Kneron Launches Automotive-Grade Chip for ADAS and AV

Kneron Unveils Automotive-Grade Chip, Pioneering Advances in Vehicle Autonomy

Kneron has announced the launch of its inaugural automotive-grade chip, a development supported by investment from prominent firms including Foxconn, Alibaba, Sequoia, Horizons Ventures, and Qualcomm. This new chip is poised to significantly alter the trajectory of autonomous vehicle technology.

Challenging Industry Leaders

The semiconductor industry has traditionally been led by established companies such as Intel, AMD, and Qualcomm. However, Kneron, having secured over $100 million in funding, has rapidly gained recognition as a noteworthy startup. Its technology is currently utilized by Otus, a provider of advanced driver-assistance systems (ADAS) for vehicles from Toyota, Nissan, and Mitsubishi.

With the introduction of the Kneron KL530 chip, the company aims to expand its presence within the automotive sector.

Key Innovations of the KL530

According to CEO Albert Liu, the Kneron KL530 incorporates two crucial advancements. Firstly, it supports Vision Transformers (ViT), a departure from conventional Convolutional Neural Networks (CNNs). Secondly, it features a 4-bit data processor.

These combined features, Liu asserts, lead to enhanced accuracy in image recognition and reduced processing durations.

The Advantage of Vision Transformers

Both ViTs and CNNs are forms of deep learning models. However, implementing transformers in hardware has historically presented challenges. Liu explained, “The transformer represents a fundamentally new architectural design, capable of surpassing the accuracy and performance of CNN chips considerably.”

He further noted the transformer’s superior capabilities in areas like voice and behavior recognition, particularly in scenarios involving time-series analysis.

Holistic Inference with ViTs

Unlike CNNs, ViTs perform inferences by considering the entire scene holistically, rather than focusing solely on individual features. Liu illustrated this with an example: a large white truck against a cloudy sky.

A CNN might mistakenly identify the truck as part of the sky, potentially leading to a collision. A ViT, however, would analyze the complete visual context, recognizing the truck’s defining characteristics like wheels and boundaries.

Enhanced Efficiency with 4-Bit Data Processing

The KL530’s 4-bit data processor offers significant improvements in data processing speed compared to 32-bit and 8-bit processors. It also enables a higher frame rate, reducing processing time by as much as 66% without a corresponding increase in power consumption.

This increased processing capability allows for the detection of more features within a given timeframe, accelerating tasks such as facial recognition by up to half a second.

Prioritizing Automotive Safety

“This translates to greater safety in automotive applications,” Liu emphasized, as faster decision-making is critical in preventing accidents.

Performance and Capabilities

Kneron states that the new chip delivers up to twice the TOPS/watt performance of its previous custom chip, the KL520. The KL530 also incorporates an image system processor, supporting 1080P resolution and enabling features like blind spot detection, object classification, distance measurement, and hazard identification.

Availability and Future Production

The chip will be available in two versions: one for the aftermarket and another designed for direct integration into vehicles. It will be utilized in upcoming vehicles manufactured by Foxconn under the Foxtron brand, in partnership with Taiwan’s Yulon Motor.

Kneron is currently fulfilling sample orders for other clients and anticipates commencing mass production in the first quarter of the following year.