The Frontline | Zhuoyu Technology Releases Medium-Computing-Power Urban NOA, Bringing End-to-End Intelligent Driving to the 100,000-yuan Class.

Written by Li Anqi

Edited by Li Qin

On December 31, Zhuoyu Technology (DJI Automotive) announced the launch of an end-to-end intelligent driving solution based on medium-level computing power. The hardware configuration of the solution is 7V + 32TOPS, that is, 7 cameras + 1 TDA4 VH chip (with an AI computing power of approximately 32TOPS), which can realize functions such as Urban NOA (Urban Navigation Assisted Driving).

This is created by Zhuoyu Technology to meet the high-level intelligent driving needs of 100,000-yuan-level vehicle models. Zhuoyu Technology stated that automotive enterprise customers who have mass-produced and equipped with its medium-level computing power solution can upgrade the service to users for free through software OTA.

According to official information, the vehicle models currently equipped with Zhuoyu Technology's medium-level computing power include Baojun, Chery iCAR, etc. Zhuoyu predicts that this function can be seen on the first batch of cooperative customer vehicle models as early as the first half of next year.

"End-to-end" is a popular technology in the intelligent driving industry in 2024. Industry leading players such as Huawei, Xiaopeng, and Li Auto have all initiated the "end-to-end" competitive mode. Generally speaking, the mainstream "end-to-end" solutions in the industry are mostly built based on the NVIDIA Orin X chip platform, starting with at least one Orin X chip, and the computing power of this chip is 254TOPS.

In other words, one Orin X is basically the lower limit of the computing power support for the industry's mainstream "end-to-end" solution, while Zhuoyu Technology wants to further lower the "end-to-end" intelligent driving to a computing power platform with only 32 TOPS, lowering the threshold for high-level intelligent driving with urban NOA capabilities.

With the arrival of 2025, many automotive enterprises are accelerating the large-scale popularization of intelligent driving products. To adapt to market changes, intelligent driving suppliers also need more inclusive product solutions.

Currently, Zhuoyu Technology has developed "end-to-end" high-level intelligent driving on two chip platforms. One is the high-computing-power end-to-end intelligent driving solution based on the Qualcomm 8650 chip, with an AI computing power of about 100TOPS, which can realize the map-free urban navigation function. This solution has been mass-produced and implemented on vehicles such as FAW Hongqi E007 and E009.

The other is the medium-computing-power end-to-end intelligent driving solution based on TDA4 VH. In fact, this solution has been in mass production since September last year, but at that time, it only had functions such as high-speed NOA. Subsequently, it has gradually evolved to memory driving and the latest end-to-end high-level intelligent driving solution.

Zhuoyu stated that through the ultimate end-side computing power optimization technology and the joint training and optimization of the model according to the chip characteristics, the end-to-end model deployment can be realized even with limited computing power.

According to the relevant technical personnel of Zhuoyu, the end-to-end algorithm architecture of Zhuoyu's medium-computing-power platform (7V + 32TOPS) is homologous to the model architecture of the high-computing-power platform (100TOPS), using one AI model to integrate multiple modules such as perception, prediction, decision-making, and planning; at the same time, the solution also retains a safety reasoning framework to ensure the trajectory output by the model.

The relevant technical person in charge of Zhuoyu Technology introduced that because the model architectures are homologous, the models of the two different computing power platforms can reuse the large-scale data of mass-produced vehicle models in training. Therefore, the urban navigation function experience of the medium and high computing power platforms is consistent in the vast majority of scenarios.

This also involves the binocular enhanced end-to-end architecture uniquely adopted by Zhuoyu Technology. The relevant technical person in charge of Zhuoyu stated that based on the image + point cloud geometric information provided by the binocular camera, the data training efficiency is increased by 10 - 20 times, and the same performance can be achieved with only about 10% of the training data volume of the conventional monocular / heterogeneous monocular solution, greatly reducing the demand for long-tail data.

Zhuoyu Technology stated that as vehicle models with urban navigation functions will be available in the 100,000-yuan price range, this will accelerate the penetration rate of high-level intelligent driving.