For the research purpose, I use nvprof (version: 8.0.27 (21)) to do the profiling work of GPU. From the documents of nvprof, it will report the power with flag system-profiling “on”.
What is this power metric stands for? I’m wondering if it’s only the GPU chip power or includes peripherals or others. Thanks
It includes the power in milliwatts consumed by the GPU and associated circuitry (e.g. memory).
NVIDIA GPU management and monitoring tool nvidia-smi can also be used which provides more options to query GPU information. Refer link NVIDIA System Management Interface | NVIDIA Developer
Sample query to generate the power sample every 10 ms.
$nvidia-smi --query-gpu=power.draw --format=csv --loop-ms=10
1 Like
So can we say that it represent the total power consumption of GPU? I’m wondering what other stuff “associated circuitry” implies. Can you please give me more elaboration? And how do I measure only GPU chip’s power? Thanks.
Yes, it represents the total power consumption of the GPU. By associated circuitry, I meant to say all the units in the GPU including SM, memory, encoder, decoder etc. There is no control in the nvprof to measure the power consumption at the unit level.
Thanks for your nice reply.
Please let me ask a final confirmation that it does not include the power consumed by the rest of the graphics board components?
Thank you!
It does include the total power draw of the entire GPU board.
How does nvidia-smi measure total board power? An SVID or SVID-like bus could provide GPU core power, and architectural counters could probably estimate external DRAM power with reasonable accuracy, but what about the rest of the add-in card? Does the video BIOS include fields for the board manufacturer to report PSU efficiency and to account for non-GPU load on the card?
If the answer varies across GPU segments, I’m specifically interested in the Quadro P2000, N18P/M in the GB4c-128 package.