The researcher is holding a 14nm simulated artificial intelligence chip. Image source: Ryan Levine/Nature website
A study published in Nature on the 23rd reported an artificial intelligence (AI) analog chip that is 14 times more energy efficient than traditional digital computer chips. The chip, developed by IBM Research Laboratories, is more efficient than general-purpose processors in speech recognition. This technology may be able to break through the bottlenecks encountered in current AI development due to insufficient computing power and low efficiency.
With the rise of AI technology, the demand for energy and resources has also increased. In the field of speech recognition, software upgrades have greatly improved the accuracy of automatic transcription, but due to the increasing amount of operations moved between memory and processors, the hardware cannot keep up with the millions of operations required to train and run these models. calculated parameters. One solution proposed by researchers is to use "computing in memory" (CiM, or simulated AI) chips. Analog AI systems prevent inefficiencies by performing operations directly within its own memory, whereas digital processors require additional time and energy to move data between memory and processor. Analog AI chips are expected to greatly improve the energy efficiency of AI computing, but practical demonstrations of this have been lacking.
14nm analog AI chip on the detection board. Image source: Ryan Levine/Nature website After rewrite: This image shows a 14nm analog AI chip on a test board. Image source: Rain Levine's "Nature" website
The research team developed a 14-nanometer analog chip, which contains 34 modules of 35 million phase-change memory cells. The research team used two pieces of speech recognition software to test the chip's efficiency in language processing capabilities, the Small Network (Google Voice Commands) and the Big Network (Librispeech Speech Recognition), and compared them naturally with industry standards. A comparison of language processing tasks. The performance and accuracy of the small network are comparable to current digital technologies. For the larger Librispeech model, the chip is capable of 12.4 trillion operations per second per watt, and system performance is estimated to be up to 14 times that of traditional general-purpose processors
The use of 300mm wafers for manufacturing simulated AI chips. Image source: Ryan Levine/Nature website
The research team concluded that this study verified the performance and efficiency of analog artificial intelligence technology in both small and large models, and is expected to become a commercially viable alternative to digital systems
There is no need to change the original meaning. The content that needs to be rewritten is: quoted from Science and Technology Daily
Please follow the public account for more information
The content that needs to be rewritten is: "China Science and Technology Information" magazine
Supervisory unit: China Association for Science and Technology After rewriting: The agency responsible for management is China Association for Science and Technology
The hosting organization: Chinese Association for Science and Technology Journalism
Please contact WeChat ID when reprinting content: zkxxx1999
Online submission platform: www.cnkjxx.com
Please call the submission hotline: 010-68003059
If you need to seek coverage or content cooperation, please add WeChat ID: 15811564659 to contact us
The magazine is included in: "CNKI", "Chinese Journal Core Journal (Selection) Database", "Chinese Academic Journal Comprehensive Evaluation Database (CAJCED) Statistical Source Journal", "Chinese Journal Full-text Database (CJFD)" ” and included in “China Association for Science and Technology, Library Society of China (Recommended Books for Interpreting the Scientific Outlook on Development)”
The above is the detailed content of Breaking through the bottleneck of artificial intelligence development: AI simulation chips are 14 times more energy efficient than traditional chips. For more information, please follow other related articles on the PHP Chinese website!