University Research Advances Compute-in-Memory Technology

Researchers from the University of Michigan have successfully mapped complex state space models onto compute-in-memory architectures, demonstrating their potential in hardware-software co-design for edge AI applications. The newly developed system utilizes a resistive RAM crossbar array fabricated with tungsten oxide memristors, offering high accuracy while significantly reducing energy consumption. This approach addresses the limitations of traditional convolution and transformer networks in compute-in-memory systems by adapting them to be more conducive for state space models.

In parallel, researchers from the Institute of Science Tokyo and Canon Anelva have engineered an ultra-thin nonvolatile memory device measuring just 30 nanometers thick, which combines platinum electrodes with aluminum scandium nitride. This innovation showcases improved energy efficiency, maintaining high performance even at reduced thickness. Additionally, a collaborative effort between UC San Diego and Rutgers University has resulted in a brain-inspired device capable of processing signals akin to neural communication, indicating promising advancements in neuromorphic computing. These developments reflect a strategic shift towards more energy-efficient technologies that enhance computing capabilities while potentially decreasing reliance on conventional computational architectures.