In the following, we provide additional lessons learned from our own measurements of GPU energy consumption.
The bandwidth of the I2C bus used is limited. Several conclusions must be drawn from this fact:
- Do not attach and enable Tinkerforge Voltage/Current Bricklets that do not provide relevant information for an experiment.
- Measurements from bricklets attached to the bottommost master brick of the stack seem to produce more reliable samples.
- Do not sample the bricklets more often than they acutally produce new data. Sampling every 5 ms provides rather regular samples in our hardware setup.
- If you are only interested in power, do not request voltage and current from the bricklet as the power monitor can internally compute the apparent power, thus reducing the bandwidth required between the bricklets and the master bricks.
- Polling wastes bandwidth. Follow the manufacturer's recommendation of using the asynchronous API.
- Make sure to reset all bricklets at the start of each measurment series. We found that asynchronous callbacks registered from previous measurements might still be active and waste valuable bandwidth if the previous run was not cleanly shut down. Performing a reset unconditionally ensures that the bricklets are in a constistent state every time.
- Read the documentation and most notably the data sheet of the INA226 current/power monitor there. The data sheet provides valuable information about the analog/digital conversion that allow you to configure its timings to match the requirements of your measurement as closely as possible. When sampling every 5 ms, a conversion time of 588 µs and averaging over four samples provides new data every 4.7 ms, which is close to the readout rate.