SK Hynix Develops Memristor AI Chip for Edge Computing

SK Hynix has partnered with TetraMem and the University of Southern California to develop a new memristor-based system-on-chip (SoC) designed to enhance the energy efficiency of neural network inference in edge AI devices.

The chip utilizes memristor technology, which allows for data storage and computation within the same unit. This in-memory computing approach significantly reduces the latency and power consumption typically associated with data movement between processors and memory.

The newly developed SoC integrates a RISC-V embedded processor and ten neural processing units (NPU). It is designed for lightweight models and boasts a theoretical peak performance of approximately 2.54 TOPS, with an energy efficiency rating of up to 21.3 TOPS/W.

One NPU is specifically tailored for deep convolution operations, while the remaining nine NPUs handle pointwise convolution and dense computations, common in AI applications. The researchers also addressed challenges with memristor precision, achieving an end-to-end inference accuracy consistent with 4-bit software models.