Analog Chip Breakthrough for Energy-Efficient AI from IBM | Google Generative ai Course | Generative ai Benefits for Business | Generative ai in Finance and Accounting | Turtles AI
#IBM has #developed a new #analog #AI #chip using phase-change #memory that can handle #language tasks with high #speed and #efficiency. #Testing showed comparable #accuracy but greater #energy savings versus #digital #chips. Combining analog designs optimized for AI with digital #hardware could enable more #scalable and #sustainable AI. In a paper published in Nature, researchers from IBM have shown that it is possible to build analog AI chips that can handle natural-language AI tasks with significantly improved energy efficiency compared to traditional digital chips. The idea of using analog chips for AI is not new, but previous research hasn’t shown how such chips could be used for the massive models dominating AI today. The IBM team’s design uses phase-change memory (PCM) to encode the weights of a neural network directly onto the physical chip. Their chip can encode 35 million PCM devices, allowing models with up to 17 million parameters to be run, not yet comparable to cutting-edge AI models but enough to tackle real use cases when multiple chips are combined. By optimizing the multiply-accumulate (MAC) operations that dominate deep learning compute and performing MACs directly in memory, the analog chips can complete tasks faster while using less power than digital chips. Testing on keyword detection and speech transcription tasks showed comparable accuracy but far greater speed and efficiency versus MLPerf benchmark data for digital hardware. While natural language was the focus here, the team is also working on computer vision and investigating larger foundation models. Analog AI faces challenges in precision, manufacturability and integration but the scalability and efficiency demonstrated points to a future where analog and digital chips could be combined in more capable and sustainable AI systems. Highlights: - New IBM analog chip design leverages phase-change memory for energy-efficient AI - Chip tested on speech recognition tasks with accuracy matching digital chips - Estimated to have 14x greater energy efficiency than comparable digital hardware - Enables faster, lower-power AI but faces manufacturability challenges - Combining with digital chips could improve sustainability of AI systems IBM’s new analog chip design marks a significant step towards more sustainable AI hardware. While questions remain about manufacturability at scale, the estimated 14x efficiency gain over digital chips on language tasks highlights the promise of analog computing for AI’s voracious energy appetite. As climate concerns grow, creative hardware innovations like this will be key to counteracting AI’s emissions impact. What do you think of analog chips as a "green" solution for AI? How much efficiency boost is needed to make large-scale AI sustainable? Share your perspectives.