The recent publication in the journal Nature unveils groundbreaking research from IBM's labs: an AI analog chip that boasts an energy efficiency 14 times greater than traditional digital computer chips.
Reportedly, this chip significantly outperforms general-purpose processors in the area of speech recognition. It promises to break the current bottlenecks faced in AI development, which stem from a lack of computational power and efficiency.
The paper's abstract points out that contemporary AI models, containing billions of parameters, achieve high accuracy across a range of tasks but also highlight the inefficiency of traditional general-purpose processors (GPUs and CPUs). To address this, the research team introduced the concept of "analog storage computation." By parallelly executing matrix-vector multiplications directly within its storage, the chip delivers significantly enhanced energy efficiency.
Further, the team developed a 14nm analog chip comprising 34 modules that contain 35 million phase-change storage units. In a testing phase, the chip's language processing capabilities were benchmarked using Google's voice commands and the Librispeech speech recognition suite. Results indicated that the chip's performance and accuracy are "comparable to current digital technologies." More impressively, in larger-scale testing with Librispeech, the chip achieved a computational efficiency of 12.4 trillion operations per second per watt, up to 14 times greater than traditional general-purpose processors.
This groundbreaking technology could revolutionize the fields of AI and machine learning, offering an efficient and specialized alternative to conventional computing hardware.
