Aisin develops 16nm AI SoC with integrated STT-MRAM technology
Aisin, in partnership with Tohoku University, has unveiled a 16nm AI processor featuring STT-MRAM, achieving significant energy efficiency gains in object detection tasks.
Automotive parts supplier Aisin has announced the development of a 16nm AI system-on-a-chip (SoC) incorporating spin-transfer torque magnetic RAM (STT-MRAM). The project, conducted in collaboration with Tohoku University, seeks to optimize processing speeds and power consumption for localized AI tasks.
The chip utilizes a near-memory computing (NMC) design, featuring 32MB of embedded MRAM. By storing AI model weights and system startup data directly on the chip, the unit avoids the latency associated with external DRAM or SSD storage during boot sequences. Testing indicated a startup time of 65.7ms.
In performance metrics using the Tiny YOLOv3 model, the chip demonstrated a 50-fold increase in energy efficiency for object detection compared to traditional setups, consuming 50.7mJ per frame. The architecture relies on an 8/16-bit mixed multiplication system to handle data efficiently within the constraints of the on-chip memory.
The development highlights an ongoing industry trend toward integrating high-speed, non-volatile memory directly into AI processors to reduce the energy demands of edge computing.