Vehicle systems need AI ready soon after power-on, and the Aisin and Tohoku University chip treats start-up time as a core metric
Automotive supplier Aisin and Tohoku University have built what they call the first 16nm AI system-on-chip (SoC) with embedded spin-transfer torque magnetic random access memory (STT-MRAM). Holding both the operating system and AI model on-chip, the edge AI chip boots 65.7ms after power-on and runs object detection on a fiftieth of the usual energy.
As AI spreads into vehicles, robots and other edge devices, shuttling model data between processor and external memory slows both processing and start-up. Because STT-MRAM retains data without power, the SoC reads weights and boot data from memory beside the processor. Kernel boot time falls from 2,535.3ms to 65.7ms, 38.59 times faster.
A hybrid 8- and 16-bit multiplication architecture cut Tiny YOLOv3 weight data by 53% to 9.36MB, with detection accuracy slipping only from 0.325 to 0.316. Weight sharing within 4×4 tiles reduces data movement further, bringing AI processing down from 594.6ms to 58.5ms.
Across the full system, the chip runs 25.2 times faster and uses 50.7mJ per frame. Automobiles head the partners’ list of target applications, alongside robots, industrial machinery and sensing devices that need instant start-up and low power draw.
The partners now plan to expand memory capacity and computing performance. Partly funded by Japan’s New Energy and Industrial Technology Development Organization, the work is presented at the IEEE Asian Solid-State Circuits Conference in Hsinchu, Taiwan, from 1 to 4 November 2026.
Source: Aisin
Source: www.automotiveworld.com
