REAL benchmarks with REAL heatsink

Here is a copy of the tool I wrote to test boards I use, your welcome to use however you like. I have not experienced the issues others have probably because of how I do things verses how they may do things. Photo no fan of heat sink.

Photo with fan

Heat load and frequency plot from data. Ondemand Mode

Performance Mode

Raw data sets. (89.3 KB)

Performance (73.4 KB)

Actual set of REAL performance tests. (6.3 KB)

The tests are simple using stress-ng for 5 minutes and 10 minutes I warm the processor and heatsink up.

It then loads the processor with sysbench on 8 16 32 48 64 threads.

The test then measures the change in temperature of the processor during a 2 minute cool down.

Next it creates a ram disk as to not damage your emac or flash card. It then launches in the ramdisk a script that generates 10,000 unique 2048 character hexadecimal palindromes. Then it changes the case of each abcdef to ABCDEF and writes to a second file. The second part of the script then takes one line at a time from the second file finds the match inn the first file and exits. Typically you will see LOAD numbers in the mid 60s. If the numbers do not match the test fails and it means there was data corruption. This same test run on and AMD Opteron 2.1Ghz 8 core Takes about 22 minutes typically.

After all the test are done it will create the report files and graph. The script checks for the dependancies and installs them. The test should only be used with proper power and cooling it will overheat the 1.50 cent solution posted recently.

More photos

IR.mp4 (220.2 KB)


Well, those benchmarks are not so REAL but rather lightweight and with a really cheap heatsink + fan ($1.50) performance results don’t look different at all. Here is my graph even if I don’t know what it should display other than temperatures all the time remaining below throttling treshold and therefore having no influence on performance anyway:

Now using a REAL load (cpuburn-a7):

My small heatsink + mini fan are too weak, cpufreq even drops down to 1.2GHz. Possible improvements:

  • hardware (better heatsink, stronger fan, controlled airflow)
  • software (improving DVFS table, adding more cpufreq OPPs, adding a lot more intermediate steps in cooler_table)

But why? My other M3 all perform better at lower clockspeeds and have a nice fansink that is efficient for full load :slight_smile:

Lots of heat there 70C plus Should be triggering first stage of throttle

Yes I see the throttle from heat and the instability you are experiencing I am not having these issues.

Unbelievable! :joy::joy::joy:

I was running cpuburn-a7, you all the time run only light workloads. What about using some heavy stuff?! :joy::joy::joy:

You must be @sinovoip’s cousin or brother!

Your time on CPU-RAM is like 17 minutes verses 10-12 on mine Also Heat temperature is too high

No relation just not having the troubles you have doing things your way.

To recap I see that my solution works repeatedly and reliably with none of the issues complained about by Charles.

I do things my way it works other people do things their way it doesn’t.

The end of the day what I am doing is successful and I do not care what unsuccessful people do.

LOL! ‘And by showing that impressive amount of ignorance I totally reject reality’.

You’re a genius since you don’t get such simple facts that cpuburn-a7 is a heavy load and your scripted stuff only lightweight and laughable ‘load generation gone wrong’. :joy:

Good luck! You need it! :slight_smile:

Charles CPU burn only calculates PI.

You made my point with your charts and data files.

You are having problems with your hardware the way you continue to use it I am not having problems with my method.

Don’t always show your team knowledge by sneering or attacking others. You can be more friendly. :slight_smile:

haha , i also think this : You must be @TK(tkaiser) cousin or brother!