Скачать или смотреть You Won’t Believe How Much Power ChatGPT Really Uses!

Here's a breakdown of what we know about ChatGPT's (and other large language models') power consumption, focusing on recent estimates:

Key Takeaways:
Per-Query Cost is Relatively Small, But Scales Massively:
A commonly cited figure for a typical ChatGPT query (using a model like GPT-4o) is around 0.3 watt-hours (Wh) of electricity.
This is often compared to a standard Google search, which uses significantly less (e.g., 0.0003 kWh or 0.3 Wh, making ChatGPT perhaps 10 times more).
However, more complex or longer queries (e.g., 10,000 tokens, like a short paper) can increase the cost to around 2.5 Wh, and very long inputs (100,000 tokens, 200 pages) could be up to 40 Wh.
It Adds Up Quickly at Scale:

Despite the small per-query cost, ChatGPT handles an enormous volume of requests. With potentially hundreds of millions of queries daily, the cumulative energy consumption becomes substantial.
Estimates suggest ChatGPT's global energy consumption could reach approximately 621.4 megawatt-hours (MWh) per day, or 226.8 GWh annually.
Comparisons to Put it in Perspective:

Households: ChatGPT's daily energy use can be equivalent to powering tens of thousands of American homes for a day, or even a small country like Barbados for a year (based on older estimates).
Electric Vehicles: The annual energy used by ChatGPT could fully charge millions of electric vehicles.
Other Industries: While significant, its total energy consumption is still a tiny fraction (e.g., 0.0008%) of annual worldwide electricity use, much less than highly energy-intensive industries like beef production.
Not Just Electricity – Water Usage for Cooling:

There are two main phases of AI energy consumption:
Training: The initial, massive computational effort to "teach" the AI model. This phase consumes a huge amount of energy, often taking weeks or months on thousands of GPUs. For example, training GPT-3 was estimated to consume 1,287 MWh of electricity.
Inference: The ongoing energy used when users query the trained model (generating responses). While a single inference is less costly than training, its continuous, widespread use can accumulate to more than the training cost over time. Recent studies suggest that 60-70% of AI's total energy use comes from inference.

Factors Influencing Consumption:
Model Size: Larger, more complex models (like GPT-4 and beyond) generally consume significantly more energy.
Prompt Complexity: More detailed or longer prompts require more computational resources.
Server Efficiency & Hardware: Advances in GPU technology (like NVIDIA's H100 vs. older A100) and data center optimization (e.g., advanced cooling, batching requests) can reduce energy usage.
Algorithm Efficiency: More efficient AI algorithms can achieve similar performance with less computation.
The Bottom Line:

While the energy cost of a single ChatGPT query is relatively low, the sheer scale of its usage means that the cumulative energy (and water) footprint is substantial and growing. This has led to increasing discussions about the environmental impact of AI and the need for more energy-efficient models and sustainable data center practices.
Instagram   / curiousclutv  
Facebook   / curiousclutv