New algorithm could reduce energy requirements of AI systems by up to 95 percent

Researchers have developed an algorithm that could dramatically reduce the energy consumption of artificial intelligence systems.

Scientists at BitEnergy AI created a method called "Linear-complexity multiplication" (L-Mul) that replaces complex floating-point multiplications in AI models with simpler integer additions.

According to the study "Addition is All You Need for Energy-Efficient Language Models", L-Mul could cut energy use for element-wise floating-point tensor multiplications by up to 95% and for dot products by 80%. The team tested their approach on various language, vision, and reasoning tasks, including language comprehension, structural reasoning, mathematics, and answering common sense questions.

The researchers say L-Mul can be applied directly to the attention mechanism in transformer models with minimal performance loss. The attention mechanism is a core component of modern language models like GPT-4o.

Direct use in attention mechanisms possible

BitEnergy AI sees potential for L-Mul to strengthen academic and economic competitiveness, as well as AI sovereignty. They believe it could enable large organizations to develop custom AI models faster and more cost-effectively.

The team plans to implement L-Mul algorithms at the hardware level and develop programming APIs for high-level model design. Their goal is to train text, symbolic, and multimodal AI models optimized for L-Mul hardware.