Toyota-backed parts manufacturer AISIN has unveiled a system-on-chip that integrates memory to hold both artificial intelligence model weights and an operating system for on-device AI tasks. The chip, built on a 16nm FinFET process with an Arm Cortex A53 dual-core CPU and Arm MaliG31 GPU, can reach kernel startup in 65.7 milliseconds, according to the company and its research partner Tohoku University.
AISIN argues that existing processors are poorly suited to tasks such as in-car image recognition because they rely on external memory, creating heavy traffic between the processor and the model that reduces AI effectiveness. DRAM is also currently expensive. The company's solution is a system-on-chip with built-in STT-MRAM, which stores model weights and an operating system directly on the chip.
Toyota has decided to enter the market for AI processors. Toyota is a major shareholder in AISIN, an established automotive component supplier that already sells to Toyota and other carmakers. Both companies believe that robotics and other forms of "physical AI" will require components that overlap with those used in cars.
To fit the AI model into the limited on-chip memory, a machine translation of Tohoku University's announcement said that "a hybrid multiplication architecture combining 8-bit and 16-bit multiplication was developed." The approach reduced the amount of weight data by 53 percent, from 17.7 MB to 9.36 MB, while maintaining 16-bit precision. The chip can run the Tiny YOLOv3 object detection algorithm.
AISIN has not said if or when the chip can reach production. Meanwhile, Nvidia and Qualcomm want to sell their silicon to carmakers, positioning themselves as sources of technology and AI capabilities that existing automotive supply chain participants cannot match.
Analysis
Why This Matters
- The chip could bring more efficient on-device AI to cars and robotics, reducing reliance on expensive external memory and cloud connections.
- By integrating model weights and OS into on-chip memory, the design aims to solve boot-time and data traffic bottlenecks that limit current AI hardware.
- Toyota's entry into AI processors pressures established chipmakers like Nvidia and Qualcomm for the growing automotive silicon market.
Background
Toyota, through its majority stake in AISIN, has entered the market for AI processors. AISIN is a long-standing automotive components manufacturer. The new system-on-chip, built with research from Tohoku University, is the first product of this effort. It uses STT-MRAM, a type of non-volatile memory, integrated directly with the processor to eliminate external memory accesses for AI tasks.
What to Watch
- When or if AISIN announces production plans for the chip.
- How Nvidia and Qualcomm respond, particularly with their own automotive AI offerings.
- Whether the chip finds adoption in Toyota vehicles or other original equipment manufacturers.
Checked
12 claims in this story checked against 1 source: 10 supported.
Show each claim
- Partly supportedThe headline and opening attribute showing off or unveiling the chip to AISIN.www.theregister.com - Articles: “Toyota shows off SoC that stores model weights and its OS on integrated memory.”
- SupportedThe chip can go from power-on to kernel startup 38 times faster than with external memory.www.theregister.com - Articles: “AISIN and its research partners at Tohoku University say that approach means the chip can go from power-on to kernel startup in 65.7 milliseconds – 38 times faster than is possible if using external memory”
- Partly supportedThe chip reaches kernel startup in 65.7 milliseconds.www.theregister.com - Articles: “the chip can go from power-on to kernel startup in 65.7 milliseconds”
- SupportedThe chip’s model weights and operating system are stored on-chip.www.theregister.com - Articles: “The chip includes built-in STT-MRAM, which stores both model weights and an operating system.”
- SupportedThe chip is built on a 16nm FinFET process and has a dual-core Arm Cortex A53 CPU and an Arm MaliG31 GPU.www.theregister.com - Articles: “The company’s response is a system-on-chip built on a 16nm FinFET process, an Arm Cortex A53 dual-core CPU, plus an Arm MaliG31 GPU.”
- SupportedAISIN says external memory creates traffic between the processor and AI model that makes on-device AI uses less effective.www.theregister.com - Articles: “That traffic, the company contends, makes image recognition and other on-device uses of AI less effective.”
- SupportedThe story quotes Tohoku University’s announcement as saying a hybrid architecture combining 8-bit and 16-bit multiplication was developed.www.theregister.com - Articles: ““To effectively utilize the limited on-chip memory capacity, a hybrid multiplication architecture combining 8-bit and 16-bit multiplication was developed” according to a machine translation of Tohoku University’s announcement [PDF].”
- SupportedThe approach reduced AI model weight data by 53 percent, from 17.7 MB to 9.36 MB, while maintaining 16-bit precision.www.theregister.com - Articles: “By combining 16x16, 16x8, and 8x8 bit multiplication, the amount of weight data for the AI model was reduced by 53 percent, from 17.7MB to 9.36MB, while maintaining 16-bit precision.”
- SupportedThe chip can run the Tiny YOLOv3 object detection algorithm.www.theregister.com - Articles: “and run the Tiny YOLOv3 object detection algorithm on the chip.”
- SupportedToyota has decided to enter the AI processor market.www.theregister.com - Articles: “Toyota has decided to have a crack at the market for AI processors.”
- SupportedToyota is a major shareholder in AISIN.www.theregister.com - Articles: “Toyota is the major shareholder in a company called AISIN”
- SupportedAISIN has not said whether or when the chip can go into production.www.theregister.com - Articles: “AISIN is an established parts supplier and already sells to Toyota and other automakers but hasn’t said if or when it can get this chip into production.”