The Register Accessing external memory puts the brakes on physical AI, so carmaker has found an economical and speedy shortcut
Japanese carmaker -maker Toyota has decided to have a crack at the market for AI processors.
Toyota is the major shareholder in a company called AISIN, a manufacturer of automotive components. Both companies are aware that cars are increasingly reliant on tech like image recognition, and that robotics and other forms of “physical AI” will need parts that overlap with cars.
AISIN thinks existing chips aren’t a good fit for such roles, because they rely on external memory to store AI models and then generate a lot of traffic between a processor and the model. That traffic, the company contends, makes image recognition and other on-device uses of AI less effective. DRAM is also currently vastly expensive.
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. The chip includes built-in STT-MRAM, which stores both model weights and an operating system.
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 – and run the Tiny YOLOv3 object detection algorithm on the chip.
“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]. “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.”
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.
Meanwhile, the likes of Nvidia and Qualcomm very much want to sell their silicon to carmakers, and position themselves as sources of tech and AI smarts that existing automotive supply chain participants can’t match. ®
Get AI news in your inbox
Daily digest of what matters in AI.
Key Terms Explained #
GPU Graphics Processing Unit.
NVIDIA The dominant provider of AI hardware.
Object Detection A computer vision task that identifies and locates objects within an image, drawing bounding boxes around each one.
Weight A numerical value in a neural network that determines the strength of the connection between neurons.