Toyota shows off SoC that stores model weights and its OS on integrated memory

What happened

AISIN, a car parts maker that Toyota owns a large share of, has shown a new AI processor. It is a system-on-chip, which means a whole computer packed onto one small piece of silicon. The chip stores its AI model and its operating system in built-in memory, instead of in separate memory chips outside the processor.

Key numbers

• The chip starts up from power-on to kernel startup in 65.7 milliseconds.
• AISIN and Tohoku University say this is 38 times faster than using external memory.
• The AI model's weight data shrank by 53 percent, from 17.7MB to 9.36MB, while keeping 16-bit precision.
• The chip is built on a 16nm FinFET process, with an Arm Cortex-A53 dual-core CPU and an Arm Mali-G31 GPU.
• The chip can run Tiny YOLOv3, a fast object detection algorithm.

Why the chip is different

Most AI chips need separate memory to hold AI models. The processor must keep moving data back and forth to that memory, and this slows things down. AISIN says this back-and-forth makes tasks like image recognition work less well on devices. Memory is also very expensive right now. The new chip uses STT-MRAM, a type of memory built into the chip, to avoid these problems.

Why it matters

Cars, robots and other machines are starting to use AI to see and react to the world. This is often called physical AI. If AI can run quickly on a small chip without extra memory, it may cost less and work better in real machines. Companies like Nvidia and Qualcomm also want to sell chips to carmakers, so this shows that parts makers may try to compete too.

What to watch

• AISIN has not said if or when this chip will go into production.
• Watch whether Toyota or other carmakers start using the chip in their vehicles.
• Watch how the chip compares with chips from Nvidia and Qualcomm in real tests.