Researchers at Cornell University have found a clever trick to make AI models work better while using less energy.
The Problem
- AI models use a lot of complex math operations called “floating point multiplications.”
- These operations require a lot of energy and computing power.
The Solution
- The researchers created a new Linear-complexity Multiplication method: ” L-Mul “
- L-Mul replaces difficult multiplications with simpler addition operations.
Benefits
- Energy Savings: L-Mul could reduce energy use by up to:
- 95% for certain calculations
- 80% for more complex operations
- Precision: Despite being simpler, L-Mul is often more accurate than current methods.
- Versatility: The new method works well for various AI tasks, including:
- Understanding language
- Solving math problems
- Answering common sense questions
Real-World Impact
- When tested on popular AI benchmarks, L-Mul performed just as well as current methods.
- In some cases, L-Mul even outperformed existing techniques while using less computing power.
In simple terms, this research shows a way to make AI models smarter and more energy-efficient at the same time, which could lead to faster, cheaper, and more environmentally friendly AI systems.
Why is it important?
The popularity of Large Language Models (LLMs) has rapidly increased the demand for electricity in data centres. Data centres account for approximately 1-2% of global electricity use but they also host more servers than those used for LLMs, in fact, Bitcoin mining is one of the heavyweights in usage. By the end of the decade, Goldman Sachs estimates that data centre energy consumption will rise to 3-4%.
LLM queries are significantly more energy-intensive than traditional Google searches.
Here’s a breakdown of the energy consumption:
Google Search Energy Use
- A single Google search uses about 0.0003 kWh of electricity.
- This is equivalent to powering a 60W light bulb for 17 seconds.
- 30 Google searches consume roughly the same energy as boiling one litre of water.
ChatGPT Energy Comparison
- On average, a ChatGPT query consumes nearly 10 times more electricity than a Google search.
- This means that just 3 ChatGPT queries use approximately the same energy as boiling one litre of water.
It is easy to see that investment into renewable energy, adopting more efficient cooling technologies and developing more energy-efficient algorithms for LLMs is crucial to handle future power requirements.
References
https://arxiv.org/abs/2410.00907 & https://arxiv.org/pdf/2410.00907
https://googleblog.blogspot.com/2009/01/powering-google-search.html
https://www.goldmansachs.com/insights/articles/AI-poised-to-drive-160-increase-in-power-demand