AI researchers have made another breakthrough.

The study, published in the journal Nature Computational Science, suggests a new way of computation that enables large language models (LLMs) to run 100 times faster and 10,000 times more energy-efficiently than they are today.

The point is simple. Existing GPUs have to constantly exchange data between the calculator and memory. It’s like reading a sentence by going back and forth between a bookcase and inserting it back into place. However, the proposed method processes storage and operation simultaneously in the same place. This simple method reduces delay and almost eliminates power consumption.

More surprisingly, we have already implemented GPT-2-level performance on this hardware without re-learning.

If this technology becomes a reality, we can expect a future where models like GPT-5 will soon operate in my hands, offline, and with very little energy than now, not in the data center.

Even if the technology is fast, it is too fast.

출처: https://www.nature.com/articles/s43588-025-00854-1