sign of failure.

The little plastic jig held the heads in place inside the end of a snip of 3/8" silicone tube while I put on a second coat.
The pictures of the next step didn't turn out. That was inserting the unit above barely into a 1/2" pipe end, troweling in Arctic Alumina epoxy, and popping the probe heads out with a pencil eraser before the epoxy set up fully, & trimming off the slop with a hobby knife. This step put a round flange around the probe heads, as was the intent of the nylon bushings of my first post. Even the tiniest bit of sealant's thermal capacitance degrades the probes responsiveness, accounting for a much of the deviation in probe readings. I carefully scrape off even those little spots.

By repeating the previous step with silicone and letting it dry before cutting away the nylon tubing, I had the four plugs pictured above. Quite sloppy I think to myself. The tinge of orange is from the next step, where I mounted all four plugs in a single chamber temporarily with the mildest most removable silicone I had handy.

Shows the inside of my temporary flow chamber & the cheap seal job on the outside. That particular silicone is actually disc brake quiet, a particularly thin and runny silicone that stays slimy a long time and peels and wipes off most things easily, notably things other than the white silicone my previous step used.

The four plug sets in their sputnik-looking test chamber on the workbench. This was a simple loop with a pump, a reservoir / bubble trap, the heating element out of a broken coffee pot on a variable resistor, and since I lack a water chiller, a bucket of ice cubes to drop in the reservoir for cooling.

full size image
A chart of the 12 temps over time, with the heater at thirty volts as the water circulates through the loop. Sensors in the same plug set are in the same color here. The idea behind logging lots of data on the 12 running together -- several different runs heating and cooling -- is to see just how closely the probes follow each other or how much they vary under known and controlled conditions before putting them into something unknown and complex. I was pleased to see 12 probes reading within only six or seven 0.0625 steps from each other. I suspect that most of the spread in the readings comes from how well each probe is mounted -- in what proportion is each probe head reading the water temperature versus the ambient temperature through its' mounting.

A jig to hold the probe heads at a constant insertion depth and distance from the walls for the final sealing. I put a plugged hose on one barb and used gentle suction on the other barb to draw the silicone slathered plugs into the jig's cup.

The prettiest of the finished four triplets.

full size image
This is a chart of the thermal flux during a run of 3DMark03. (1753 3D marks on the default test, with stereoscopics enabled. For this recording the 633 was run from another computer so that I could log data with this computer shut down. The time course of this trace is: computer off with pump circulating, about two minutes; boot windows and idle, about two minutes; run the default 3DMark03 test (looks about 11 or 12 minutes); idle about another two; shutdown and record with the pump circulating about one minute. The basic relationship (without conversion to watts; corrections for the observed sampling rate; or debate on mean, observed, or corrected sensor error; issues of rounding and significant digits) can be described in reference to the more simple previous chart as "the waste heat absorbed through the CPU block (dark blue here) is the average of the three pink lines minus the average of the three blue lines."
The circulation loop here is: radiators -> visual paddle wheel -> mechanical thermometer -> mechanical temperature switch -> probe set one (probes 1,2,3) -> cpu cooler -> probe set two (probes 4,5,6 -> graphics cooler - > pump -> probe set three (7,8,9) -> motherboard cooler -> probe set four (10,11,12) -> power supply cooler -> radiators. This is in a distinctly crowded mini-atx case, with the probe sets spliced into a preexisting loop that was designed to keep the total length of hose and number of turns and extra loops in the case to a minimum -- lumping together (power supply cooler -> radiators -> visual paddle wheel -> mechanical thermometer -> mechanical temperature switch) is sloppy; being able to measure the total cooling of the radiators and the heat from the power supply separately is only sensible.

The equation for the first point (row 6) in the CPU series above is:

=( ROUND((AVERAGE(H6:J6)-AVERAGE(E6:G6))/0.020833,0)*0.020833 )*C6*4.184/R1

Total elapsed time: 19:49.02; 1189.02 seconds; 1241 original data points; so R1=1.043717 records / second

c6=flow in ml / record

0.0625 / 3 = 0.020833, the precision to which the average of three probes is rounded

4.184 calories to watts conversion constant.

columns E,F,G are the sensors before the CPU block, H,I,J the three after

The CPU block (dark blue series) is neatly bracketed by two probe sets with little tubing either side of the block. The pink series is the graphics cooler and the pump. Yellow is the motherboard cooler. The bottom light blue series is everything else -- most significantly heat lost the air through the radiators.

I speculate that the main reason that the idle values vary is that the air warms by two or three 0.0625 temperature steps over the 14 inch height of the computer case, so as the coolant circulates up and down through the case it warms and cools. The variation while the computer is off is quite in proportion to the height of the probes inside the case. Also, the pump puts a little bit of heat into the system, as do both the power supply (2 amps x 5 volts = 10 watts waste heat potentially available) and the montherboard (I haven't measured its' actual power consumption) even when shut down.

I chose to use 3Dmark for this after reading


or to be honest, reading


by the way.
Last edited: