Big Models pollute Earth
Recent progress in hardware and methodology for training neural networks has ushered in a new generation of large networks. These models have obtained notable gains in accuracy across many NLP tasks. However, these accuracy improvements depend on the availability of exceptionally large computational resources that necessitate similarly substantial energy consumption. As a result, these models are costly to train both financially and environmentally.
This paper: Energy and Policy Considerations for Deep Learning in NLP, published in 2019 by the University of Massachusetts Amherst, is explaining that these big models are costly environmentally due to the carbon footprint required to generate electricity as it’s the main source of electricity in the top three cloud service providers according to the following table:
To heighten the awareness of the NLP community to this issue, researchers of this paper have characterized the dollar cost and carbon emissions that result from training popular off-the-shelf NLP models. They have done that by estimating the kilowatts of electrical energy generated to power the required hardware which can be calculated via the following formula:
\[p_{t} = \frac{1.58t\left( p_{c} + p_{r} + gp_{g} \right)}{1000}\]Where:
-
$p_{t}$: total power consumption (in watts).
-
$p_{c}$: the average power draw (in watts) from all CPU sockets during training.
-
$p_{r}$: the average power draw from all DRAM (main memory) sockets during training.
-
$p_{g}$: the average power draw of one GPU during training. For $g$ GPUs, the average power draw will be $gp_{g}$.
-
$1.58t$: the Power Usage Effectiveness (PUE), which accounts for the additional energy required to support the compute infrastructure (mainly cooling). In 2018 (one year before the publish date of this paper), the global average PUE for data centers was 1.58 per hour.
-
$\frac{1}{1000}$: to convert the power to kilo watt.
According to EPA (Environmental Protection Agency) in 2018, each one kilowatt produces 0.95 pound of CO~2~. Then, the average CO~2~emissions will be:
\[Co_{2}e = 0.95\ p_{t}\]In the paper, they analyzed five different models as shown in the following table. These four models were BERT, ELMO, standard Transformer (T2T), Evolved Transformer (NAS), and finally GPT-2. The following table shows the estimated cost of training a model in terms of CO~2~ emissions (lbs) and cloud compute cost (USD) according to the previous equations:
Based on the previous table and knowing that an air travel from New York to San Francisco emits around 1984 lbs of CO~2~, we can clearly see that training BERT~base~ for around 90 hours will have the same effect. And don’t get me started on the fact that training NAS on 8 Tesla P100 GPUs emits the same CO~2~ as approximately 315 flights from New York to San Francisco.
You can use this tool: experiment-impact-tracker to track energy usage, carbon emissions, and compute utilization of your system.