MoS2 signal-folding neuromorphic chip slashes AI compute power by up to 90%

Artificial intelligence (AI) demands substantial computing power, driving continuous innovation for more efficient hardware solutions. Neuromorphic chips, designed to mimic the human brain's neural networks, offer a promising path to lower energy consumption for AI tasks.

These specialized systems integrate computation directly into memory, a concept known as compute-in-memory. They often utilize advanced materials like two-dimensional molybdenum disulfide (MoS2) to build their core components.

A new signal-folding technique for MoS2 neuromorphic chips now significantly reduces the power consumption required for vector–matrix multiplication, a fundamental mathematical operation in artificial intelligence.

This innovation cuts power use by up to 90% while maintaining high processing accuracy. Crucially, it also eliminates the need for calibration, a common and often power-intensive step in traditional chip optimization.

The method employs two distinct folding mechanisms. Input signal folding lowers the chip's operating voltage, directly contributing to energy savings during computation.

Weight conductance folding addresses device variations inherent in the hardware itself. This technique enhances the precision of stored 'weights,' which are the crucial data points AI models use to learn and make decisions.

This advancement tackles a central challenge in neuromorphic hardware: balancing high data precision with energy efficiency. Historically, improving one aspect often compromised the other, or required energy-intensive calibration schemes.

By achieving both simultaneously and removing calibration needs, this new method supports scaling up neuromorphic hardware for wider application.

The development could lead to AI systems with significantly lower energy footprints. This enables broader deployment in 'edge artificial intelligence' platforms, such as smart devices or local sensors, where power resources are often limited.

Future developments will focus on integrating this signal-folding technique into wider AI architectures. Researchers will next explore its impact on the energy demands of complex, real-world AI applications.