Engineers at Tohoku University and Aisin have built a 16nm FinFET system-on-chip with 32MB of embedded STT-MRAM, enough non-volatile storage to hold both a compact Linux image and a model’s weights. No DRAM or SSD sits outside it.
The research comes out of Tetsuo Endoh’s group at CIEs, Tohoku’s integrated electronics centre, working alongside Aisin engineers.
The numbers, and what was measured
Power-on to operating-system kernel took 65.7ms, roughly 39 times faster than a conventional layout. Time to the first object detection result was 124.2ms, about 25 times better.
Those figures come from running Tiny YOLOv3 on an evaluation board, not from a production automotive controller. Inference energy was 50.7 mJ per frame, with the team claiming roughly a fiftieth of comparable mobile silicon.
Fitting a model into 32 megabytes
Squeezing the design into on-chip memory required weight sharing and mixed-precision arithmetic. The compute block pairs dual Cortex-A53 cores with a Mali-G31 GPU and a dedicated accelerator, aimed at vehicles and robotics.
For a surround-view camera or a factory robot, the appeal is not raw throughput. It is waking up and answering in a fraction of the time while drawing almost nothing from a battery in standby.
The open questions are automotive grade: temperature range, write endurance, and whether MRAM yield holds at volume. A prototype does not answer any of them.
