🤖 Is The Ghibli Trend Harmful for the Earth?

Image generation is much more complex and requires more computing power than text generation, which relies on a language model predicting the next word in the sequence.

📊 Computational Complexity:

Images contain more detailed descriptions than text, requiring more power. As such, an image may include details such as shape, colour, texture, structure, and size. These details require intricate processing and consume immense energy due to the complex mathematical operations and algorithms required to generate them.

A Carnegie Mellon University study found that writing a 1,000-word text consumes only 16% of the energy needed to charge a smartphone. On the other hand, as per MIT Technology, generating a single image consumes as much energy as fully charging the smartphone.

💻 Data Volume and Processing Power 

Notably, AI is first trained to process images. The more data-rich images are, the more power they need to process them. AI models undergo training on millions of images to comprehend visual structures. Each generated image requires processing massive datasets in real time, placing a heavy burden on GPUs.

The difference in power generation is stark–text generation requires around 0.042 kWh, while image generation consumes 1.35 kWh per instance.

🧮 Advanced Algorithms at Work

Generative Adversarial Networks and Diffusion models generate realistic images in the desired resolution. But even with these models, making images with a lot of data needs a lot of processing power to work well.

For example, generating 1,000 images with the powerful model, the Stable Diffusion XL, produces emissions equivalent to the amount of carbon dioxide emitted by an average gasoline-powered car when driven 4.1 miles (6.6 km).

This is much more than the least carbon-intensive text production model, which can generate emissions equivalent to 0.0006 miles of driving in the same vehicle. Of course, the power requirement is huge.

Besides this, there are additional environmental costs associated with image generation.

🏭 The Environmental Cost of AI Image Generation

Such colossal power demand does indeed raise some environmental concerns.

⚡ High Energy Consumption:

An investigation shows that training GPT-3, which has 175 billion parameters, consumes 1287 MWh of electricity and emits 502 metric tons of carbon, equivalent to the annual emissions of 112 gasoline-powered cars. Not only that, just a 100-word text prompt in ChatGPT-4 can consume enough electricity to power 14 LED bulbs for an hour. This huge need for energy also needs ways to keep things from getting too hot, which uses water.

💦 Water Usage for Cooling

According to Moody's Ratings, high-performance computing workloads generate more heat than traditional tasks, increasing the need for cooling and water consumption. For example, Google used 5.6 billion gallons of water in 2022, a large portion of which was due to Google's growing AI investments. And just a 100-word email for ChatGPT-4 can consume about 500 ml of water.

Notably, most of this water was clean enough for drinking. Still, new cooling technologies are starting to come out with partial solutions that can cut water use by up to 50% while keeping energy efficiency the same. This surge in water consumption puts pressure on data centres.

❄ Increased Strain on Data Centers

Data centers use air-based cooling to prevent their servers from overheating. Thus, more computational power, such as image generation, can lead to more heat, which can pressure the data center. Such developments can also increase competition for water resources and create tensions between communities, policymakers, and stakeholders, requiring government regulation.