Fujitsu develops reconstruction technology for Generative AI aiming at power-efficient AI models based on Takane LLM

Fujitsu has announced a new reconstruction technology designed for Generative AI (GenAI) that aims to improve the efficiency of its Takane Large Language Model (LLM). This innovation is particularly relevant for IT decision-makers focusing on energy efficiency and operational effectiveness.

Technology Overview

The reconstruction technology incorporates quantization techniques and specialized Artificial Intelligence (AI) distillation. Fujitsu's 1-bit quantization method achieves a reduction of memory consumption by 94%, while maintaining an accuracy rate of 89% when compared to unquantized models.

Operational Efficiency

This improvement in memory efficiency leads to a three-fold increase in inference speed, enabling large GenAI models to function on a single low-end Graphics Processing Unit (GPU). This efficiency is beneficial for enterprises that rely on AI models for various applications.

Deployment and Future Plans

Fujitsu's specialized AI distillation reduces the size of models while enhancing accuracy by isolating task-specific knowledge. This enables advanced AI deployment on edge devices, which promotes real-time response capabilities and optimizes data security alongside reduced power usage.

In the second half of fiscal year 2025, Fujitsu plans to establish trial environments for the Takane model using this new technology and make models of Cohere's quantized research available through Hugging Face. The initiatives aim to advance GenAI to tackle multifaceted social issues.

Conclusion

This update highlights Fujitsu's focus on developing energy-efficient AI solutions tailored for varied applications. The strategic advancements in AI technology showcase Fujitsu's intent to address relevant societal challenges through enhanced computational capabilities. This Blog Signals brief reflects a timely, fact-based summary of the original blog post.