On Tue, 17 Jun 2014 08:46:24 -0700 (PDT)
Tong Zhang <[email protected]> wrote:

> On Tuesday, June 17, 2014 3:36:21 AM UTC+8, Siarhei Siamashka wrote:
> >
> > On Mon, 16 Jun 2014 06:03:18 -0700 (PDT) 
> > Tong Zhang <[email protected] <javascript:>> wrote: 
> >
> > > Hi Siarhei, 
> > > 
> > > Thank you for your reply. 
> > > 
> > > I attached a ds18b20 thermal sensor on top of a10 to measure the 
> > > temperature. 
> > > 
> > > And I drew a graph using following formula (temp_data-1447)*/20. It 
> > seems 
> > > that the two line correlate well on the right(minutes after attached) 
> > > rather than the left(not long after attached). 
> > > 
> > > <
> > https://lh3.googleusercontent.com/-VX9j3c6EWnk/U57oUI6gB-I/AAAAAAAABxM/7FfvvskqmzQ/s1600/sensor.png>
> >  
> >
> >
> > Thanks. While you are at it, can you try to do the measurements for a 
> > wider range of temperatures? And report the actual numbers for a few 
> > points. So that we can calculate the correlation coefficient and find 
> > reasonably accurate linear approximation coefficients. 
> >
> > You can heat up the CPU by using the cpuburn-a8 program as explained at: 
> >     http://linux-sunxi.org/Hardware_Reliability_Tests 
> >
> > And if you set the userspace cpufreq governor, then you can manually 
> > change the CPU clock frequency. Basically, first you let it idle for 
> > a while, then get the readings from the a10 temperature sensor (the 
> > raw number) and the ds18b20 thermal sensor. As a next step you set 
> > some low CPU clock frequency, run cpuburn for a while and also record 
> > the sensor information. Repeat this until you reach the 1GHz CPU clock 
> > speed. 
> >
> > If you also have some A20 hardware, doing the same for it would be 
> > very interesting too. 
> >
> > This would allow us to finally resolve the temperature sensor 
> > calibration mysteries. 
>
> 
> Hi Siarhei, 
> I measured the temperature of a10 using ds18b20 while set a10 at different 
> freq.
> 
> I think this is a rather rough measurement, because I noticed that a10 has 
> big cap area(~ 1cm^2 ). but the ds18b20 is in TO-92 package (~10mm^2).
> 
> The heat is not evenly spread among the chip surface. the PCB may also act 
> like a heatsink. i.e. the temperature inside the chip is not the same as 
> the surface.
> 
> Think twice about the RTP sensor, is it actually measuring the core 
> temperature? Maybe it is placed somewhere far from the core. I have no idea.
> 
> Anyway, I carefully glued ds18b20 in the chip center. and I fitted a line 
> using linear regression.
> 
> got 
> b =[ -198.1115
>     0.1014]
> 
> <https://lh5.googleusercontent.com/-_f728NID9BI/U6BhE9UprxI/AAAAAAAABxg/Xstil8H_URU/s1600/res.png>
> 
> 
> some raw data FYR
> 
> Freq/MHz      ds18b20         a10-raw-reg
> 1010  53.8125         2480
> 912   52.625  2468
> 816   51.75   2464
> 720   50.875  2457
> 600   50.125  2447
> 528   49.6875         2442
> 408   49      2437
> 300   48.5    2433
> 204   48.25   2430
> 120   48.125  2427
> 
> Hope this help.

Thanks a lot! This really helps and the data is perfectly presented.
Based on this data, looks like the steps are the same 1/10 degree as
on A20, but just with a different starting offset.

However the temperature range seems a bit too low (just changing from
~48 degrees to ~54 degrees). Have you used the cpuburn-a8 program
during the test? It should increase the current draw almost up to 1A
when running at 1GHz and the temperature should rise very quickly. I
would have expected that the temperature would swing in a much wider
range (up to at least ~70 degrees). If just cpuburn-a8 alone is
not enough and you are using the sunxi-3.4 kernel, then you can
additionally stress the Mali GPU by using the 'textured-cube'
program from
    https://github.com/ssvb/lima-memtester/
Just run the 'build-textured-cube-static-binary.sh' script to
compile it.

Using a wider temperature range in the data set should improve the
accuracy.

And if this is still not enough, then we also have G2D and CedarX
hardware units, which could be also stressed :) All of this can
increase the power consumption really a lot. This is one of the
reasons why having properly working temperature sensor makes a lot
of sense. Sooner or later, the mainline kernel must implement a
thermal throttling support.

-- 
Best regards,
Siarhei Siamashka

-- 
You received this message because you are subscribed to the Google Groups 
"linux-sunxi" group.
To unsubscribe from this group and stop receiving emails from it, send an email 
to [email protected].
For more options, visit https://groups.google.com/d/optout.

Reply via email to