Detailed study on AS5 break-in time; it only takes a few hours

[graysky]

Purpose
According to the manufacture of AS5, there can be up to 200 hours of break in time required for temps to stabilize. This claim has been tested the results are shared in this thread.

Materials, software, and experiment
A delidded i7-3770k (over-clocked to 44x100 with fixed vcore) was treated with a fresh line of AS5 both under and above the IHS per the manufacture recommended procedure for application to Ivybridge i7 chips (see above link). The Intel stock HFS was then attached and the system was stressed using either mprime (prime95 for Linux) or burnP6 immediately after the HSF was seated to establish as baseline (time = 0) steady-state temperature for each stresser.

Core temperatures, vcore, and CPU fan RPMs were measured using a simple shell script harvesting the linux kernel's lm_sensors readings. Ambient temperature also also recorded using the average of two external thermometers.

Once the baseline temps were established, the experiment was conducted as follows: the stress tests were repeated every 2 hours for several days. This means 42 min of stress (12 min for burnP6 and 30 min for mprime) followed by 78 min of idle all day everyday for 3 days. Then the interval was lengthened to once every 4 hours (42 min of stress, 198 min of idle) for a few days. Finally, the interval was lengthened to once every 6 hours (42 min of stress, 318 min of idle) for the rest of the experiment.

The resulting data were normalized to the average readings for the initial run in order to study the temperature decrease and rate of decrease.

Mprime version 28.5 x86_64 running 8 threads preforming a FFT of size 440k.
cpuburnP6 has no settings for multiple cores but it was run in 8 independent processes so all cores would be stressed equally.



Conclusions
The temps basically stabilize after 8 hours -- no where near the 200 hours the manufacture claims -- in this particular test. The magnitude of the decrease was 3-4 C relative to the initial temps depending on the stress methods which is pretty good. How general is this result? The hallmark of a scientific experiment is a reproducible outcome. This is just an n of 1 but no plans are in the works to repeat the experiment. Perhaps others will do it and report back. Feel free to use this thread if you do.

Extra details
Calculating zero
Zero was calculated simply by stressing with the program (either mprime or burnP6) and waiting for the temps to stabilize. In the plot of core temp adjusted for ambient vs. time (seconds), the pink color indicates data points I excluded since the CPU was coming up to temperature while the blue color denotes datapoints that were included. The higher temp/shorter line is the temperature profile for burnP6 and the lower temp/longer line is the profile for the mprime run. Basically, cpuburnP6 was run for 12 min and mprime for 30 min, excluding the first 400 sec from each one. The 2nd plot just shows the same data unnormalized for ambient temp and split out by individual core:

Data
This plot shows the time since the HFS was seated (days) vs. delta temperature "dT" above ambient for each of the stress programs. Initially, they were run every 2 hours, but after 3 days or so, the interval was lengthened to every 4 hours, and finally to every 6 hours until approx 200 hours had elapsed. Both stress programs gave similar results with the average temp drop occurring sometime around 8 hours and stabilizing around 3-4 degrees less than the initially measured temps.

The first plot shows the average of all cores and the 2nd one shows the values for the individual cores.

Data table
The data table generated by the script linked above is available in csv format. A more elaborated table where I preformed some calculations needed for the plots (difference from ambient temp, which points are included or excluded, average for time=0 for each stresser, etc) is also available in csv format should you wish to plot anything yourself.

 

[graysky]

OK, ran out to ~200 hours and as expected, no change in the conclusion. The original post was updated with these results.

 

[Sasqui]

Nice analysis and presentation. Appears 1/4 day will do the trick!

 

[Aquinus]

It's good except AS5 does call for multiple heating/cooling cycles and I found with my 6870 that it took 3 heating/cooling cycles after running the GPU at 80*C for ~20 minutes each run got temps to drop. To maintain 80*C I adjusted voltage to maintain a steady temperature. Each test required more voltage to maintain 80*C so I would make sure you examine the entire problem, not just the first run. So I still don't think this is giving the data you're really expecting. I think you have a couple more tests to do.

So more tests and continuing each test beyond only one heating cycle is required before you can say for certain that you know what's going on with respect to empirical data.

 

[EarthDog]

Didn't he do that? With decreasing frequency?

 

[MT Alex]

Nice job, and a lot of effort put into your study, but I didn't know people still used AS5.

 

[Steevo]

Nice job, and a lot of effort put into your study, but I didn't know people still used AS5.

Its still a very very good TIM.


I believe the 200 hours may be for very low temperatures, since the phase change requires heat to squish the extra out if you have a CPU running under just above ambient water cooling I found it took awhile to get it to set. I used a hair dryer back int he day to see what the spread looked like on a lapped IHS, it melted quickly with 1500W of heat blowing into it.

 

[hat]

Man AS5 has been around since I got into this hobby...

Then I heard about MX-2. Better thermal transfer than AS5 and non-conductive. Also easier to apply and clean. Now I believe MX-4 is the newest variant of it so I would just go with that if buying TIM. There's other weird TIM around like liquid metal whatever with better thermal transfer than MX-4 but it could also be dangerous.

 

[Aquinus]

Didn't he do that? With decreasing frequency?

I only see 0 - 400 minute runs. I don't see heating and cooling cycles anywhere. It appears to be one long loaded test.

It's important to note that letting the hardware cool down is an important part of the curing process and I feel like that important detail was ignored. While covering all the bases might take time, we're not getting better results than your last attempt because it's still only one heating cycle. At least that's what I gathered from the data the OP provided.

I would also like a list of all the variables that were made to be constant because fewer constants means more variability in the results. For example, were all fans set to a static speed and was the speed consistent through all the tests?

 

[graysky]

I only see 0 - 400 minute runs. I don't see heating and cooling cycles anywhere. It appears to be one long loaded test.

No, you don't have that quite right, see here:

Basically, cpuburnP6 was run for 12 min and mprime for 30 min, excluding the first 400 sec from each one.

and, here:

Initially, they were run every 2 hours, but after 3 days or so, the interval was lengthened to every 4 hours, and finally to every 6 hours until approx 200 hours had elapsed.

So collectively, it was 42 min stressing, 78 idle, then repeat for 3 days, then once every 4 hours for a few days, then once every 6 hours for a few days. Many power cycles in other words.

I would also like a list of all the variables that were made to be constant because fewer constants means more variability in the results. For example, were all fans set to a static speed and was the speed consistent through all the tests?

Look at the data table, fan RPM speed was maxed, vcore was constant, ambient was controlled for by difference. The data are very tight. Note the error bars in the plots (they are tiny).

 

[graysky]

Now I believe MX-4 is the newest variant of it so I would just go with that if buying TIM.

I tested MX4 multiple times and was disappointed with the results. Always hotter than AS5 without curing. Maybe I got a bad tube? Others have reported better.

 

[Aquinus]

No, you don't have that quite right, see here:

and, here:


So collectively, it was 42 min stressing, 78 idle, then repeat for 3 days, then once every 4 hours for a few days, then once every 6 hours for a few days. Many power cycles in other words.



Look at the data table, fan RPM speed was maxed, vcore was constant, ambient was controlled for by difference. The data are very tight. Note the error bars in the plots (they are tiny).

I stand corrected. I misunderstood the post. Good work. Maybe it's how you represented it.

If you have the data handy could you create a different graph? I'm interested in not the delta T over ambient, but rather the delta T with respect to the result before it. So instead of averages, I want a rate, so a differential on temperature.

For example, if you have three temperature readings over 3 seconds, 40, 43, and 46, you would have two points of 3 and 3 because the temperature over the last result changed by that much. I feel that would be an easier visualization for when the CPU is truly done curing as temperatures aren't changing any further. You can still factor ambient temperature into that, but really what I would like to see is a differential on temperature as the rate of temperature change over all the tests is really what we want I think.

The idea is when that kind of graph approaches zero, you know when it's done curing, but while it's curing numbers will be negative and approach zero as the rate of delta T should be, for the most part, negative.

 

[graysky]

I stand corrected. I misunderstood the post. Good work. Maybe it's how you represented it.

Not a problem, perhaps I didn't make it clear... I'm not a very good writer.

I'm interested in not the delta T over ambient, but rather the delta T with respect to the result before it. So instead of averages, I want a rate, so a differential on temperature.

Not exactly sure how to plot that... I have made both the raw data and the expanded data tables available. Perhaps you can plots it and post the result.

 

[Aquinus]

Not exactly sure how to plot that... I have made both the raw data and the expanded data tables available. Perhaps you can plots it and post the result.

Certainly but I can't find the data. If you can direct me to it, I should be able to generate the graph in open office or something.

Edit: Never mind, I see it now. Give me a little bit to do the calculations and plot it.

Edit 2: Open office doesn't let me work with the data you gave me in the way I want it to work. I'm going to try loading the CSV into Clojure and writing a little bit of code to do the conversion for me then spit it back out into a CSV document. It might take me a little bit. I can't write code without having coffee in the morning. I'll provide the program on github, the output, and the graph when I'm done.

Edit 3: I was successfully able to load the CSV into a sequence of hash maps in Clojure. I'm going to start munging the data in groups by each run, pulling out the max CPU and ambient temperatures for each run and using that to calculate the difference between each successive run.

For clarification: I'm more interested in the aggregate numbers for each run compared against each other as opposed to individual runs. If this data were in a relational database like PostgreSQL, I could have written some SQL to calculate this a lot easier, but that's a different kind of discussion for another time.

 

[graysky]

You might need to unpivot the data with respect to core0-core3 which is how I worked with it in the plots. No idea why openoffice can't read a csv file... did you uncompress it first? It is compressed by gzip.

 

[GreiverBlade]

I tested MX4 multiple times and was disappointed with the results. Always hotter than AS5 without curing. Maybe I got a bad tube? Others have reported better.

i wonder ... i tested both and the result where inverse for me ... and with the Gelid GC Extreme now : AS TIM are outdated and either Arctic (ex Arctic Cooling) and Gelid TIM does same or better for a lesser price. (personal opinion ofc)

 

[Aquinus]

You might need to unpivot the data with respect to core0-core3 which is how I worked with it in the plots. No idea why openoffice can't read a csv file... did you uncompress it first? It is compressed by gzip.

Yeah, it read the CSV fine. I used 7zip to decompress it on my tower and gzip on my server which is where I'm working with it. It's that OpenOffice doesn't do the kind of computations that I want to do to the data.

I do have a question, max ambient temps appear to be in the low "60's". That seems really hot for ambient temps. How is ambient determined?

Side note: From the data I've generated so far, it doesn't appear that much of a change has occurred. These are ordered max temps. The first is the max of core0-3 achieved for any "time" and the second is the max of ambient at any "time". These are in order from time 0 to time 62 I believe (63 items).

([86 61] [85 61] [84 61] [84 61] [83 61] [83 61] [84 62] [83 62] [83 61] [83 61] [82 61] [83 61] [82 61] [82 61] [81 61] [82 61] [81 61] [82 61] [82 61] [82 61] [82 61] [82 61] [83 61] [82 61] [82 61] [81 61] [80 61] [82 61] [83 62] [83 62] [83 62] [83 62] [84 63] [84 63] [84 63] [84 63] [84 63] [85 63] [84 63] [86 63] [85 64] [86 64] [86 64] [84 64] [84 64] [85 63] [84 63] [85 63] [84 62] [85 62] [86 63] [86 64] [85 63] [84 63] [85 63] [85 62] [85 62] [85 63] [85 64] [84 63] [84 63] [84 63])

I'm still working on it but so far, I don't see much change in temps after each run but I'm also not done. If ambient temps are in Fahrenheit and core temps in Celsius, I need to know.

 

[graysky]

Low 60s in F scale... there is a col that converts it to C since the coretemp values are in C.

 

[Aquinus]

So far these are the numbers that came up when I calculated the difference between each run.

[[-1.0 0.0]
[-1.0 0.0]
[0.0 0.0]
[-1.0 0.0]
[0.0 0.0]
[1.0 0.5555555555555571]
[-1.0 0.0]
[0.0 -0.5555555555555571]
[0.0 0.0]
[-1.0 0.0]
[1.0 0.0]
[-1.0 0.0]
[0.0 0.0]
[-1.0 0.0]
[1.0 0.0]
[-1.0 0.0]
[1.0 0.0]
[0.0 0.0]
[0.0 0.0]
[0.0 0.0]
[0.0 0.0]
[1.0 0.0]
[-1.0 0.0]
[0.0 0.0]
[-1.0 0.0]
[-1.0 0.0]
[2.0 0.0]
[1.0 0.5555555555555571]
[0.0 0.0]
[0.0 0.0]
[0.0 0.0]
[1.0 0.5555555555555536]
[0.0 0.0]
[0.0 0.0]
[0.0 0.0]
[0.0 0.0]
[1.0 0.0]
[-1.0 0.0]
[2.0 0.0]
[-1.0 0.5555555555555571]
[1.0 0.0]
[0.0 0.0]
[-2.0 0.0]
[0.0 0.0]
[1.0 -0.5555555555555571]
[-1.0 0.0]
[1.0 0.0]
[-1.0 -0.5555555555555536]
[1.0 0.0]
[1.0 0.5555555555555536]
[0.0 0.5555555555555571]
[-1.0 -0.5555555555555571]
[-1.0 0.0]
[1.0 0.0]
[0.0 -0.5555555555555536]
[0.0 0.0]
[0.0 0.5555555555555536]
[0.0 0.5555555555555571]
[-1.0 -0.5555555555555571]
[0.0 0.0]
[0.0 0.0]]

This represents the change between each pair of runs. Since they're integers but the conversion for Fahrenheit to Celsius requires a decimal number there will be a lot of similar looking numbers after decimal points. I'll be adding the difference between the two deltas and spitting it out for a simple graph. I've already spent too much time on this I feel.

 

[Vario]

Give it a few months and see if its lower. Its still premature to reach conclusions. I've found that I still got improvements on AS5 months later.

 

[graysky]

So far these are the numbers that came up when I calculated the difference between each run.
...

Still not sure what you're plotting.. you want the ave diff between the endpoints in my graphs above?

time = time in hours since seating
burnP6 = average temp accounting for ambient from burnP6 stress
mprime-440k = average temp accounting for ambient for mprime stress

time    burnP6    mprime-440k
0    0.00    0.00
1    -1.37    -1.65
3    -2.25    -2.63
5    -2.81    -3.42
7    -3.20    -2.80
9    -3.38    -2.40
11    -3.83    -2.91
13    -4.19    -3.15
15    -3.30    -2.33
17    -3.47    -3.01
19    -4.18    -3.49
21    -4.22    -3.39
23    -4.08    -3.89
25    -4.62    -4.11
27    -5.60    -4.79
29    -5.17    -4.15
31    -4.64    -3.98
33    -4.24    -3.38
35    -3.99    -2.70
37    -4.39    -3.16
39    -4.43    -3.31
41    -4.52    -3.77
43    -3.76    -2.87
45    -3.71    -3.27
47    -4.14    -4.09
49    -4.98    -4.58
51    -5.22    -4.19
53    -3.90    -4.33
55    -4.30    -4.03
57    -4.40    -3.46
59    -4.27    -3.04
61    -3.52    -2.60
63    -4.61    -3.82
65    -4.46    -3.44
67    -4.09    -3.23
69    -4.05    -3.55
71    -4.19    -3.31
73    -3.61    -3.20
75    -3.85    -3.10
79    -3.59    -2.74
83    -4.40    -3.85
87    -3.04    -3.92
91    -4.32    -4.55
95    -4.91    -4.95
99    -5.14    -4.88
103    -3.44    -3.75
107    -4.40    -4.14
111    -3.36    -3.65
119    -3.61    -3.57
125    -3.04    -2.88
131    -3.58    -3.83
137    -3.83    -3.85
143    -3.51    -3.54
149    -3.75    -3.64
155    -3.48    -2.80
161    -2.06    -2.18
167    -2.77    -3.23
173    -3.07    -2.70
179    -3.55    -4.01
185    -3.55    -4.01
191    -3.82    -2.86
197    -3.53    -2.95

 

[Aquinus]

There are aggregate delta T numbers that I'll be graphing shortly:

(-1.0
-1.0
0.0
-1.0
0.0
0.44444444444444287
-1.0
0.5555555555555571
0.0
-1.0
1.0
-1.0
0.0
-1.0
1.0
-1.0
1.0
0.0
0.0
0.0
0.0
1.0
-1.0
0.0
-1.0
-1.0
2.0
0.44444444444444287
0.0
0.0
0.0
0.4444444444444464
0.0
0.0
0.0
0.0
1.0
-1.0
2.0
-1.5555555555555571
1.0
0.0
-2.0
0.0
1.5555555555555571
-1.0
1.0
-0.4444444444444464
1.0
0.4444444444444464
-0.5555555555555571
-0.44444444444444287
-1.0
1.0
0.5555555555555536
0.0
-0.5555555555555536
-0.5555555555555571
-0.44444444444444287
0.0
0.0)

Still not sure what you're plotting.. you want the ave diff between the endpoints in my graphs above?

time    burnP6    mprime-440k
0    0.00    0.00
1    -1.37    -1.65
3    -2.25    -2.63
5    -2.81    -3.42
7    -3.20    -2.80
9    -3.38    -2.40
11    -3.83    -2.91
13    -4.19    -3.15
15    -3.30    -2.33
17    -3.47    -3.01
19    -4.18    -3.49
21    -4.22    -3.39
23    -4.08    -3.89
25    -4.62    -4.11
27    -5.60    -4.79
29    -5.17    -4.15
31    -4.64    -3.98
33    -4.24    -3.38
35    -3.99    -2.70
37    -4.39    -3.16
39    -4.43    -3.31
41    -4.52    -3.77
43    -3.76    -2.87
45    -3.71    -3.27
47    -4.14    -4.09
49    -4.98    -4.58
51    -5.22    -4.19
53    -3.90    -4.33
55    -4.30    -4.03
57    -4.40    -3.46
59    -4.27    -3.04
61    -3.52    -2.60
63    -4.61    -3.82
65    -4.46    -3.44
67    -4.09    -3.23
69    -4.05    -3.55
71    -4.19    -3.31
73    -3.61    -3.20
75    -3.85    -3.10
79    -3.59    -2.74
83    -4.40    -3.85
87    -3.04    -3.92
91    -4.32    -4.55
95    -4.91    -4.95
99    -5.14    -4.88
103    -3.44    -3.75
107    -4.40    -4.14
111    -3.36    -3.65
119    -3.61    -3.57
125    -3.04    -2.88
131    -3.58    -3.83
137    -3.83    -3.85
143    -3.51    -3.54
149    -3.75    -3.64
155    -3.48    -2.80
161    -2.06    -2.18
167    -2.77    -3.23
173    -3.07    -2.70
179    -3.55    -4.01
185    -3.55    -4.01
191    -3.82    -2.86
197    -3.53    -2.95

As it stands, everything. I can limit it to one or the other since temps seem to be different.

 

[graysky]

Give it a few months and see if its lower. Its still premature to reach conclusions. I've found that I still got improvements on AS5 months later.

I will check it again in a few weeks.

 

[Aquinus]

So I pulled the two apart and they don't seem to be all that much different. I'll graph them individually. I just need to dump it into some CSV. The code isn't something you can just run. I've been writing code in a file and evaluating it as I go in the REPL. I'll put it somewhere public when I'm done.

Code is on github, here's the core source file that has everything in it.
https://github.com/jrdoane/graysky-analysis/blob/master/src/calc/core.clj

Rest is coming soon.

 

[Aquinus]

The graphs are here. Judging by the look of them, max temperatures (with ambient temperate factored in) seems to be hard to judge. To me, it seems that looking at the data this closely doesn't give us any picture of what's going on. A good scientist doesn't jump to conclusions, but to me, the lack of downward trend to me stays the the individual runs are either not controlled or the paste hasn't cured yet.

Remember, these are aggregate max temperatures with ambient factored in used to sample rate of change (differential calculus if you will) to sample the change in max temperature between runs. As you can see, there isn't an obvious downward trend of max temperatures over time.

These graphs represent what I would call ΔTmax-adj if I had to give it a name.

Here is the graph for MPrime:

Here is the graph for burnP6:

To me, these graphs look balanced and is telling me that the change in temperature over time is practically zero as the number of lines above zero are almost equal.

Edit: Including the CSV that was generated. Col1 is ΔTmax-cpu, Col2 is ΔTmax-ambient, and Col3 is ΔTmax-adj.