I am a research scientist developing a model of predictive performance using my theory a mathematics of arrays (MoA). I want to obtain power, heat etc and store to csv files. Any and all methods are of interest with a goal of finding the best.
I’m currrently working with @lenore.mullin on writing a script for this sole purpose.
So far, we’ve seen three possible ways:
- CUPTI using the
CUpti_ActivityEnvironmentstruct: CUPTI :: CUPTI Documentation
- NVML using the
nvmlDeviceGetPowerUsage()function: NVML API Reference Guide :: GPU Deployment and Management Documentation
nvidia-smi --query-gpu='timestamp,power.draw' --format=csv loop-ms=1000 -f output.csv
I want numeric data for power and heat as well as time to start and I want to save in csv files.
You can either use NVIDIA GPU management and monitoring command line utility nvidia-smi, or obtain these stats programmatically using the NVML APIs.
Does nvidia-smi command provided by you not serve the purpose? I noticed that double hyphen is missing before the option loop-ms., probably it is lost in the formatting.
nvidia-smi --query-gpu=‘timestamp,power.draw’ --format=csv --loop-ms=1000 -f output.csv