Analog Chip Has Paved the Way for Sustainable AI

Excessive energy usage for AI research was a growing concern. IBM Researchers have developed a device that offers a sustainable AI future.

An analog computer chip can run an AI speech recognition model 14 times more efficiently than traditional chips. According to the researchers, this chip can reduce the challenges thus far faced in AI development.

Since most of the energy utilized by AI models was derived from fossil fuels, the development of analog chip will be groundbreaking. Another cumbersome issue that analog chip resolves is the constant bottlenecks due to the back and forth from memory to processors during operations. The analog computer-in-memory (CiM) chip performs calculations directly within its memory, as demonstrated by IBM.

IBM’s developed analog chip carries 35 million phase-change memory cells set to one of two states. These varied states can represent the synaptic weights between artificial neurons in a neural network. This AI type models how the linking between neurons in a human brain varies when absorbing new information or skill. This was stored as a digital value in computer memory in a traditional chip setting. However, the new CiM can hold and process these weights without performing millions of recall or data storage operations.

The researchers further tested the analog chip on speech recognition tasks, demonstrating an efficiency of 12.4 trillion operations/s/w. Compared to conventional processors, this is over 14 times more efficient.