A few days ago, Qian Jianfeng, vice president of technology of Aixin Yuanzhi Automotive Division, was invited to give a speech at the sub-forum of the 2024 China Automobile Forum and expounded his thoughts on the development of high-level intelligent drivers from the perspective of chip enterprises.
He believes that the performance of the end-to-end technology route far exceeds the regulatory effect of rule-based, and gives a very clear judgment: the end-to-end of the high-level intelligent driving scheme is the only way.
In the speech, he also analyzed the application of end-to-end technology behind Tesla’s FSD capability improvement, and put forward the stage theory of ADAS1.
0-ADAS4.
0 development.
at present, the industry is at a critical moment of breakthrough to the 4.
0 stage.
End-to-end autopilot can not be achieved without the support of key computing chips, including architectural innovation, core IP breakthroughs and leaps in performance.
As a chip company, Aixin Yuanzhi has launched a series of smart driving chip products adapted to the evolution of the intelligent driving algorithm architecture, and provides it with a rich development tool chain that empowers all partners and car companies.
Aixin Yuanzhi is the fastest and most efficient supplier of domestic smart car chips in mass production, and it is not only the efficient development of car companies, but also the second largest smart car chip manufacturer in China, with shipments of hundreds of thousands of pieces.
customers of car companies have covered new car-building forces, mainstream joint venture car companies and domestic TOP-level brands.
Last November, Tesla released the V12 version of FSD, which was amazing.
Compared with the previous smart driving system, the V12 version has a strong driving generalization ability and anthropomorphic driving style.
Take-over mileage has also increased significantly, see on the right, V12 version compared with the previous version of the takeover mileage, a significant increase, more importantly, it contains a high proportion of urban conditions.
Critical Disengagement-safety / critical takeover,% of CriticalDE-no critical takeover rate, no safety takeover rate, the second chart shows no serious takeover success rate for drivers from point A to point B, and no takeover success rate at all.
As you can see, with the iteration of the version, it is rising steadily.
especially in V12.
3, the proportion of no takeover at all is rising rapidly.
From 47% of the V11 version to 70% of the V12 version, the user experience has also increased significantly.
Therefore, it can be seen that the performance of the end-to-end technology route far exceeds the regulatory effect of rule-based.
In the middle of last year, the Shanghai artificial Intelligence Laboratory published the first best paper on autopilot in the history of CVPR.
This paper mainly introduces the first large model of autopilot, UniAD, which integrates perception and decision.
This algorithm scheme, the traditional rule-based algorithm, has done a comprehensive replacement.
It shows great advantages in the perception, prediction and planning of open data sets.
Tesla FSD V12 version, and this algorithm scheme is also relatively similar, it can be said to be very amazing.
With the popularity of the end-to-end technology paradigm, academic institutions also jointly published DriveVLM at the beginning of this year, further introducing the generative large model module.
This algorithm scheme is also very interesting.
In this system, there are two subsystems, and the main system is an integrated system of perception and decision like UniAD, which acts as the cerebellum.
That is, relatively speaking, quick response, good real-time, try to be unconscious, just like driving a manual car to shift gears.
It can be considered that Tesla FSD V12 version is to replace the role of the human cerebellum.
In addition, there is a large visual language model based on generative AI to act as the brain.
This system may be slow to respond.
In other words, the real-time performance is poor.
However, in the face of very complex scenarios, such as sudden emergencies, complex traffic or unfamiliar roads, drivers need to pay more attention to make thoughtful decisions.
Most of us do not think actively when we drive, and most of us use habits and “subconscious”.
When we encounter special circumstances, the brain begins to think about reasoning and judgment and artificial ways to deal with the problem of long tail.
This is the interesting part of the DriveVLM algorithm.
Finally, in May, under the end-to-end craze, self-driving Wayve raised $1.
05 billion, the largest single financing ever raised by AI in the UK.
As a leading company engaged in the research of embodied intelligence in the field of self-driving, the GAIA and LINGO architectures released by Wayve correspond to the end-to-end architecture of UniAD and the large model architecture of visual language, respectively, which are highly consistent with the academic understanding of the direction of self-driving.
This case, in fact, is to echo the mainstream of the above two algorithm schemes.
With the landing of end-to-end technology, it will also have a certain impact on the current intelligent driving scheme.
First of all, there are many ways to classify the smart driving scheme, such as L2, L2 +, L2.5, L2.9 and so on.
Of course, there are other ways of classification.
Aixin Yuanzhi belongs to the chip manufacturer, so our perspective is partial to the hardware form to describe.
So we divide it according to the load, that is, the number of sensors, especially the number of cameras.
Specifically, there are three types: all-in-one machine, medium-sized domain controller and large-scale domain controller.
However, due to the intervention of end-to-end technology, there are obviously four technical stages.
First of all, in the ADAS1.
0 phase, the solution is actually an all-in-one machine.
In this scheme, only the perceptual algorithm is based on AI.
There are even many traditional CV algorithms in it.
Of course, this is also sufficient for the current all-in-one positioning.
Our chip manufacturers will also make continuous product investment on this kind of solution.
So, in the ADAS2.
0 phase, the solution is actually a medium-sized domain control of 5V to 7V.
In this scheme, the perceptual algorithm is also based on AI.
However, in terms of the specific technical route, the technical route of discrete sensing and cross-camera tracking is gradually converging to the technical route of BEV/Transformer to ensure that better sensing results can be obtained in the case of limited sensors.
The biggest change currently occurs in the large domain control aspect of 11V-13V.
In the past two years, the high-speed and urban NOA, that is, the ADAS3.
0 phase, increased the detection of network occupation and general obstacles, integrated with lidar features, and greatly improved the perception accuracy.
But in terms of regulation and control, rule-based is still the mainstream.
By the end of last year, with the release of Tesla’s V12 version of FSD, that is, the ADAS4.
0 phase, large-scale domain controlThe end-to-end trend is already obvious, and the entire autopilot technology paradigm is also rapidly iteratively upgrading.
Whether it is a modular algorithm scheme similar to UniAD, or an integrated generative AI with the concept of fast and slow systems.
EVLM scheme will have a great impact on the design of the chip.
The autopilot scheme has iterated rapidly in the past few years and has gone through many stages: our view is that from the early stages of 1.
0 to 2.
0, from discrete perception dominated by CNN to BEV perception with transformer structure, the efficiency has been improved.
In stage 3, occupation network is introduced, lidar features are integrated, and perception accuracy is further improved.
At the same time, AI planners gradually replace the traditional rule-based planning scheme in the planning control module, but so far, the transmission of perceived regulation and control is still the defined display interface, that is, the transmission barrier frame, the coordinate location of the lane line, etc.
To the end-to-end period, that is, stage 4.
0, the characteristic feature of the model is transmitted between different modules.
As we mentioned earlier, UniAD is used to transmit between different “former” modules, which maximizes the transmission of information and reduces the loss.
In the end-to-end stage, the introduction of large models of visual language, that is, generative end-to-end models, can better solve complex scenes and truly “explain” autopilot to some extent.
Is the introduction of the fast and slow system the final solution for autopilot? (at least at present, the end-to-end solution has reached a certain degree of consensus among practitioners and researchers.
), the above describes the algorithm characteristics in four different ADAS phases.
Aixin Yuanzhi also iterated different versions of NPU according to different stages to cope with the evolution of technology trends.
Among them, it should be emphasized that in the ADAS2.
0 phase, that is, 5V-7V ‘s smart driving program, we introduced the third-generation NPU architecture that supports BEV/Transformer to further improve the performance of the system.
At the same time, an end-to-end algorithm scheme similar to FSD V12 is provided in the fifth generation NPU.
In the latest generation of NPU, it supports generative AI to increase the brain of the intelligent driving system and improve the ability to understand complex scenes.
From the above explanation, we can know that there are only two key requirements of end-to-end algorithms for smart driving chips: high memory and multi-core computing power.
Here I would like to focus on the fifth generation of our Aixin NPU.
In the architecture design, it adopts a distributed heterogeneous multi-core architecture, and introduces inter-chip interconnection and self-developed instruction set (ISA) to improve the efficiency of computing.
At the same time, for the prediction of high memory requirements for large models, we continue to break through the computing ceiling by implementing near-memory computing.
In short, Aixin’s fifth generation NPU can support large-scale reasoning of end-to-end models on our big computing chips.
High-speed data transmission, Transformer extreme special optimization and efficient deployment.
Note: chip ISA (Instruction Set Architecture) is an abstract model description of a computer.
As a service interface of computer hardware and software, it defines the information needed by a hardware programming engineer (CPU designer), including supported data types, storage systems, registers and their corresponding operations (addressing, reading and writing), instruction sets and instruction set coding.
Aixin Yuanzhi has an efficient, easy-to-use and stable software tool chain that supports both end-to-end and LLM models at the software level.
So on the left is support for ADAS algorithms, mainly based on UniAD, as well as important operators and substructures.
on the right is support for large model algorithms (lightweight version of LLM).
After introducing Aixin’s latest NPU technology and tool chain, we will report to you the latest product signpost of Aixinyuan Speed.
Last year, we quickly hit the ground and mass produced M55 and M76 chips.
In terms of M55, we have mass-produced the all-in-one machine scheme and CMS scheme.
in M76, the related 6V small domain control scheme is also in mass production.
This year, we continue to work on the all-in-one solution, iterating over another M57 chip with built-in MCU and supporting 8 million monocular cameras.
Through this chip, to help our partners to further consolidate product advantages.
What I want to emphasize here is our M77.
In fact, M77 was already under development last year, but due to the sudden change in the technology route of smart driving at the end of last year, with the launch of Tesla FSD V12 version, we saw a huge iterative trend of technology route, so we redesigned the NPU of M77, which also caused the delay of M77 in time.
M55H is the first mass-produced smart car chip in Aixin Yuansu.
since its mass production in July 2022, it has shipped more than hundreds of thousands of pieces.
it ranks second in the domestic ADAS all-in-one market.
It is also the fastest chip for mass production in the industry.
And this year, we have also gained many new customers, including new power car companies, independent brand car companies, joint venture car enterprises and so on, and the pace of mass production is advancing rapidly.
At present, for M55 and M76 chip products, many mass production projects are under development, including electronic rearview mirror system (CMS), integrated travel and parking system and medium-and high-level intelligent driving system.
Finally, make a conclusion: Aixin has always been a company advocating technological innovation, and we also uphold the concept of innovation in the field of research and development of on-board SoC.
As I shared earlier, each generation of chips has made an advance technology layout, which is very consistent with the development trend of the smart driving industry and conforms to the direction of technology evolution, which is also the reason why we can catch up from behind.
In the future, we will continue to promote the iteration of intelligent computing technology.
There are innovative technologies and customers need to use them in order to realize their value.
Thanks to our deep experience in the field of large-scale chip shipments, we have also rapidly formed a scale in smart driving chip shipments, helping to promote the large-scale landing of autopilot.
Of course, this high efficiency also benefits from the support and trust of our customers.
Finally, I would like to reiterate our position that we firmly do Tier 2, uphold the concept of openness, and work with our ecological partners and customers to accelerate the development of China and the global smart car industry.
, return to the first electric network home page >.
