search

mlcommons / power-dev

Dev repo for power measurement for the MLPerf™ benchmarks
https://mlcommons.org/en/groups/best-practices-power
Apache License 2.0
16 stars 22 forks source link

Issues with measuring power of low power devices #280

Closed dmiskovic-NV closed 1 year ago

dmiskovic-NV commented 1 year ago

Important note:

Problem statement: There are some issues and potentially erroneous (under reporting) power reporting from PTD when measuring power at range near max current of low power devices. Root cause: Meters by default use crest factor of 3. Crest factor is ratio of peak value of current and its RMS value. If current value is overshot, PTD will not detect it because meters state it one status register. (next version of PTD will have that check) Area of exposure: DUTs that have low power factor that are operating below 75W

Proposed fix: change ranging code in manner that if max power consumed is under 75W (and power factor is in at least one sample below 0.8) that requested current range is 40% above highest value of current seen in ranging run. 40% is number i pulled out of thin air, so it is very open to discussion.

Rationale: IEC 61000-2-3 (https://emcfastpass.com/wp-content/uploads/2017/04/Applicability_flow_chart.pdf) exempts <75W devices from adhering to any limitations with regards to number of harmonics and waveform of consumed current example of current waveforms of units with lousy and with decent power factor image ref: OnSemi PFC handbook (https://www.onsemi.com/pub/Collateral/HBD853-D.PDF)

arjunsuresh commented 1 year ago

Thank you @dmiskovic-NV for raising this issue. Can you please do a PR for the change seeds for inference 3.0 is going to be released soon?

dmiskovic-NV commented 1 year ago

more reading material https://assets.testequity.com/te1/Documents/pdf/power-measurements_AC-DC-an.pdf pg 16: Power supplies can often exhibit a current crest factor of around 4 and up to 10. since we want to emulate that with static crest factor of 3, it appears we need to set range 3.33 over highest seen (this will apply only to power supplies under 75W)

s-idgunji commented 1 year ago

If current value is overshot, PTD will not detect it because meters state it one status register. (next version of PTD will have that check) Let's implement a consistent methodology for all submitters once PTD will have it.

Agree fully that we should not go to manual range settings or such approaches.