Toyota supplier AISIN shows an AI chip that keeps model weights and its OS in on-chip MRAM
Built with Tohoku University on a 16nm process, the chip boots to kernel in 65.7 milliseconds and runs Tiny YOLOv3; AISIN has not said when or whether it will reach production.
Why it matters
Most edge AI chips fetch model weights from external DRAM, which adds traffic, power use and cost at a time when DRAM prices are high. AISIN's system-on-chip puts weights and the operating system in built-in STT-MRAM, and a mixed 8-bit and 16-bit multiplication scheme cut the model's weight data by 53%, from 17.7MB to 9.36MB. Boot to kernel is 38 times faster than with external memory. The design targets small vision models in cars and robots rather than large language models; watch for a production commitment.
Key facts
- 16nm FinFET SoC with built-in STT-MRAM
- Boot to kernel in 65.7 milliseconds, 38 times faster
- Weight data cut 53%, from 17.7MB to 9.36MB
- Developed with Tohoku University