"Nature" magazine published the results of IBM's "AI simulation chip", which can reach 14 times the performance of traditional chips

IT Home News on August 24, the journal Nature published the latest research results of IBM Research Laboratory on August 23: an AI simulation chip that is 14 times more energy efficient than traditional digital computer chips.

It is said that the efficiency of this chip in speech recognition exceeds that of general-purpose processors, and is expected to break through the bottleneck of current AI development caused by insufficient computing power and low efficiency.

The abstract of the article states that current AI models with billions of parameters can achieve high accuracy in a range of tasks, but also highlight the performance of traditional general-purpose processors (including graphics processors, central processing units, etc.) Low question. To this end, the research team proposed a "simulated memory computing" solution to provide stronger energy efficiency by executing matrix-vector multiplication in parallel on its own memory.

The research team also developed a 14nm analog chip containing 35 million phase-change memory cells in 34 modules. During the test session, the research team used Google voice commands and Librispeech speech recognition to test the efficiency of the chip's language processing capabilities. After using Google voice testing, they found that the chip's performance and accuracy were "comparable to current digital technology";On the larger Librispeech, the chip can achieve 12.4 trillion operations per second per watt, which is up to 14 times the performance of traditional general-purpose processors.

IT Home attaches the original text of "Nature" magazine as follows:

refer to

The above is the detailed content of "Nature" magazine published the results of IBM's "AI simulation chip", which can reach 14 times the performance of traditional chips. For more information, please follow other related articles on the PHP Chinese website!