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
18 stars 24 forks source link

Measuring high-power systems #237

Open psyhtest opened 3 years ago

psyhtest commented 3 years ago

A single-channel analyzer such as YOKOGAWA WT310E with a breakout box such as VOLTCRAFT SMA-10 can measure systems up to 10 Amps (~2.5 kW).

Similarly, a multi-channel analyzer such as YOKOGAWA WT332E and WT333E can support up to 20 Amps (~5 kW) in total (or higher for WT333E?)

Measuring higher-power systems would require several analyzers which the workflow does not support currently.

psyhtest commented 3 years ago

WT1800E supports up to 6 channels.

WT1800E_Rear_no_BOX_M_copy_LG_1

dmiskovic-NV commented 3 years ago

WT333E can support around 7kW powered from 120Vac if using start topology. in case of 240Vac, around 14kW. (appendix3 here https://cdn.tmi.yokogawa.com/IMWT310E-01EN.pdf)

What is the power level we are targeting here? Is it really 36/72kW?

Keep in mind that there are meters (LMG) that if using current shunts can do 100A per element (and they have up to 6 elements)

araghun commented 3 years ago

Got one feedback from Sanjay (SPEC/Intel).

He pointed out to Section 2.4 in this doc. https://spec.org/power/docs/SPEC-Power_Measurement_Setup_Guide.pdf

This says PTD supports single analyzer for single SUT, multiple analyzers for single SUT and multiple analyzers for multiple SUTs. Of course, these need to be ingrained into the power flow to ensure things go as planned + validation, but there shouldn’t be any limitations based on PTD alone. PTD has the capability to collate data from multiple sources and give one output. He mentioned SPEC allows for power measurements on racks as well and for these they use multiple analyzers on multiple systems (scale-out).