Researchers at IBM’s Almaden research center in San Jose, California, have unveiled a groundbreaking development in artificial intelligence (AI) with their NorthPole processor chip. This innovative chip, inspired by the human brain, promises to supercharge AI by operating significantly faster while consuming remarkably less power compared to existing architectures.
The breakthrough, which has been published in the journal Science, is set to challenge conventional thinking in computer architecture. NorthPole is designed to run neural networks, complex arrays of computational units used for pattern recognition in data.
While some existing chips can perform these calculations efficiently, they still rely on external memory, known as RAM, for each layer of computation. This data transfer between chips results in a bottleneck, slowing down processing and causing energy inefficiency, commonly referred to as the Von Neumann bottleneck.
NorthPole, consisting of 256 computing units or cores, each embedded with its own memory, mitigates this bottleneck by design. This innovative architecture interconnects the cores in a network inspired by the white-matter connections in the human cerebral cortex, effectively eliminating the Von Neumann bottleneck within each core.
According to Dharmendra Modha, a computer engineer at IBM and a study co-author, NorthPole outperforms existing AI chips by a substantial margin in standard benchmark tests of image recognition and utilizes only one-fifth of the energy, despite not employing the most recent manufacturing processes. The research team estimates that if NorthPole’s design were implemented with the latest manufacturing processes, its efficiency would be 25 times better than current designs.
The NorthPole chip, however, comes with a limitation: its 224 megabytes of RAM may not be sufficient for large language models, such as those used by chatbots like ChatGPT, which require several thousand megabytes of data. Additionally, NorthPole can only run pre-programmed neural networks that need to be trained in advance on a separate machine. Despite these limitations, the authors of the paper suggest that the NorthPole architecture could be especially valuable in speed-critical applications, such as self-driving cars.
This development marks a significant advancement in AI technology, bringing memory units as close as possible to the computing elements within the core. While other researchers have been exploring innovative approaches using new materials and manufacturing processes, such as in-memory calculations using memristors or changing a circuit element’s crystal structure to store information, the NorthPole chip’s groundbreaking design is set to revolutionize the AI field and lead the way to more energy-efficient and faster AI computing.
It remains to be seen how these newer approaches can be scaled up economically, but the possibilities they present are undoubtedly exciting for the future of artificial intelligence.
Reference: https://www.nature.com/articles/d41586-023-03267-0
