The Pulse
Tohoku, Aisin Build 16nm AI Chip That Boots in 65.7ms
Tohoku University and Aisin developed a 16nm AI system-on-chip with 32MB of embedded STT-MRAM, storing both AI model weights and operating-system boot data on the chip. In a Tiny YOLOv3 demonstration, the system booted in 65.7 milliseconds
AI.info Team · Written with AI · fact-checked ·
Tohoku University and Aisin have developed a 16-nanometer AI system-on-chip that boots its operating system in 65.7 milliseconds, with 32MB of embedded spin-transfer torque magnetic random-access memory (STT-MRAM). The university announced the result on November 2, 2026, describing the chip as a near-memory computing system designed to run object detection without external DRAM or SSD storage. It reports energy use of 50.7 millijoules per frame for Tiny YOLOv3 object detection.
32MB of MRAM holds the model and boot data
The chip stores both AI model weights and operating-system boot data in its embedded STT-MRAM. That design removes the need to transfer those data from an external SSD or DRAM when the system starts, according to the university. STT-MRAM retains data without power, a property the announcement says the team used to reduce standby leakage while keeping memory access fast.
The researchers combined near-memory computing with an 8/16-bit hybrid multiplication architecture and weight-sharing techniques. Those methods let the system run Tiny YOLOv3 within the chip's limited 32MB of embedded memory. Tiny YOLOv3 is a smaller version of the YOLO object-detection model, which identifies objects and their locations in images.
The chip uses a 16nm FinFET process and integrates a dual-core Arm Cortex-A53 CPU, an Arm Mali-G31 GPU, an AI accelerator and the STT-MRAM. In the conventional arrangement described by the university, model weights are transferred from flash storage to DRAM before the AI accelerator processes them. The new design instead reads the required weights directly from on-chip STT-MRAM.
To fit the model into the available memory, the team combined 16-by-16, 16-by-8 and 8-by-8-bit multiplications. The university says this reduced the model's weight data from 17.7MB to 9.36MB, a 53% decrease, while the reported mean average precision measurement changed from 0.325 to 0.316. The researchers also used weight sharing within 4-by-4 tiles, reusing the same kernel weights across multiple clock cycles to reduce data movement and local bandwidth requirements.
Boot time and detection speed measured separately
Tohoku reports that the operating system starts 65.7 milliseconds after power-on, a 38.59-fold improvement over the conventional approach used for comparison. The release gives the conventional time to kernel startup as 2,535.3 milliseconds. The announcement separately says AI processing time fell by a factor of 10.16, while the full system achieved a 25.2-fold time reduction. Those are distinct measurements: the 65.7-millisecond figure describes startup through OS boot, not the time to detect objects in a frame.
For Tiny YOLOv3 object detection, processing time fell from 594.6 milliseconds to 58.5 milliseconds, according to the detailed release. In the conventional system, 12MB of model weights had to be transferred from SSD to DRAM before AI processing. The chip's direct access to weights in STT-MRAM avoids that transfer in the demonstration.
The team measured energy consumption at 50.7 mJ per frame for Tiny YOLOv3 object detection. Tohoku says this is one-fiftieth of the conventional system's energy use. The release does not provide the conventional system's absolute energy-per-frame figure, so the comparison is expressed as a ratio.
A-SSCC presentation follows the November 2 announcement
The university said the results would be presented at the IEEE Asian Solid-State Circuits Conference, or A-SSCC 2026, scheduled for November 1–4 in Hsinchu, Taiwan. The paper is titled “A 65.7ms Quick-instant-on and 50.7mJ/frame Tiny YOLOv3-based Object-detection 16nm-AI-SoC with 32MB-embedded-MRAM and Hybrid-multiplication-architecture.” Its listed authors are Tetsuo Endoh, Takanobu Naruse, Tao Li, Shoji Ikeda, Mitsuo Yasuhira and Hiroshi Yamada.
Endoh is identified as a professor and director of Tohoku University's Center for Innovative Integrated Electronic Systems; Naruse and Yamada are listed with Aisin. The university says part of the work resulted from a project commissioned by Japan's New Energy and Industrial Technology Development Organization. The announcement describes possible future applications in vehicles, robots, industrial equipment and sensing devices, where quick startup and low power use are sought. The reported demonstration itself centers on Tiny YOLOv3 object detection and its measured startup, processing-time and energy results.