Tohoku University and AISIN Build AI Chip With 65.7 Millisecond Startup

Tohoku University and AISIN Corporation developed what they describe as the first 16 nanometer AI system on chip with embedded STT MRAM, according to International Business Times Japan. The chip contains 32 MB of nonvolatile memory and stores operating system data and AI model weights on the chip, removing the need for external DRAM or solid state storage during demonstrated AI processing.

The system reaches operating system startup in 65.7 milliseconds, compared with 2,535.3 milliseconds for the conventional setup. AI processing time fell from 594.6 milliseconds to 58.5 milliseconds, while overall system processing was about 25 times faster.

The chip runs Tiny YOLOv3 object detection using a dual core Arm Cortex-A53 CPU, an Arm Mali-G31 graphics processor, and an AI accelerator. It consumed 50.7 millijoules per frame, one fiftieth of the conventional approach, while mean average precision declined from 0.325 to 0.316. The research is scheduled for presentation at the IEEE Asian Solid State Circuits Conference in Taiwan, which runs from 1 to 4 November.