Computing Carbon Footprint of AI Model Training

Опубликовано: 26 Ноябрь 2024
на канале: Anuj shah

2,009

This video discusses why computing carbon footprint of AI Mdel is important and how we can do that using a super cool tool MLCO2 Calculator.
The accompanying blog is here - / computing-carbon-footprint-of-training-ai-...

The Machine Learning Emission calculator -
application - https://mlco2.github.io/impact/#co2eq
Paper - https://arxiv.org/abs/1910.09700

Live Electricity Map App - https://app.electricitymap.org/map

Other Tools
CodeCarbon - https://github.com/mlco2/codecarbon
experiment-impact-tracker - https://github.com/Breakend/experimen...

Othe Papers
Energy & Policy Consideration for Deep Learning - https://arxiv.org/pdf/1906.02243.pdf

Blog from Rob Toews about startups working on climate change
https://www.forbes.com/sites/robtoews...

Climate Change AI Workshop - https://www.climatechange.ai/events/n...

Climate Change AI paper - https://arxiv.org/pdf/1906.05433.pdf

Limiting Power in GPU
• This can be done with the nvidia-smi command. First, enable 'persistence mode' for your GPUs: nvidia-smi -pm 1. Then, you can set the power limit for a specific GPU by using the -pl option, for example: nvidia-smi -i 0 -pl 200 (this limits the power consumption of GPU 0 to 200 Watts).
• The number after parameter -i is the GPU number. Further you can see the current power limit of a GPU by using nvidia-smi -q -d POWER -i 2 -l 1 | grep "Power Limit" and the amount of power being drawn using nvidia-smi -q -d POWER -i 2 -l 1 | grep "Power Draw" (for GPU 2)
• If you are worried about how power limitation might affect your training performance, you can find a very good analysis here: https://www.pugetsystems.com/labs/hpc...

Mixed Precsion Training
https://www.tensorflow.org/guide/mixe...

#CarbonFootprint #ClimateChange #AI for ClimateChange #CarbonEmission #DeepLearning