With the rise in use of artificial intelligence and advancements in technology, the concern about energy consumption by AI has also grown. Artificial intelligence systems consume a varying amount of energy based on their complexity and usage, but generally they require a significant amount of energy to process and analyse data efficiently. Innovative solutions like BitEnergy’s L-Mul are emerging to address this challenge, offering energy-efficient technologies designed to optimize AI processes. For instance, responding to a query can consume about 10 times more electricity than a simple Google search.
According to the World Economic Forum, with Microsoft’s investments in AI and making generative AI a key part of its services, there has been an increase of 30% in CO2 emissions since 2020. Google’s GHG emissions were also almost 50% higher than they were in 2019. So, while AI tools have the potential to aid in energy transitions, they require a lot of computing power too. Keeping these complexities in view, scientists at BitEnergy have unveiled an exciting new technique known as ‘Linear-complexity multiplication’ (L-Mul). It simplifies the complex floating-point multiplications in AI models into straightforward integer additions.
The Impact of L-Mul on AI Efficiency (BitEnergy’s L-Mul)
In the study “Addition is All You Need for Energy-Efficient Language Models”, researchers have found that L-Mul can slash energy consumption for element-wise floating-point tensor multiplications by an incredible 95% and by 80% for dot products. This new approach has been tested on various tasks, ranging from language comprehension to structural reasoning to maths and common sense questions.
Another interesting part is that researchers believe L-Mul can be seamlessly integrated into attentional mechanisms in transformer models like GPT-4, without significantly impacting the performance. BitEnergy AI envisions L-Mul enhancing both academic and economic competitiveness, contributing to AI sovereignty and enabling large companies to create custom AI models faster and cost effectively.
Moving ahead, the team plans to embed L-Mul algorithms at the hardware level and create easy-to-use programming APIs, aiming to train text-based, symbolic, and multimodal AI models that are perfectly optimized for L-Mul technology.
