IBM in the United States has recently launched a brain like chip called "Arctic", which runs image recognition algorithms driven by artificial intelligence at a speed 22 times faster than similar commercial chips and has 25 times more energy efficiency than similar chips. The relevant research paper was published in the journal Science on October 19th.
It is reported that inspired by the way the human brain works, the 'Arctic' chip interweaves computing and storage modules, allowing each computing core to easily access remote storage blocks like adjacent storage blocks, significantly improving the speed of information exchange between computing and storage units. IBM had previously manufactured chips known as "True North" based on this idea, but "North Pole" transformed this technology into a digital architecture compatible with silicon chip technology used in contemporary computers. The head of the research team, Dalmendra Modha from IBM Research Laboratory, stated that this is a new approach to looking at computer architecture.
The North Pole chip is composed of 256 computing units, each of which has memory. By eliminating off chip memory, computing is intertwined with on-chip memory. These computing units are connected together in a network inspired by the white matter connections between the human cerebral cortex. The paper mentions that on the ResNet50 benchmark image classification network, compared to GPUs using similar 12 nanometer technology processes, the North Pole achieves an energy metric of 25 times higher per watt of FPS (frames per second), a spatial metric of 5 times higher per transistor FPS, and a delay time metric of 22 times lower
However, the 'Arctic' cannot perform other tasks, such as artificial intelligence training tasks, nor can it easily run larger artificial intelligence models. The research team plans to prove whether multiple 'Arctic' chips can support large language models. Although the prototype of the "Arctic" chip is unlikely to be commercialized immediately, the digital architecture of the "Arctic" chip is extremely innovative, which is crucial to the efficient operation of AI on the computing hardware of autonomous vehicle and aircraft.
The copyright of this article belongs to the original author. The reprint of the article is only for the purpose of disseminating more information. If the author's information is marked incorrectly, please contact us immediately to modify or delete it. Thank you for your attention!