The energy efficiency of AI simulation chips is 14 times that of traditional chips, and the efficiency of speech recognition is extremely high.

[CNMO News] In the rapid development of AI technology, a new analog chip is changing our understanding of computing power. In a research report published on August 23, Nature magazine introduced a new artificial intelligence (AI) simulation chip developed by IBM Research Laboratory. The new chip is 14 times more energy efficient than traditional digital computer chips, and its efficiency in speech recognition even exceeds that of general-purpose processors. This breakthrough may solve some of the current bottlenecks in AI development.

14nm Analog AI Chip

With the widespread application of AI technology, the demand for energy and resources is also growing. Especially in the field of speech recognition, although software upgrades have greatly improved the accuracy of automatic transcription, the hardware cannot keep up with the data required to train and run these models due to the large amount of computational data required to be moved between memory and processor. Millions of parameters. To solve this problem, researchers have proposed a solution called "computing in memory" (CiM, or simulated AI).

A simulated AI system avoids the inefficiencies of moving data between memory and the processor by performing operations directly within its own memory. This technology is expected to greatly improve the energy efficiency of AI computing. In this regard, IBM's research team successfully developed a 14-nanometer analog chip that contains 35 million phase-change memory cells and can operate in 34 modules.

The research team conducted tests using two speech recognition software, one is Google's small network and the other is Librispeech's large network. When applied to the larger Librispeech model, the performance of the analog chip exceeded imagination, reaching 12.4 trillion operations per second per watt, and the system performance is estimated to be up to 14 times that of traditional general-purpose processors.

The research team concluded that this study simultaneously verified the performance and efficiency of analog AI technology in both small and large models, and is expected to become a commercially viable alternative to digital systems. This breakthrough discovery not only opens up a new path for the development of AI technology, but may also provide a more efficient and environmentally friendly computing method.

The above is the detailed content of The energy efficiency of AI simulation chips is 14 times that of traditional chips, and the efficiency of speech recognition is extremely high.. For more information, please follow other related articles on the PHP Chinese website!