From Algorithms to Energy: A few fun facts regarding AI's Environmental Impact

Today, we’re going to explore some incredible figures related to AI energy consumption.

1. AlphaGo Zero: Energy Consumption of One Move

AlphaGo Zero, the awe-inspiring Go AI that has left the world amazed, has an energy consumption that’s nothing short of astonishing when it’s training against itself. Every time AlphaGo Zero trains to play itself, it uses about 1 to 2 kilowatts (kW) of electricity. To give you a better idea of its energy consumption, picture the energy equivalent of a human brain. The average human brain uses about 20 watts (W), which means AlphaGo Zero uses the energy of 10 to 20 brains! And get this – every time AlphaGo Zero makes a move, it uses about 200 to 400 Watts (0.2 to 0.4 kilowatts) of energy, which is equivalent to a 60-watt light bulb staying on for 3.3 to 6.7 hours straight. This shows that while AI can perform complex calculations in a very short period of time, it also consumes far more energy than our everyday common sense might suggest.

2. AI Generates Text: Energy Consumption in the Reasoning Phase

In the inference phase of AI, the power consumed per generated text is relatively small, but when this technique is widely used, the cumulative effect of energy consumption cannot be ignored. Take GPT-3, for instance – it uses around 0.1 kWh of power for each request in the inference phase. While this may seem like a small amount, if thousands of users are making queries every day, the collective energy consumption of all these AI inference requests is significant. If we assume that 10 million query requests are generated per day, the daily energy consumption would be:

10,000,000 requests ✖️0.1 kWh = 100,000 kWh

This is equivalent to the electricity used by 5,000 households in a year. This situation, if it continues to happen, will result in significant energy consumption and carbon emissions behind AI.

3. Computational Consumption of Self-Driving Cars

The development of self-driving technology is an exciting area, and it’s great to see the application of AI in cars becoming more popular. Self-driving cars need to constantly collect environmental data, analyze road information, and make decisions, all of which require a lot of computing support – and that takes energy. Each hour of self-driving computing consumes about 3 kilowatt hours (kWh), which is equivalent to two hours of electricity consumption for an average household. Autonomous driving has the potential to change the way we travel and reduce carbon emissions, which is fantastic news. However, the computational consumption behind it needs our attention as well. If thousands of self-driving cars were on the road, overall energy consumption and carbon emissions would increase significantly.

4. Energy Consumption of AI Video Processing

The potential of AI in the film and television industry is truly remarkable, especially in video processing. AI technology is used in various areas, such as video rendering and special effects production, which require substantial computing resources. For instance, rendering one hour of 4K HD video requires around 200 kilowatt hours (kWh) of computational energy, which is equivalent to the electricity consumption of 100 households for an entire year. As AI technology continues to make inroads into the media and entertainment industry, the energy consumption generated during video processing and rendering will gradually increase, further driving up carbon emissions.