In the case of “software-defined cars”, what should be the core competencies of car companies?

Recently, a news shook the industry. Hu Chengchen, chief engineer and laboratory director of Xilinx’s Asia-Pacific laboratory, confirmed to join NIO as chief expert and assistant vice president in the field of technical planning.

Behind this news is the acceleration of NIO’s self-developed AI chips. The purpose of NIO is very clear, just like Tesla, to establish a closed loop of autonomous driving capabilities. To establish this closed loop, the chip is a hurdle that cannot be bypassed.

In fact, since last year, we have found that the field of smart chips (AI chips) has once again entered a bright moment. Except for NVIDIA, which has a monopoly in the cloud, both Horizon and Black Sesame have attracted great attention in the industry. The continuous improvement of computing power has also made major car companies anxious about entering the “arms race” of computing power.

For example, this year Nvidia released the industry’s first 1000TOPS SoC, which is more than an order of magnitude higher than Tesla’s FSD single-chip computing power of 72TOPS. In China, there are also Horizon Journey 5, which has a maximum computing power of 128TOPS, and the Black Sesame A1000Pro, which has a computing power of 106TOPS and so on.

However, one of our questions is, is it really that important to pursue TOPS computing power? Is it possible to achieve the goal with the computing power of stacked chips? The industry seems to have entered the misunderstanding of “only computing power theory”. So, here is a brief discussion.


Computing power vs software

Yu Kai, founder and CEO of Horizon, made an analogy, “If the power battery is the heart of the future car, then the smart chip is the brain of the future car.” As the core of the future in-vehicle computing center, the AI ​​chip is of course very important.

At present, the structural form of these automotive main control chips is from MCU to SoC heterogeneous chip (ASIC structure). According to the forecast of Guanyan Tianxia, ​​the global market size of AI chips (inference) in autonomous vehicles will increase from US$142 million in 2017, with an average annual growth rate of 135% to US$10.2 billion in 2022, far exceeding AI chips. The market size of (mobile phone side) is 3.4 billion US dollars.

The market share of AI chips/built-in units deployed at the edge (like Horizon) will also increase from 21% in 2017 to 47% in 2022. Its average annual growth rate of 123% exceeds the 75% average annual growth rate of AI chips deployed in the cloud. The GPU (Graphics Processing Unit) market share will drop from 70% in 2017 to 39% in 2022.

However, in the case of “software-defined cars”, what should be the core competencies of car companies? This is a question that the industry is thinking about. Is it only the chip computing power that is ahead of the curve?

Not really. Still need to look at it dialectically. We say that “data is the means of production”, and the chips that provide data processing are tools. It is impossible for tools to be customer-oriented and become the core. Tools are essential, but the more important core is the software that runs on it. With the rapid increase in computing power of various chip companies, this problem will soon become a non-issue.

In addition, do car companies need such high-end computing power immediately on the consumer side? Not necessarily. At present, car companies claim that the computing power of 8-core chips is much stronger, but is the car system really smooth and easy to use?

We know that in recent years, the trend of “software and hardware decoupling” and “software and hardware integration” have been mentioned. In fact, software and hardware have never really been separated, they have always worked together.

Take the WinTel Alliance in the PC era as an example. Under the WinTel architecture, Intel chips and Windows operating systems are highly synergistic, and can ultimately achieve the effect of monopolizing market share. Both are indispensable.

Therefore, Yu Kai, the founder and CEO of Horizon, put it very well. The chip is the stage of the software. The standard for measuring the quality of the chip depends on whether the software on the chip can maximize its effect. Of course, it does not mean that computing power is not important, and there needs to be an effective match between computing power and software. Comparing two chips with the same computing power, the chip that can make the software run more efficiently is the “good chip”.

Moreover, as a car company, there is also a cost issue of chips. One of the current trends is “L4 hardware + L2 software”, the hardware is “pre-buried” to meet or exceed the standard, and the software is slowly accumulated. But on the other hand, is this a waste? I’m afraid, it is still necessary to “use every TOPS carefully.”

For example, Liang Shuang, co-founder and chief technology officer of Chaoxing Future, said in a recent forum that the arms race of computing power has already started, but the computing power of chips is essentially a necessary and insufficient condition for intelligent driving systems. , “Now everyone is talking about peak computing power. We often see a poorly optimized chip claiming 10TOPS computing power, but the actual application is equivalent to only 3-4TOPS computing power.”

In the final analysis, it is necessary for the AI ​​algorithm to run smoothly on the chip. In the end, this becomes a very complex problem that requires system optimization design.


The “shroud” of computing power

As the master of the modern technology industry and the chip of the “infrastructure” of the digital economy, it brings together the most complex, cutting-edge, and sophisticated basic technologies, as well as high-end talents and funds, which will undoubtedly be the focus of future competition.

However, due to the increasing complexity of chip manufacturing, tens of billions of dollars are often invested in each generation of chip manufacturing. We can see that chip manufacturing is gradually concentrated in a few companies such as TSMC and Samsung. Correspondingly, many established chip companies have given up manufacturing and focused on design.

Therefore, the innovation ability of chip design companies has become more important. With it, there are disputes between AI chip companies and AI algorithm companies. However, for a TOP-level chip company like Nvidia, there are actually more software engineers than hardware engineers. In other words, the underlying technology of a chip company contains both hardware and software.

Moreover, we say that the chip ultimately serves the in-vehicle computing platform of the car company. Therefore, a question that the industry needs to think about is whether to solve the problem of supporting the computing platform of the intelligent driving system, can it only be achieved by stacking the computing power of chips?

The answer is obviously no. Although car intelligence requires stronger computing power, industry experts also said: “The computing power cannot be said to grow infinitely, and the chip PPA (power consumption, cost and area) is very fatal.”

This is because, for automotive AI chips, the computing power index is important, and the energy efficiency ratio is more important. In the traditional chip industry, PPA is the most classic performance measure. However, due to the pursuit of computing power in autonomous driving, the industry still regards “peak computing power” as the main indicator for measuring AI chips, which leads to a biased “only computing power”.

In this regard, Horizon proposed a new method MAPS (Mean Accuracy-guaranteed Processing Speed, the average processing speed within the guaranteed accuracy range) to evaluate the real AI performance of the chip. In the absence of a unified evaluation standard in the industry, it can only be regarded as a single word.

However, the Horizon still has a huge advantage in terms of power consumption. Let’s take Horizon’s first commercial mass-produced Journey 2 chip in 2020 as an example. It is equipped with the self-developed computing architecture BPU2.0 (Brain Processing Unit), which can provide an equivalent computing power of more than 4TOPS, and the typical power consumption is only 2 watts. Moreover, the AI ​​capability output of each TOPS can reach more than 10 times that of a GPU with the same computing power.

For car companies, in the highest performance mode, if the power consumption level of the chip of the automatic driving controller is high, even if its own performance is strong, it will also cause some unpredictable hidden dangers, such as the calorific value increases exponentially, the consumption The electricity rate has increased exponentially, and these results are undoubtedly a “thunder” for smart electric vehicles. Therefore, car companies will fully consider their power consumption indicators in the selection of autonomous driving chips.

We say that a large number of edge AI applications (smart electric vehicles are edge applications) in the AIoT era put forward higher requirements for edge intelligent computing. The general conditions of the edge end will be relatively poor, requiring low power consumption. What AI edge computing needs to solve is to provide the best computing power support under the power consumption limit, as well as supporting memory support and connection capabilities.

In other words, car companies are not likely to worry about the problem of car power consumption, but issues such as chip heat dissipation and power consumption must still be considered. According to the analysis of industry insiders, the infinite expansion of chip computing power and hardware pre-embedding will not be the future trend, and the hardware also needs to match the actual situation. Computing power to adapt to changes in electrical and Electronic architecture.”

Another possibility is that in the future, more and more car companies may choose the same SoC chip in the selection of main chips in the smart cockpit domain and autonomous driving domain. Save development cycle and cost. For example, many car companies now choose Horizon’s journey chip, which is the best example.

Finally, from the concept of “mother ecology” recently proposed by Dr. Lu Qi, the former president of Baidu, smart cars will be a bigger mother ecology after PCs and smartphones, and also the largest in China’s auto industry and technology industry. Where is the opportunity. Moreover, one of the signs that the technology industry where the chip is located is gradually maturing is the formation of a complete ecology. Due to the competition for the future ecology, chip companies are also required to pay more attention to the matching of computing power and software.

The Links:   A2475 LC260W01-A5K2