VS919, the single-chip smart driving chip for Weijing Yuxing

VS919, the single-chip smart driving chip for Weijing Yuxing

Automotive-grade chips

Unique advantages:

The VS919 series chips officially released by Weijing Technology accurately enter the high-cost-effective L2+/L2++ single-chip integrated market. Main advantages of the product:

First of all, the balanced architecture, energy-efficient computing, high-quality images and low power consumption support a variety of in-vehicle application scenarios.

Taking into account that the current intelligent driving application scenarios are becoming more and more complex, the types and quantities of data collected by diversified sensors are increasing, and the requirements for computing resources are becoming higher and higher. In this regard, Weijing Technology has adopted a balanced architecture design. VS919 has built-in multiple different computing units, including CPU, GPU, DSP, and Weijing's self-developed AI processor - Weijing Tianquan NPU, which can fully meet the diverse data processing needs of the intelligent driving system.

It supports most of the common operators in the industry, including mainstream models such as BEV+Transformer, optimizes Transpose-conv and DW-conv convolution operations, supports data formats of different precisions such as INT8/INT16/FP16, and calculates and processes multiple image formats such as YUV/RGB/RAW.

In addition, in order to significantly improve computing efficiency, the Jingtianquan NPU also adopts an advanced near-memory computing architecture. By built-in 4MB SRAM and custom dedicated storage interfaces, it achieves high-bandwidth, multi-parallel access to on-chip memory, effectively reducing the latency and power consumption caused by frequent data exchange between computing and storage units. It is worth mentioning that Tesla's D1 chip also uses a near-memory computing architecture to achieve excellent energy efficiency. In addition, major manufacturers such as Alibaba Damo Academy and Samsung have also carried out similar research and development layouts.

Through actual performance comparison, under the same conditions, the computing efficiency and power consumption of Weijing Tianquan NPU are far superior to those of competing products. With extremely high energy efficiency, it can effectively cope with the stringent requirements on chip computing power and power consumption in actual application scenarios.

On the other hand, Weijing Technology's fully self-developed Weijing Yaoguang ISP supports the automotive field's needs for high dynamics (140dB), global noise reduction, high-light suppression, dehazing, etc., and can fully meet the image restoration under low light, strong light and various weather conditions, greatly reducing the difficulty of machine reasoning, and improving the overall perception accuracy of the system, fully meeting the increasingly high requirements of automotive vision solutions for image quality.

The Weijing Yuxing series chips also have built-in multi-level low-power management. The chip design power consumption is controlled within 10W. Through natural heat dissipation, the chip cooling needs can be met without water cooling, air cooling and other equipment. It is not only more convenient to install on the vehicle, greatly reducing the engineering implementation threshold, but also because there is no need for external cooling devices, it helps vehicle manufacturers to further reduce costs.

Second, the independent safety island design replaces the external MCU and can truly support single-chip integrated driving and parking.

Many existing smart driving chips have limited built-in MCU storage space and are relatively lacking in functional safety, real-time and reliability. Although many smart driving technology companies have announced the launch of single-chip integrated driving and parking solutions this year, they must actually be equipped with an additional MCU, which does not truly meet the single-chip integrated driving and parking requirements.

To address this industry pain point, VS919 integrates an independent safety island internally, supports ASIL-D level functional safety through dual-core lock-step, successfully replaces the external MCU, and provides an overall AutoSAR solution on the safety island, helping vehicle manufacturers and Tier 1 intelligent driving to effectively reduce development cycles and costs, and accelerate the rapid large-scale implementation of integrated driving and parking.

At the same time, for the single-chip driving and parking solution, the Jingyuxing VS919 series chip has achieved a highly integrated product design. In addition to the safety island, it also has a built-in HSM module, supports various encryption algorithms such as national encryption, and can fully meet the high-level information security requirements and standards at home and abroad. In addition, by integrating a variety of peripheral interfaces such as dedicated vehicle Ethernet interfaces, CAN-FD, PCIE and other high-speed interfaces, it can meet complex applications in multiple scenarios.

Third, Weijing's self-developed Star Chart tool chain redefines the chip usability paradigm

In order to improve the development efficiency of users based on the Weijing chip platform, Weijing Technology provides a complete set of tool chains - Weijing Star Chart, which can provide developers with all-round technical support.

The Jingxingtu toolchain supports PTQ (post-training quantization), a variety of quantization algorithms, and automatic mixed-precision quantization; at the same time, it also supports QAT (quantization-aware training) to meet the accuracy requirements of different scenarios. In addition, it imports training framework models such as PyTorch, TensorFlow, and Paddle-Paddle through ONNX, and can access mature AI reasoning ecosystems such as TVM, OpenVX, NNAPI, and Paddle-Lite.

Weijingxingtu uses advanced compilation algorithms to perform tiling and efficient scheduling of models to fully utilize on-chip resources, improve MAC utilization, give full play to the computing power of NPU, and significantly reduce memory access bandwidth. At the same time, it can support direct interaction between NPU and DSP, effectively improve the efficiency of heterogeneous computing and reduce model reasoning delay jitter. In addition, Weijingxingtu tool chain also provides NPU performance simulator and precision simulator under x86, which greatly improves the deployment efficiency of algorithm models.

Application scenarios:

Can be widely used in driving scenarios such as L2/L2+/L2++ driving, parking, and driving-parking-in-one

Future prospects:

In the application field of L2+ integrated driving and parking, there is no cost-effective chip that truly meets the application requirements. The solutions available on the market are: one is to use a single-chip solution for L2 low-level intelligent driving. Due to insufficient chip computing performance, this solution can only realize time-sharing multiplexing of driving and parking functions, resulting in poor user experience; one is to use a cascaded solution of multiple chips for L2 low-level intelligent driving. This solution is relatively complex in terms of software and hardware implementation. At the same time, multiple chips will also increase the BOM cost of peripheral devices, and the cost-effectiveness is low; another is to use a chip solution for high-level intelligent driving. Due to the overall performance redundancy of the chip and the high power consumption, this solution often requires additional air cooling and water cooling devices for heat dissipation. The system cost is high and it is not conducive to mass production.

Based on the above industry pain points, Weijing Technology has launched a true single-chip mobile and parking integrated chip, providing a highly integrated, high-performance, and low-power solution. It integrates a safety island internally and provides an overall Autosar solution, saving customers' MCU development cycle and costs. The single-chip solution effectively integrates different customer teams, effectively reduces customer product costs and development costs, and enables rapid large-scale implementation of mobile and parking integrated.