AI: More Energy-Efficient and Sustainable than Humans?
Carbon Footprints: Humans vs. AI Models
The figures paint a clear picture of the actual cost of energy consumption.
Let's start with the human brain, a marvel of energy efficiency, using a mere 12 watts of power.
Contrast this with consumer computers equipped with high-performance GPUs like the NVidia 3090, which can gulp up to 650 watts when engaged in computationally intensive tasks. In a single computer setup, that's over 50 times the energy the brain requires.
However, the stakes soar when we shift our gaze to the business realm.
Consider an NVidia h100 residing in a data centre, demanding roughly 700 watts of power. With the power of 10,000 h100 GPUs combined, a cluster of supercomputers can send this energy demand soaring to more than 500,000 times what your brain needs.
These numbers underscore the escalating energy needs of advanced computing and lay bare the sustainability challenge at the core of the AI revolution.
We're confronted with a fundamental question:
How do we balance AI's rapid growth and our pressing environmental responsibilities?
As we hurtle toward an increasingly digital future, we must grapple with the actual costs of the energy our technology craves.
The Energy Efficiency Dilemma for AI and Human Endeavors
As the adage goes, you can't improve what you don't measure. However, comparing the carbon footprints of humans and AI models presents challenges, given the rapidly evolving field and the need for precise data.
Comprehensive information about hardware, energy consumption, and energy sources makes it possible to accurately estimate AI's carbon footprint.
A study by researchers from the University of California-Irvine and the Massachusetts Institute of Technology (MIT) sparked significant discussion among AI experts. This study challenges conventional wisdom about the energy consumption of generative AI models like ChatGPT.
According to the findings, when generating a page of text, ChatGPT releases 130 to 1500 times less carbon dioxide equivalents (CO2e) than a person performing the same task. Similarly, AI models like Midjourney or OpenAI's DALL-E 2 produce 310 to 2900 times less CO2e than humans when generating images.
The study concludes that AI technology holds the potential to accomplish a wide range of tasks with substantially lower carbon emissions compared to human activities.
Why Comparing AI and Human Abilities Is Complex
The debate surrounding the energy and resource usage of AI versus humans is far from straightforward. It's a nuanced landscape with varying benchmarks. In certain domains, such as chess, AI systems have been faster and more energy-efficient than humans for a considerable time.
However, this only provides a partial view of AI's capabilities across diverse human tasks.
Consider an AI model trained for a year and then deployed billions of times for different individuals, while a human may have honed the same skill over 30 years but applied it only dozens of times. In this scenario, AI might appear more efficient, yet it's vital to acknowledge that humans perform many tasks that AI can only replicate once.
Moreover, besides physical disparities, ethical considerations further complicate energy use comparisons. Humans have multifaceted lives outside their specific "tasks," consuming resources that aren't typically factored into a direct comparison.
In the age of AI, more questions than answers emerge regarding sustainability.
The juxtaposition of AI and human resource usage is fraught with scientific and moral complexities. AI excels at specific tasks but needs to replicate the full spectrum of human existence. The comparison isn't merely about energy consumption; it delves into the intricate reasons behind our choices.
The question transcends energy efficiency; it encompasses the profound implications of our decisions. If AI is to shape our future, we must choose it fully aware of its costs, not just in terms of energy consumption and emissions but also in terms of its impact on our humanity and stewardship of the planet we call home.