The Cost of Generating an AI Ghibli Studio Image

As the world crazes over Studio Ghibli images generated by AI, let’s put the environmental costs into perspective.

Below we have highlighted the technical aspect of arriving at these calculations. You may prefer to go straight to real world examples.

• High-end GPUs like NVIDIA A100 or V100 (commonly used for AI workloads) consumes around 250-400 watts or power during their inference.

  
  

• Assuming an image generation process time of 30 seconds and a 300-watt GPU, energy consumption during this time would be :

  
  

  300 watts x (30/ 3,600) hours = 2.5 watt-hours (Wh)

• Data centers operate with an efficiency ratio measured by the Power Usage Effectiveness (PUE) that account for energy used to power cooling systems, services and other infrastructure.

  
  

• A PUE of 1.3 (meaning that for every watt used by the server, an additional 0.3 watts are used for cooling and other infra needs), each 2.5 Wh of energy consumed by the GPU, actual energy consumption would be closer to:

  
  

  2.5 Wh x 1.3 = 3.25 Wh

  
  

  3.25 Wh is equal to typical home light bulb running for about 19.5 minutes.

Real-world usage examples:

In collaboration with InvestorSight by 7i Ventures, we ran the following real-word energy usage examples. This puts the use of energy into perspective.

This is not mean to be a buzz kill and make people feel bad about getting fascinated by cool stuff, but a little perspective and knowledge can help us understand the consequences of our action.

Also read: A Timeline of Actions Against Sustainability and Climate by Trump

Subscribe to our newsletter to receive monthly updates on the sustainability industry.