The network rumors that the ideal car lays off 15%, is the life of the new energy vehicle company not good?

This interpretation is quite due to blowing Stine.

Let's start with the rumors of ideal layoffs. First, it's an unconfirmed rumor; second, it's an expected operation, if it's true.

Wei Xiaoli has developed to a new stage, and it is necessary to move from 100,000 vehicles per year to a higher goal. In order to support this goal, the organizational structure must be innovative. The first stage of crazy recruitment, the current timing is just to make an adjustment, but also in anticipation.

Then say That Zhang Jianyong left, how can this news be put together with layoffs?

The NIO ES8 was the first car equipped with MobileyeQ4 chip, and the hardware was the most advanced at the time, but the software has not been done well. Zhang Jianyong was ordered to do two things in the face of danger:

The first is to use the hardware on the ES8 and ES6 vehicles at the fastest speed, with basic auxiliary driving capabilities - so that the early management of slightly chaotic Weilai can deliver the auxiliary driving function at the normal time node. The second is to deliver NIO NOPs a quarter earlier than Xiaopeng NGP with a minimum computing power of 2.2TOPS (Xiaopeng 32TOPS, Tesla 72TOPS) - let the second-generation platform delay the release of 1-2 years, and the weilai platform that is barely supported by the first generation platform has not lagged behind in the competition of advanced assisted driving technology.

Judging from these two things, first of all, Zhang Jianyong is a hero, and the humble gentleman Li Bin is like a person who unloads and kills donkeys? Secondly, under the unfavorable conditions of time and hardware, the delivery of qualified functions is the most lacking talent in the automotive industry, who will be stupid enough to quit him?

Zhang Jianyong resigned because he wanted to join a startup company with an autonomous driving AI chip - not only can not explain that the new car-making forces are not having a good time, but it shows that the investment of the new car-making forces in the smart track is optimistic.

If the end of the new car-making power track is Tesla, then the end of the autopilot AI chip track is Nvidia, and both tracks are wide enough!

Since the track is as wide, it is a reasonable choice to join the AI chip company as a founding member. From the perspective of personal interests, the identity of the founding members has more room for development; from the perspective of big dreams, China lacks excellent car companies and chip companies.

In addition, there is one of the most important reasons: although the AI chip is hardware, it pays great attention to the "combination of soft and hard", and the most needed is algorithm talents and talents who are proficient in application scenarios!

The difference between AI chips

Let's start with someone familiar: Steve Jobs.

When Steve Jobs visited the farm as a child, he was shocked to see calves struggling to start walking a few minutes after birth: "It's not a skill acquired through learning, it's innate", "It's like it's designed, the body and brain work together immediately after birth".

This is probably the earliest enlightenment brought to Jobs by the idea of "combining soft and hard".

Left: Giraffe Right: Calf, learns to walk a few minutes after birth Left: Giraffe Right: Calf, learns to walk a few minutes after birth

Decades later, at the launch of the first generation of the iPhone in 2007, Jobs emphasized the importance of the idea of "hardware and software" in a phrase from Turing Award-winning inventor Alan Kay, who said: "People who really take software seriously should make their own hardware."

Steve Jobs at the first iPhone launch in 2007

At the end of 2020, the M1 chip launched by Apple can achieve both violent performance and ultra-low power consumption, unified memory architecture design and software and hardware collaborative optimization are indispensable, which is the concentrated embodiment of the idea of "combination of soft and hard".

Can you explain the idea of "soft and hard combination" in layman's terms?

It is very simple, that is, to design the software according to the application scenario, and to design the chip according to the software!

As we all know, those who do small videos like to use Macbooks (application scenarios), Apple designed Final Cut Pro (software) for it, and designed M1 chips (hardware) for this software.

To sum it up in one word, it is "specialization." Speaking of specialization, it is easy to understand an example: the CPU is already very strong, why play games with a GPU graphics card? Mainly because in the field of image processing, the general-purpose CPU has more than enough and is insufficient, while the GPU of the "special office" is in hand.

The "combination of soft and hard" in the field of image processing Left: CPU serial operation, slow speed Right: GPU parallel computing, fast speed

The "specialization" from CPU to GPU is: from serial computing to parallel computing. The difference can be seen in the following two figures:

Schematic diagram of how the CPU works

Schematic diagram of how the GPU works

So what is an AI chip? It is to take the "specialization" a step further: the same parallel computation, but the design of specialized hardware for the MAC multiplication accumulation operation of cnn convolutional neural networks.

CNN calculation features: "multiply and accumulate" MAC operation

So it becomes an AI chip:

"Soft and hard combination", can it go further?

To make a summary, I believe everyone understands:

CPU: General-purpose serial computing. GPU: Parallel computation of a general-purpose algorithm. AI chip: Parallel computation of specialized algorithms.

So, can the specialization of the combination of soft and hard go one step closer? Yes, and it can go two steps further.

The first step is to need an algorithm master

Among traditional chip companies, the most important thing is the hardware master. Their job is to perpetuate Moore's Law: doubling in 18 months.

Moore's Law has been very fast, but it still can't keep up with the growth in the demand for computing power by AI algorithms: 7 orders of magnitude in 7 years, 10 times in 12 months.

Algorithm engineer you are running too fast, please wait for your hardware engineer!

At this time, someone found a problem: although the computing power of Tesla FSD is only 3 times that of the Nvidia Drive PX2, the actual performance is 7 times!

The reason is simple: Nvidia is still biased towards hardware companies in general, designing good chips for car companies to adapt to algorithms; Tesla has long seen through this and decided to make its own chips - first think about what algorithm to use, and then optimize the hardware for the algorithm.

It's not easy! Because of this idea of "developing hardware according to algorithms", chip development takes 2-3 years, and AI algorithms are developing rapidly. If there is no AI algorithm expert, it is likely that the trend prediction is wrong, and the chip is obsolete before it is mass-produced!

Therefore, AI chip companies are best dominated by software masters, at least in today's fast-evolving algorithms.

China already has one such company: Horizon.

Horizon's founder, Yu Kai, is from Baidu and is entirely a talent with a software background, but may be one of the first people in China to recognize that the combination of soft and hard is the only way for AI chips:

Horizon Quest 5 hasner capacity lower than Nvidia Orin, but has higher actual efficiency. The reason is simple: Nvidia's orin is certainly optimized for the algorithm, but the degree of combination of soft and hard is not as high as the horizon.

Let's look at the second set of comparisons: Horizon Journey 5, NVIDIA Xavier, and Orin

From the TOPS point of view, Journey 5 is about half of Orin's. From MAPS, the frame rate of Journey 5 is slightly better than Orin. Considering the MAPS performance per watt, Journey 5 is far superior.

Since Yu Kai is an algorithm master who makes chips, zhang Jianyong is a graduate of the Automotive Engineering Department, and he is not from an algorithm background, can he have any advantages in making chips?

This brings us to the second step: a master who is proficient in the application scenario is more needed

From general-purpose computing to AI computing, the degree of specialization has increased a lot, but it is not enough!

Voice interaction, autonomous driving, and face recognition in the car all need to use AI algorithms, but different AI algorithms are needed!

Autonomous driving is also a very broad concept, which can be further refined into dozens of sub-functions, each with specific performance goals. Let me give you an example:

While a typical AI chip needs to obtain a complete image before starting processing, the Horizon Quest 5 chip selects the keyframe of the input frame to implement instant data processing. In addition, data processing is prioritized for scheduling critical tasks through time slicing, which can significantly reduce latency by tens of milliseconds or even hundreds of milliseconds.

In most application scenarios, saving 100 milliseconds does not make much sense. However, in the application scenario of emergency braking, 100 milliseconds means a braking distance of 1.7-3.3 meters, which may determine the life and death of a life. Based on the measured 5 of The Journey, the 8M single sight perception structured output delay is less than 60 milliseconds, which is far superior to the industry and will greatly improve driving safety.

In this example, "starting from the scene" means that intelligent driving is particularly sensitive to low latency, so when designing the "algorithm", it is necessary to consider the instant processing mode, flexible priority scheduling mechanism, and the "hardware" should also be optimized accordingly with the algorithm.

If you can optimize the algorithm for every performance goal of each function, and then optimize the chip, it will definitely achieve better results! To achieve such a development idea, who is the most needed talent?

It is Zhang Jianyong!

With a computing power of 2.2TOPS to achieve a Weilai NOP function comparable to Tesla NOA and Xiaopeng NGP, with his toes, he can also think of how much suffering Zhang Jianyong has suffered in the chain of "chip computing power - software algorithm - application scenario"!

Finally, Zhang Jianyong started a business to create an autonomous driving chip, and it is also possible to use it on the Weilai car, which is much less risky and win-win than Weilai taking money to make its own chips.

From any point of view, I can't read that the new forces are having a bad time!