In a significant stride toward sustainable computing, Fujitsu has announced the development of a new generative AI reconstruction technology designed to create highly energy-efficient AI models. The breakthrough, which is a core component of the Fujitsu Kozuchi AI service, directly addresses the growing environmental impact of large language models (LLMs) and other AI applications, which require immense computational power. This technology promises to make powerful AI more accessible and sustainable by radically reducing the hardware and energy needed for operation.
The new technology is underpinned by two key innovations. The first is a proprietary 1-bit quantization method that significantly compresses the data within an AI model’s neural network. This technique has enabled an impressive 94% reduction in memory consumption for Fujitsu’s Takane LLM, while maintaining a remarkable 89% accuracy retention rate—far surpassing the performance of conventional quantization methods. This efficiency allows large generative AI models that once needed multiple high-end GPUs to run on a single low-end GPU, a major reduction in hardware and power requirements.
The second core innovation is a “specialized AI distillation” process. Inspired by the human brain’s ability to specialize, this method efficiently extracts and condenses only the knowledge required for a specific task from a massive, general-purpose model. The result is a specialized AI model that is not only lightweight and efficient but can also achieve higher accuracy than the original model. This approach moves beyond simple compression to create smarter, more focused AI that consumes less energy.
The implications of this technology are far-reaching. By allowing advanced AI to run on edge devices like smartphones and factory machinery, Fujitsu is enabling a future of improved real-time responsiveness and enhanced data security. The radical reduction in power consumption for AI operations contributes significantly to a more sustainable society. Fujitsu plans to roll out trial environments for this technology and will continue to advance its research, aiming for even greater memory reduction while maintaining accuracy. This positions the company as a leader in a new era of “green AI,” where intelligence and sustainability can advance hand-in-hand.
