Toyota shows off SoC that stores model weights and its OS on integrated memory
Reported by The Register1 source ↓
Toyota-affiliated supplier AISIN and Tohoku University have developed a chip that stores AI model weights and an operating system in on-chip memory. The partners say this reduces reliance on external memory for on-device AI tasks.
Key points
- The chip combines a dual-core Arm Cortex A53 CPU, an Arm MaliG31 GPU and built-in STT-MRAM
- AISIN and Tohoku University report kernel startup in 65.7 milliseconds, 38 times faster than with external memory
- A hybrid multiplication design reduced model weight data by 53%, from 17.7MB to 9.36MB
- The chip ran the Tiny YOLOv3 object-detection algorithm on-device
- AISIN has not said whether or when the chip could enter production
Sources
Details
AISIN, an automotive components manufacturer in which Toyota is the major shareholder, developed the system-on-chip with Tohoku University researchers. The chip is built on a 16nm FinFET process and includes built-in STT-MRAM to store both AI model weights and an operating system.
The partners say the chip reaches kernel startup in 65.7 milliseconds, 38 times faster than a configuration using external memory, and can run the Tiny YOLOv3 object-detection algorithm. They also report that a hybrid design using 8-bit and 16-bit multiplication cut model weight data from 17.7MB to 9.36MB while maintaining 16-bit precision.
AISIN says external-memory traffic can hinder image recognition and other on-device AI uses, and notes that DRAM is expensive. The company already supplies Toyota and other automakers, but has not announced plans or a timeline for producing the chip.
Written from coverage as of
Images
- The Register
Stocks in this story
Change since open: 20 Nov to 21 Nov 2026
Last refresh