news

this article is the 835th original work of shenshen atom

after a long and tortuous process of self-developed intelligent driving, the ideal road to intelligent driving seems to have finally been on the right track.

recentlyideal autointernally determined to establish an "end-to-end autonomous driving" entity organizationthe team has more than 200 people, and other intelligent teams in the company are required to flexibly support the project. this is enough to prove that ideal's transformation is ruthless.

as for the end-to-end solution being the “right track for intelligent driving”, it is not because of the industry giants.teslaother new energy vehicle companies are all developing it first, but this is indeed the most reliable approach among all current technical routes.

and most importantly,"end-to-end" has become the technology that can break through the ceiling most among all current technical routes.。

the old intelligent driving solution "rule control" supported by code has almost reached its limit. going forward will only add to the heavy code, which is prone to bugs and will affect the memory and operating efficiency of the car computer. in addition, everyone is reducing costs now. the costs of laser radar, high-precision maps, sensors and cameras are very high, and end-to-end also has great advantages in terms of cost.

after all, the underlying logic is to train ai to drive and just feed it data.

they are both end-to-end solutions, but their ideals are different.

at the recent "ideal auto smart driving summer conference", ideal auto publicly demonstrated their end-to-end autonomous driving technology architecture.the end-to-end model, vlm visual language model and world model are composed of three parts.li xiang said that ideal auto will officially launch an end-to-end + vlm autonomous driving solution by early next year at the latest.

in fact, end-to-end is just a template concept. generally speaking, it is "data-driven", but the models for obtaining data are very different. for example, tesla uses the bev+transformer model.xiaopengxngp uses a model consisting of the neural network xnet + the large regulatory model xplanner + the large language model xbrain.

but "the fewer models, the more accurate the output."

isn’t that right? isn’t it true that the more models and the more data there are, the more accurate the output will be?

to explain this problem clearly, a lot of professional terms are needed. the "explicit output" between models will lead to the loss of some information. to put it simply, it is like a game of "passing the message". the more people pass the message, the more likely it is to go off topic in the end. the most accurate game is played by two people.

one-model is the route that all smart car companies are pursuing. the ideal goal is to achieve it in the next two years, but it may take another three to five years to fully achieve it and then put it on the car.

in the past three to five years, what everyone has been competing for is the amount of data.

as we said before, the underlying logic of end-to-end is to train ai. how to make ai smarter? the only way is to feed it with data.

in other words, in the past three to five years, what everyone has been competing for is "resources", data resources.

musk once gave a precise description of the amount of training required for an end-to-end model.

"training with 1 million video cases is barely enough, 2 million is slightly better, 3 million will feel wow, and 10 million becomes unbelievable."

ideal aims to develop an intelligent driving solution trained with 10 million video clips by the end of this year.

to obtain data, we must rely on the driving data of our own users, and secondly on training chips.

musk went on a buying spree of nvidia h100 ai chips in april, increasing the purchase volume from 35,000 to 85,000.

ideal, which realized the situation later, has recently started purchasing chips from nvidia at high prices.

in addition to data, computing power is also an important factor affecting intelligent driving capabilities. at present, the computing power of the car computer chip is the ceiling of the orin x chip. if you want to break through this ceiling and further reduce costs, you have to compete with the cloud computing power of the intelligent driving computing center.

in terms of computing power, huawei is still in the first echelon in china, with a computing power of 3.5eflops, while xpeng is only 0.6eflops.nioa slightly better performance is 1.4 eflops, and the ideal performance is 2.4 eflops.

in the "li xiaowei" camp, ideal has already approached the top level in the industry.

from various perspectives, ideal has already made preparations for this "end-to-end war" by recruiting troops and preparing sufficient food and supplies.

it is even more prepared than ever for intelligent driving, because it is urgent for ideal to catch up with the end-to-end model.

ideal is a latecomer to both "noa with map" and "noa without map". noa without map was only officially launched to ad max users in july this year. at this critical moment when tesla fsd is about to land in china, if we don't develop an end-to-end model, we will be left behind by the industry again.

by increasing investment and building a team, it can be said that ideal auto has placed the importance of end-to-end models at the highest level in the entire company. its goal is to put its intelligent driving route on the so-called "right track."

the old traditional code solution that can be "rule-controlled" was once the "right track", and end-to-end is just the "most promising" solution at present.

no one knows what new technology routes will emerge next.investing heavily in end-to-end solutions may be futile, but no car company can afford to miss this opportunity.。

ideal has made all the preparations, but it is just one of them, just following the trend. it is too early to talk about being on the "right track".