The current "controversial" topics include the current hardware arms race, the commercialization thinking under the Robotaxi boom, and the dispute over the route under the current "Hundred Boats" of many car companies....
Based on this, we gathered with Xiaopeng Automobile Huang Xin and many netizens of the new travel community to launch a "new travel open class" on intelligence.
Focusing on the discussion of the hardware arms race, Xiaopeng's views on the lidar route, and the difficulties of urban NGP development, the arguments are fierce and fruitful.
Now,Let's witness it ！
<h1 class="pgc-h-arrow-right" data-track="5" >. </h1>
In the new market, we have announced that more than 1000 TOPS computing chips are about to be launched, including new forces and traditional car companies.
How to view the current increasingly fierce "hardware arms race", Huang Xin mentioned: The hardware arms race has just begun, and the second and third rounds of the arms race will be more intense.
On the topic of "weight is not important", Huang Xin talked about the current hardware competition from the two points of "computing power" and "software and hardware".
<h1 class="pgc-h-arrow-right" data-track="9" >1, "hash rate"</h1>
"For autonomous driving, vehicles need a good computing power platform."
Huang Xin first "confirmed" the importance of the computing power platform.
In terms of "the computing power of the platform", Huang Xin believes that "the difference between the hash rate of 200 TOPS and the hash rate of 800 TOPS will not be too large, but the gap between 200 TOPS and 2.5 TOPS will be more obvious." ”
Huang Xin believes that the computing power is not simply the accumulation of numbers, and it is the most difficult to break through the first level at a lower hash rate, and in fact, most of the problems have been solved at the second level. For 800TOPS or more, most of the current hash power has not really used the excess hash rate.
At present, the core of the cost of "software computing power" lies in how to process the scene with the "lowest" computing power value, and at present, most car companies have announced the "soon to get on the car" of 1000TOPS computing power, in fact, the allocation of computing power has not yet been substantially applied, and the hardware arms race will obviously lose its meaning.
Huang Xin mentioned at the scene: "Optimize the algorithm on the basis of the current chip".
High computing power does not mean that the ability to process scenes is stronger, and adding new functions does not mean that only the hashing power needs to be increased.
Instead, it is necessary to optimize the chip of the current computing power and truly "utilize" the computing power.
<h1 class="pgc-h-arrow-right" data-track="18" >2, "hardware serves software"</h1>
Huang Xin's views are similar to those of many current engineers.
The hashrate size number is the easiest criterion for "users" to grasp to measure the performance of a chip, but it is not.
For example, for smart driving and cockpit chip developers, the chip is not a general-purpose chip built for the pursuit of high computing power, but to maximize the energy efficiency of the chip and consider how to adapt to the software needs of future customers.
For car companies or L4 unmanned vehicle technology companies that "develop their own routes", it is also particularly important to choose the right chip instead of a high-computing chip, which includes the ability of their own algorithms and the solution of high costs.
The logic of this is that the hardware is attached to the software, the chip is designed around the software system, the algorithm requirements itself, and by understanding the market demand, it is necessary to make a forward-looking layout in the next few years to ensure that the chip will not face elimination after the market is launched.
<h1 class="pgc-h-arrow-right" data-track="24" > second, the core competitiveness of autonomous driving: deal with shortcomings</h1>
"Barrel Effect"
Huang Xin mentioned that automatic driving is a systematic project, and what car companies need to overcome is the "short board" in the barrel effect, and the existence of a "short board" will make the entire system unable to run.
Obviously, the current market's vision is still accustomed to focusing on the more attractive "long board".
This also returns to the problem of "holistic system architecture design" that Xiaopeng insisted on mentioning earlier, the unbalanced relationship between sensors and computing power, and one of the defects or shortcomings will make the system "lose completely".
It is precisely by using the positive "system architecture design" that Xiaopeng can solve China's complex special scenarios, such as gase, cross, etc., becoming its own unique advantage.
<h1 class="pgc-h-arrow-right" data-track="30" >3. What is the commercialization of Robotaxi? </h1>
Huang Xin believes that in the future, L4 unmanned vehicle companies will show polarization.
In the future, there are a small number of head players in this field, of course, more funds and resources need to be invested. Another part of the player will do support to the main engine factory to achieve commercial escape.
The decentralization of Robotaxi will also have a new problem, that is, the "contradictory relationship" between consumer demand and commercial demand.
Huang Xin mentioned that the commercial end of Robotaxi can open up to a hutong in the city, but the consumer side is distributed throughout the country, and the road needs of individuals are different.
<h1 class="pgc-h-arrow-right" data-track="35" > fourth, the emergence of Huawei supporting, is there a non-self-developed, semi-self-developed car company will also have new advantages? </h1>
As an extension of the topic, although Huawei does not do unmanned vehicles, we also see Huawei's supporting views on intelligent driving, and is it possible for car companies to get new advantages based on this?
Huang Xin gave an example of "climbing stairs", which is quite graphic.
The team on the market about self-development leads from the first floor to the second floor, and the subsequent team may use ladders, ropes and other tools to climb to the second floor.
But the difference between the two is "time" and "speed of development".
In the "time dimension", the entry of later car companies is different from the time node of the comers, and in the intelligent track of the "minute and second dispute", the latecomers are slower.
In terms of "development speed", the former's self-research ability will bring stronger differences in products in the second half, such as the second floor to the third floor, and this differentiation is constantly increasing with the higher the stairs and even the ceiling.
"Mr. Airplane" tries to summarize as follows: 1 to 2 floors is a 0-1 breakthrough, in this scenario all car companies to bring out the driving system must be enough "stunning", but in fact, from the 2-3 floor and higher floors, it is based on a set of forward system architecture for further development, not just the hardware arms race. With the deepening of self-research and the acceleration of iteration speed, later car companies will obviously fall behind.
<h1 class="pgc-h-arrow-right" data-track="43" >5. What are the dilemmas that Xiaopeng thinks is above the second floor? </h1>
<h1 class="pgc-h-arrow-right" data-track="44" >1, the ability to realize "product value". </h1>
In breaking through the next layer, Huang Xin mentioned that the new challenges are not only based on technical problems that can be solved, but also come from many aspects.
"There are features that no one uses after you launch them."
For example, in the urban scenario, the most important thing is how to really solve the product value capability.
Huang Xin mentioned "garbage time".
In the urban large-scale traffic jam scenario, the "efficiency" at this moment does not play any role, whether it is "assisted driving" or "artificial driving" does not play any role, and at this time, if the "assisted driving" can release time, it is the optimal item, which is also the "value" mentioned by Huang Xin.
Another "value" comes from decision-making. The scene encountered by the city is not like the highway section, for example, the system only needs to capture the information of the next high-speed port in the high-speed, and in this range, NGP only needs to find the right time to change lanes to complete the low-speed action.
Reaching the destination in urban NGP requires encountering many "unpredictable" emergency environments, including lane change lanes becoming smaller, lane change lanes being a right turn or U-turn lane, and traffic jams in lane change lanes.
The logic in the city needs to add how to "avoid obstacles" and "predict", including intersections of urban roads, dedicated lanes, U-turns, etc. This decision is not a "mechanical decision".
Xiaopeng's development logic is to take technology as the "foundation" and to have the ability to think about the value of the product, which directly determines whether the frequency of use of this product is high or not, whether it is good or not.
So the entire urban NGP system is not just about establishing a strong perception, not that adding "lidar" can solve any problem.
<h1 class="pgc-h-arrow-right" data-track="55" >2, encounter a sudden scene. </h1>
Huang Xin mentioned that for thinking about the city, the development logic of the system is based on traffic rules, but in real life, many unexpected scenes such as people who "violate traffic rules" appear.
Although the probability of such scenes is relatively low, the situation is complex, including pedestrians crossing the lane, straight vehicles in the left turn lane, malfunctions and illegal vehicles, especially in China's complex working conditions There is great uncertainty.
<h1 class="pgc-h-arrow-right" data-track="58" >6. How does Xiaopeng solve these problems in urban NGP development? </h1>
<h1 class="pgc-h-arrow-right" data-track="59" >1, urban NGP system development is benchmarked against people. </h1>
The benchmarking here does not mean that a training system for drivers is equipped in the system, but in the development of ideas, Xiaopeng uses the driving behavior, thinking and decision-making ability of "people" as a decision-making reference.
"The driver is like a gold mine, it can give you a lot of interesting things that you can do."
If it is not based on the "people" way of thinking to solve the scene, this function will not be used.
Obviously, this confirms the problem of "product value" expressed by Huang Xin.
"Our product is not to be just suitable for the city of 2 a.m. without a car, if you make such a product is a failure."
<h1 class="pgc-h-arrow-right" data-track="65" >2, "Be sure to learn to split the scene."</h1>
In the face of most of the "sudden scenes" in the city, Huang Xin mentioned that the most important thing is how to split the scene and play the thinking mode of "people".
"When people are driving, they don't have to just think about what your eyes see. You might envision which lane will be faster and how it will change lanes. ”
By splitting the scene, Xiaopeng constantly deduces and splits new data and completes new upgrades to achieve decision-making ability similar to "people".
This also benefits from Xiaopeng's full-link self-developed algorithm. After the system completes the identification, the system will predict whether the object will have an impact on its own driving trajectory, and also determine its own planning and control, including the ability to avoid obstacles for Multiple Scenarios in China.
From identification, fusion algorithms, trajectory prediction, planning and control, this link will make decisions for different scenarios, and if there are shortcomings along one of the roads, it will also lead to the "failure" of the system.
<h1 class="pgc-h-arrow-right" data-track="71" >3, lidar first</h1>
Huang Xin mentioned that lidar is incomparable to being able to compensate for the advantages of data in the short and medium term, but at present, every sensor is improving.
Huang Xin stressed that lidar is indeed very important, but lidar is not "everything is fine" after it is on.
However, in the current complex road conditions in China, it is still very difficult to abandon the pure visual scheme of lidar in the foreseeable short and medium term.
<h1 class="pgc-h-arrow-right" data-track="75" >7, "Mr. Airplane" point of view summary:</h1>
Whether it is Xiaopeng, Ideal or Weilai and other new force car companies, simply from the basic AP to automatic light lane change and other functions, the actual solution is only a single function, based on L2, constantly optimize or change the user's experience, this is the 1-2 floor players.
However, with the climbing of the floor and even reaching the ceiling of the current intelligent driving, with the simple hardware and computing power stacking, in fact, it will encounter an embarrassing situation, that is, the hardware and computing power are not used in practice, but only for the role of a single function, which is almost zero in the second half of the follow-up.
The car companies that take the lead in launching full-stack self-research will play an advantage on the new track.
The Chinese market has begun to usher in a new transformation, and the demand for user personalization is gradually being decentralized, and different technical routes are also "blooming". However, pure vision or lidar, high-precision map route is still a "tangled" dispute, but it is undeniable that there is more than one road to automatic driving, but the decisive way is never the hardware dispute including computing power, but the dispute of optimization algorithms and thinking, and all competitiveness will return to system infrastructure, user experience, and ultimately realize product value.
With the arrival of P5 and the landing of the city NGP, Xiaopeng once again continued to run wild on China's intelligent track.