2 Replies Latest reply on Jul 18, 2013 7:36 AM by mwvantol

    Strange, but regular, long-term power fluctuations on SCC

    mwvantol

      Hi everyone, it has been a while,

       

      We have recently been working on gathering power measurements from the SCC. As everyone else who has worked on power measurements on the SCC knows; the resolution of the ADC on the 3.3V SCC rail can only produce current measurements at a granularity of approximately 0.1A and which results in a power granularity of 0.33W. Therefore we developed a technique to read out the power statistics at a high rate and interpolate these measurements over larger timeframes. This way we could manage to quite accurately detect changes in power when changing the computing load on only a single core.

       

      However, when we were making measurements of the power while changing frequency/voltage at the same time, we measured over longer timescales. We noticed two effects;

      1) after the voltage has been changed, the power still increases/decreases gradually over the next ~150 seconds, we expect that this effect is caused by the chip heating up/cooling down.

      2) after the power of 1) has settled, we notice a strange, but regular, power fluctuation over a long period of time. (see attached image). The power slowly ramps up for about 500 seconds, and then drops again in 250 seconds, creating a slow moving ~750-second period sawtooth form, which has an amplitude of as much as 1 W.

       

      The graph we attached here is a measurement on a completely idle SCC; freshly initialized in the 533/800/800 and nothing booted on the cores. We currently have no idea how to explain the shown behavior, except that we noticed that it comes from a fluctuation in current, and not in voltage. This excludes the power supply from causing this. We have also excluded things like interference between different measurement frequencies; we could observe this wave in each experiment. We have even run one experiment with all the cores of the SCC completely switched off (setting all power domains to 0 V, except the mesh), and we would still observe this, albeit with a much smaller amplitude and more noise.

       

      Is there any explanation of this behavior? Has anyone else observed this on their SCC? Is this an error in the ADC that measures the current? It almost looks like a charging/de-charging effect.