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

[graysky]

Guess I'm not sure what you're plotting. The trend is pretty clear to me: represented by the dashed green line on my plot.

 

[Aquinus]

Guess I'm not sure what you're plotting. The trend is pretty clear to me: represented by the dashed green line on my plot.

I'm plotting the change in max temperature (with ambient factored in) over the "time" column. It's the first derivative of max temperature over both types of runs.

Also, I thought my code was clear... well, I guess that's if you know Clojure. Also I tend to understand my own code.

(defn csv->by-time
  [parsed-csv]
  (let
  [time-map
   (into
     {}
     (map
       (fn [i] [(:time (first i)) (second i)])
       (clojure.set/index parsed-csv [:time])))
   ordered-times (sort < (keys time-map))
   max-temp-ordered-times
   (map
     (fn [i]
       (let [items (get time-map i)]
         [(apply
            max
            (concat
              (map :core0 items)
              (map :core1 items)
              (map :core2 items)
              (map :core3 items)))
          (apply
            max
            (map fahrenheit->celsius
                 (map :ambient items)))]))
     ordered-times)]
  (loop
    [delta-seq []
     last-values (first max-temp-ordered-times)
     remaining-times (rest max-temp-ordered-times)]
    (if (empty? remaining-times)
      delta-seq
      (let [current (first remaining-times)
            delta-cpu (- (first current) 0.0 (first last-values))
            delta-ambient (- (second current) 0.0 (second last-values))
            delta-delta (- delta-cpu delta-ambient)]
      (recur
        (conj delta-seq [delta-cpu delta-ambient delta-delta])
        current
        (rest remaining-times)))))))

It's not polished. I just wanted to get it done. I've spend too much time writing code today which is something I already do in my day job.

Edit: This is instantaneous delta T, not delta T based off the first test. I could include that but I'm done programming for the time being.

 

[graysky]

If you're only extracting the max value from each run I don't believe that is a sound strategy. Select either the median or mean and be sure to exclude the first 400 data points for each run. I don't understand your code at all (not a database guy).

 

[Aquinus]

If you're only extracting the max value from each run I don't believe that is a sound strategy. Select either the median or mean and be sure to exclude the first 400 data points for each run. I don't understand your code at all (not a database guy).

A median or mean would describe the entirety of the graph without specifics of it. This describes the changes over time instead of the values over time. The total of all of these values would describe the total change in temperature.

I did the math and it comes out to be a total change of -3.11 degrees Celsius for the burnP6 runs and mprime is -1.11 degrees Celsius but once again the change in temperature's variability over time is a little confusing and might mean that the criteria for logging may need to be adjusted to reflect a large period of time for runs and for time to cool. I would also suggest that when cooling that the machine is completely shut off to give the paste time to return to or close to room temperature.

 

[graysky]

Sorry dude, it's been years since calculus. My simplistic take on it is: how much has the mean temp for a stress event decreased over time. Set a zero point, test at known intervals, report the mean result and compare back to the zero point. Your rate of change calculation and what it tells me is lost on me I'm afraid.

I have shown a data set that shows a statistically significant decrease in mean stress temps over time. After approximately 8 h, the average temp is more or less still under the starting point by a good 3 degrees. This hasn't really fluctuated much out to 200 h.

 

[Aquinus]

Sorry dude, it's been years since calculus. My simplistic take on it is: how much has the mean temp for a stress event decreased over time. Set a zero point, test at known intervals, report the mean result and compare back to the zero point. Your rate of change calculation and what it tells me is lost on me I'm afraid.

If the paste was curing the general trend would be that the maximum temperature attained by any core on the CPU would decrease. The problem is that the trend isn't clear and the change in max temperature from beginning to end isn't a whole lot and is almost so small it could be accounted for by error.

how much has the mean temp for a stress event decreased over time.

Mean temp isn't a good measure because the number of data points for each test is not consistent and median can be greatly impacted by that difference. Also by measuring the mean once it plateaus isn't much different than taking the max and will result in a smaller number and a smaller change over time. Max temperature doesn't change, even more so if temperatures for any given run have plateaued. The idea is if the paste doesn't transfer heat as well because it has to cure, there would be a marked reduction in maximum temperature of the device as the paste cures and thermal conduction improves if thermal output of the device remains constant. The numbers I calculated from your data is how much that maximum changed from the first run to the last run.

 

[graysky]

I sort of follow your thinking, but have you looked at the distribution of temps per run? By selecting the max, you are intentionally increasing the amount of bounce in the data by picking the highest flyer in the distribution. For example, there are roughly 1400 data points in the prime95 runs. If I understand what you're doing, you are letting a single datapoint from the set represent the entirety of your conclusion for that run.

 

[Aquinus]

I sort of follow your thinking, but have you looked at the distribution of temps per run? By selecting the max, you are intentionally increasing the amount of bounce in the data by picking the highest flyer in the distribution. For example, there are roughly 1400 data points in the prime95 runs. If I understand what you're doing, you are letting a single datapoint from the set represent the entirety of your conclusion for that run.

Yes, just because you have a ton of data doesn't mean that all of it describes what you're looking for. If you're not looking for the change in maximum temperature attained by the CPU then what are you looking for? I can run it based on averages or medians but the problem is that you're introducing variability over time. By having the highest temperature attained you have a point that wasn't exceeded and will be correct once the temperatures plateau. The change in temperature for this to matter is large enough where sensor fluctuation once max temperature is reached won't vary much more than 2*C.

If the maximum can't be trusted as a true maximum and the amount of variance once the device has reached a steady state of thermodynamic balance where temperature doesn't "change" and variance is due to uncontrollable factors like the sensor itself then you can't trust the data either because you're precision is garbage. Regardless, over time as the paste seemingly cures, that max temperature should drop regardless of variation. The slow rate of change and large amount of variation tells me that the sample intervals are too short or not enough is occurring to allow the paste to cure between runs to show a considerable difference.

At least a general downward trend would occur with the total change over time which is the red line I've added to this burnP6 graph:

 

[graysky]

Sorry, I still don't understand your reasoning for using max rather than mean or median for this particular question. Most of the math I do with large sets excludes outliers but your strategy is the polar opposite: you are looking for trends based on them :/

 

[Aquinus]

Sorry, I still don't understand your reasoning for using max rather than mean or median for this particular question. Most of the math I do with large sets excludes outliers but your strategy is the polar opposite: you are looking for trends based on them :/

That's because outliers implies poor precision. Maximum excludes temperatures that don't represent the device getting up to temperature. Even if I were to do as you suggest, it would only make the difference smaller. The trend wouldn't change unless the sensor is really bad and inaccurate (instead of just imprecise) which is a different problem all together.

 

[graysky]

That's because outliers implies poor precision.

Gotta disagree with that... there are distributions of data in many MANY populations. No sensors is perfect. My point is basing an entire run of say 1400 samples on 1 sample or a 0.0007 % of the samples doesn't make sense to me. Also, I do believe that I sampled out enough based on the plots (blue/pink) I showed. Basically, there isn't much if any change in the slope of the lines after 400 seconds. So taking an average of 400-600 should be more or less the same as 600-800 for example.

 

[Aquinus]

Gotta disagree with that... there are distributions of data in many MANY populations. No sensors is perfect. My point is basing an entire run of say 1400 samples on 1 sample or a 0.0007 % of the samples doesn't make sense to me. Also, I do believe that I sampled out enough based on the plots (blue/pink) I showed. Basically, there isn't much if any change in the slope of the lines after 400 seconds. So taking an average of 400-600 should be more or less the same as 600-800 for example.

Right, but you're looking for something specific. I don't think that taking the average of a select portion of the set is a better option in this regard. The question is where to you start the range for calculation of a mean when each run has a different number of samples and might take different amounts of time to heat up? If you want to remain consistent you need solid criteria that's consistent across all runs for determining when heated readings begin and warm up ends. So if you can say that the first x percent of each run is warm up, then maybe but that's still not perfect. Even with variation on maximums, a reduction in temperatures overall would also be followed by a marked decrease in maximum temperatures, variation or no variation. While maximums might not be perfectly accurate, it will yield consistent results over time since the number of samples per run no longer is a variable. Also a maximum is not "one point", it's the highest recorded point which could have several points at that same temperature. The point is that it's an upper bound that should decrease over time and the data only shows a weak correlation to that. Sensors don't swing that high, so I think your concern is unfounded, but we can always do an analysis of variation in temperature over time for each run to see how bad that swing really is.

 

[Arjai]

Nice debate.

I can see both arguments, understand them.

Aquinus is correct.

But OP is not wrong.

Neither of you have a data set that shows the possible TRUTH. But, it is a good start. And a Good Run for averages.

Kudos, for the both of you!! I'm not sure I would devote this amount of time to anything, unpaid.

 

[Aquinus]

Kudos, for the both of you!! I'm not sure I would devote this amount of time to anything, unpaid.

That's why I didn't invest more than a few hours into writing code to handle it. Granted good problems to solve keeps the mind sharp.

Edit: HOWEVER! It's almost like double time because it's not just my day off, but my birthday too. Sometimes I have that programming itch (which is probably 90% of the time for me) so I couldn't help myself. Also don't put it past me to do some variance analysis on the individual runs to see if it really would make a difference or not.

 

[Arjai]

Happy Birthday!!