As artificial intelligence drives an explosion in data processing, interest is growing in integrated next-generation chips such as neuromorphic semiconductors. Alongside neuromorphic designs, the leading examples of chips that combine computing and memory functions on a single device are processing-in-memory (PIM) and compute-in-memory (CIM).
PIM is a next-generation memory that places compute units inside the memory itself, academic sources said on the 16th. The Korea Advanced Institute of Science and Technology (KAIST) is widely regarded as leading domestic research on PIM chips. The PIM Semiconductor Design Research Center run by KAIST aims to build PIM intellectual property and platforms, support PIM process and design technology, and train specialized researchers in the field. It has also worked with major semiconductor companies.
A team led by Professor Kim Dong-jun of KAIST's School of Electrical Engineering developed technology that sharply raises PIM communication speeds, in an international collaboration with Northeastern University and Boston University in the U.S. and the University of Murcia in Spain. Because PIM must exchange signals with an external central processing unit during computation, bottlenecks can occur. To minimize communication between PIM and the CPU, the team implemented a network in which the compute units inside PIM communicate directly with one another without going through the CPU. The approach raised PIM communication speeds 11 times above previous levels. Kim, an authority on PIM, is the first Korean to serve as a fellow in the computer architecture field of the Institute of Electrical and Electronics Engineers (IEEE).
A new design automation technology that processes data directly in memory, rather than sending it to a separate compute unit, has also emerged. Developed last month by a team led by Professor Park Hee-cheon of the Department of Electrical Engineering at Ulsan National Institute of Science and Technology (UNIST), the technology can cut the total number of operations by nearly 20% from previous levels. The team used memristors — semiconductor devices that retain resistance values based on current flow — to arrange circuits so that computation is possible even within a small chip area.
More recently, integrated next-generation chips one step beyond PIM have appeared. CIM is a cutting-edge technology in which memory cells themselves are designed to perform computation as well as their original storage role. The standout performer in this field is U.S. semiconductor startup EnCharge AI. The company's CIM chip, the EN100, can run AI inference and computation instantly on laptops and workstations without relying on external cloud infrastructure or large data centers. EnCharge AI has raised $100 million (about 136.3 billion won) and is said to be working with Taiwan's TSMC based on its design technology. In an interview with foreign media, EnCharge AI Chief Executive Naveen Verma said AI computation involves enormous volumes of data, making the movement of that data a major constraint, and described CIM as one of only a few architectures capable of solving the problem.
Analysts say dedicated software must follow for such integrated next-generation chips to see wide adoption. Nvidia's software platform CUDA played a large part in allowing the company to deploy graphics processing units at scale across the AI infrastructure market.
