British Weida’s “seat anxiety”

The smart driving calculation platform equipped on the North Auto Arcfox Alpha S is Huawei’s MDC Pro 610, which is designed for this car, with a force up to 400+ TOPS. Next, Huawei MDC will combine the universal intelligent driving force demand in the industry, and the underlying capacity of MDC Pro 610 is created into platform-based standardized product MDC 810 for all car companies customers.

How much power does it require L4 automatic driving? There is no clear answer, and the car is not too easy to conclusion, and the caliber of the chip manufacturer has been changing.

A few years ago, some people said that it was 300 TOPS; later, some people said it was 500 TOPS; then, some people said it is 1000 TOPS … Recently released 1000 TOPS’s AI chip atlan from Ying Weida to see, automatic driving The domain controller should soon be “smiled” to 2000 TOPS.

And car companies, whether they really need such algorithms in the next few years, it really needs so much force, but also or passively involved in this “Power Climbing Competition”, selling personnel with a automatic driving chip company. In the words, this is also “internal volume”. But there is no way, in the current large environment, it seems that only a large strength can give the car enterprise enough “safe sense”.

However, at the MDC 810 (GM 400+ TOPS) conference on April 18, Huawei emphasized that automatic driving domain controllers, can’t be placed, but also to see the engineering capabilities behind. It is this ability. Huawei has a bottom to announce “release, mass production” on the date of the press.

Currently, the supplier of 200 TOPS or more high-profile driving chips is only the British Weida, Gao Tong, Huawei, from the currently public information, the latest automatic driving chip is higher, and Ying Weida is also The most mass production order, but the positioning of the two companies is Tier 2. They only sell chips, but lack system integration experience, and the Huawei of Tier 1 has already accumulated powerful machine engineering capabilities.

Of course, under the big trend of hardware and software decoupling, there will be more and more car companies chooses to circulate Tier 1 directly with the chip manufacturer to cooperate, but at this stage, most car companies still do not have their own automatic driving domain controller. Ability – Even with this ability, current opportunities cost is also relatively high, so they still have to entrust a Tier 1 to integrate work.

Therefore, if the chip vendor can provide a domain controller, the workload of the vehicle enterprises will have a lot less.

Comparison and British Weida’s “Size Anxiety”

On September 29, 2016, at the GTC European conference, Ying Weida officially announced the next generation of TEGRA processors, code Xavier. According to the plan, Xavier will be produced based on the TSMC 16N process, integrated up to 7 billion transistors, can reach 20 TOPS at 20 W. According to the initial plan of Yingda, Xavier will be at the TSM in the end of 2017.

Prior to this, according to the disclosure information, the highest power consumption in the automatic driving chip in planning is the highest power consumption 5 W, and the force 12 TOPS. From this, Ying Weida’s Xavier is aiming at the “first” position.

However, in December 2017, Mobileye, which has been acquired by Intel, announced that EYEQ 5 will launch two versions of 12 TOPs and 24 TOPs. Eyeq 5,20 TOPS with 24 TOPS is difficult to keep the “first” seat.

In fact, the British Weida, which is not convinced, has long been “first man”. On October 11, 2017, the Xavier parameter published by Ying Weida is upgraded to: 12nm, 9 billion transistors, 30 TOPS Accordingly, the power consumption is also upgraded to 30 W.

At this conference, Ying Weida also launched a computing platform called “L4-level automatic driving”, Drive Px Pegasus, Drive Px Pegasus, including two Xavier SoCs and two graphic architecture GPUs, energy 320tops. Although this power consumption is more than 500W computing platform is not suitable for mass production, but due to its “king” in that stage, there is still a strong interest in many companies.

Unfortunately, PEGASUS that disdaining everything pursuing the biggest force did not stay in the “first” seat – October 2018, at the General Assembly of Huawei, Huawei launched MDC 600 integration up to 352 TOPS And power consumption (352 W) is lower than pegasus.

In particular, the British Weida is nervous. The main chip used by Huawei MDC 600 (Technical Pre-research) is up to 16 TOPS, but the power consumption is only 8 W, energy efficiency is 2 TOPS / W, far higher than The same is based on the 12 NM process Xavier (energy efficiency 1 TOPS / W).

While being robbed by Huawei to automatically driving, Ying Weida Titla’s cooperative relationship began to end. Tesra officially recognized the self-developing automatic driving chip at the end of 2017, and in August 2018, Tesla said that the self-research chip is about to be completed. On the “Automatic Driving Day” on April 22, 2019, Tesla officially announced the parameters of the self-developed FSD chip: 16 nm, 72 TOPS, 36 W. Whether it is still a power efficiency, it is a larger than Xavier.

It is true that considering the high energy efficiency of ASIC is based on sacrificing a certain flexibility, the power of the ASIC chip and the energy efficiency with the GPU ratio are “unfair” for the GPU core-based Xavier, but in force comparison Under the background, few customers will rationally see this problem, even Ying Weida is also kidnapped.

So, in September 2019, in the case of Xavier, I only won a quantitative customer of Xiaopeng, and Ying Weida still could not wait to release 7 nm automatic driving chip ORIN, which was up to 200 TOPS.

According to the plan, ORIN is going to 2023. Some people guess that Ying Weida is so anxious to launch ORIN, which is “stimulating” by Huawei’s MDC 600 (technical pre-research).

However, ORIN did not sit on the first chair of the rankings – it quickly became a Qualcomm “.”

On the CES in January 2020, Qualcomm also released the automatic driving calculation platform Snapdragon Ride, which can achieve 700 TOPS power consumption at 130 W, and there is news that Cruise and Argo are in use. Qualcomm automatic driving chip is tested.

After CES, many Chinese car companies have begun to contact Qualcomm. According to the informed person, Qualcomm’s automatic driving chip has two two models, 200 TOPS and 300 TOPS, like Ying Did ORIN, SA9000 is also based on the 7nm process, but the energy is up to 5 TOPS / W (Orin initial The version energy efficiency is 3 TOPS / W).

Qualcomm’s chip has strong competitiveness in the mobile phone market and the smart cockpit market, but it is completely a newbie in the automatic driving market. The previous accumulation may not be able to play out – Qualcomm’s strength is CPU, but In the GPU, the capacity accumulation is insufficient, and in the calculation of automatic driving, the weights of GPU account for a relatively large weight.

Xiaopeng Automatic Driving Vice President Wu Xinzhou once served as Qualcomm, automatic driving, and said that Xiaopeng has also negotiated with Qualcomm. Xiaopeng finally did not choose Qualcomm’s Snapdragon Ride or SA9000, but chose to choose Yingda. ORIN.

Since September 2020, ideals, Xiapeng, Wei, SAIC, SAIC, SAIC R brand, etc. have chosen ORIN when doing “hardware pre-buried” programs. It seems that British Weida has achieved stage victory, but in the big background of the hardware arms race, the force is “not enough”, so British Weida still can’t keep it off.

There are a few small episodes:

The original plan is produced in 2023, but in order to cooperate with the ideal X01’s listing rhythm, Huang Renxun personally filled the board decided to submit the mass production time before 2022;

2. The power consumption of theorin, initially said that it is 67 W, but it is adjusted to 45 W before and after the release of ideal cooperation in September 2020.

3. In the case of 200 TOPS, it is 200 TOPS from the release to the ideal cooperation, but it is time to upgrade to 254 TOPS in January 2021.

Originally, the mass production time was already very rushed one year in advance, and the Ying Weida has once again downs to upgrade, this is very curious – if there is a sample, the power consumption is reduced And force upgrade needs to be redesigned, re-flow (may be two times); and if there is no sample, the power consumption reduction and the force upgrade should be only “lower” or “upgrade” on the PPT.

In addition, the high-pass SA9000 is currently staying in the sample stage. At the Beijing Auto Show at the end of September 20020, the author learned on Huawei’s booth, and the chip used by the MDC 610 of 160 TOPs, which has been prepared before the US sanctions.

During the Beijing Auto Show, Momenta CEO Cao Xudong said in a speech, “MDC 610 is the highest in the automatic driving calculation platform of the quantified production.”

On April 12, 2021, British Weida, which was very heavy to “rollove the second”, released a new automatic SOC Atlan, and a single SOC’s power of 1000 TOPS, which is close to 4 times compared to ORIN SOC.

However, according to Zissi-American research recently, Altan’s premise of such a biggest force is “less considered cost, not considering power consumption”.

The author also pointed out that Ying Weida launched an intention of Atlan to “pulling up the opponent to the arms race, making a lottery in publicity, oppressing the opponent must follow the game, until the opponent” “” Other manufacturers I will not follow this Astantbed digital game, which is out of the actual demand .orin is probably the main product of British Weida in the next few years. “Although Altlan can provide samples to developers in 2023, as for the mass production loading time, then wait 2025 Year.

If you can’t provide any value for our customers if you can’t provide any value for our customers.

At that day, the author did not intend to be a friend of a friend who was responsible for marketing in Huawei MDC platform, “Not blowing the food, only spelling the goods”. Recently, the automatic driving calculation platform equipped with strong automatic driving capabilities repeatedly on the hot search, the automatic driving computing platform equipped with Huawei is the “spot” MDC Pro 610 for this car custom development.

The MDC Pro 610 has a total strength of up to 400+ TOPS. It looks, it is better than 4 British Weida’s ORIN, but it is better in “there is spot”.

In recent days, in recent days, when Huawei’s automatic driving car ran on the open road in the city, many industry insiders were very shocked. There were friends to ask me, my answer is:

“I didn’t feel unexpectedly. These years, the automatic driving circle has a bad breath. Many companies have not wanting to know if they don’t want to do something, they will start to say ‘I plan to do’, do 30 points. Time, saying “Do it right away”, only 60 points, began to emphasize that “is already the first ‘;

And I found that Huawei’s style is that if they say ‘I plan to do it, I have already achieved 30 points. If I say “I am doing”, I have already done 60 points, wait until I do it, it is almost It has already achieved 80 points, and even the mature products that have been delivered.

This is a habit of Huawei formed in the process of long-term way to do To B business. They are so doing, often make competitors difficult to follow up. “

The words retired, Huawei made a lot of stockings to MDC 810, but constipated secrets, this is also the culture of culture that has long been “less than doing”.

Next, Huawei MDC will adhere to the platform-standardized R & D layout, combined with the industry’s universal intelligent driving force demand, based on the underlying capacity of MDC Pro 610 to facilitate the platform-based product MDC 810 for all car companies.

Why is the British Weida and Qualcomm sell a chip without selling “boxes”?

The careful friend should have noticed that Huawei MDC products rarely make specific parameters of the chip in the marketing, but directly talk about the total strength, overall power consumption of the entire MDC platform. This is determined by the business model of Huawei automatic driving business – not selling chips alone, but “selling boxes”.

The so-called “selling box” refers to a full stack solution (including but not limited to domain controllers), including automatic driving chips, operating systems, algorithms, sensor solutions, and development tool chains. This full stacking program has passed ISO 26262 ASLD certification, and is about to be equipped with a mass production car in Beiqi and Changan and Hua.

Of course, while providing a box, Huawei maintains a high open open, for example, customers can decide how to write algorithms, how to match the sensor. In addition, Huawei does not compete with the control of the car.

In contrast, Ying Weida and Qualcomm mainly provide chip and development tool chain, “box” is mainly completed by Tier 1 designated by car enterprises. For example, in British Weida and Xiaopeng and the ideal cooperation, the “box” work is borne by Deye.

Under the big trend of hardware and software decoupling, there are many cars that they want to write algorithms, design sensor programs, so they also tend to buy chips from chip manufacturers, and then cooperate with Tier 1 to do “box” However, there are still many cars, and now we have an urgent festival car to carry the automatic driving kit, and self-research algorithm, it is not easy to design, then, for these customers, if the chip manufacturer directly provides the algorithm “Box”, but a better choice.

Even those who have self-study algorithm, only buy chips from chip makers, should soon realize that it is not “as long as there is a chip, it is easy to engage in a domain controller.” In fact, the process of “from the algorithm to the box” is full of various “pit” engineering challenges. If the “box integrator” (Tier 1) is not enough, the automatic driving volume of the car enterprise will be delayed; or, although it is delivered as scheduled, the effect is a gap between expected.

At this time, if the chip manufacturer has a strong system integration ability, it can do “box” very well, then the car can choose from the chip supplier here to buy back the hardware part of the whole box and the operating system, tool chain, Then inject yourself written algorithm.

Standing at the point of view of the car enterprise, the chip manufacturer “selling box” has a significant benefit compared to only selling a chip, the responsibility is relatively clear, there is a problem, they only look for the box for the vendor, not by the chip manufacturer , Operating system vendors and domain controller suppliers “kick the skin” each other.

In addition, the “hardware upgrade” trend that is about to become trend is objectively able to provide “chip manufacturers of the whole box.

Before talking about “hardware upgrade”, you need to talk about “hardware pre-burr” now. As we said in the beginning of this article, how much power is required for the L4-level automatic driving, no one is missing; in fact, what we see is that this “How much is enough” standard is continuous upgrade.

Tesla was “the hardware that supported L4” in the second half of 2016, and later found not enough, and added two chucks in the summer of 2017, upgraded to autopilot 2.5; then, then discovered 2.5 Not enough, and then Hardware3.0 hardware has been launched. Hardware 3.0 is known to meet the needs of L4, but now Tesla is in the hardware of 4.0, the new version of hardware intends to mass production before the end of 2021, which is equal to recognizing that 3.0 hardware is “not enough”.

In addition, the thin friend should pay early, and the Xaida’s Xaiver, I said it was “L4 chip”, but now I dare to play in the L2 market. Before orin said it is a L4 chip, but it may be a “L2 chip” after Atlan mass production.

Therefore, we need to imagine a question: In the past few years, if the chip manufacturers have launched a low-power chip / computing platform that intensive up to 2000 TOPS, those now fully “support L4 hardware (力 1000 TOPS “The car enterprises can still” sit “?

Since the iterative cycle of the chip is shorter than the life cycle of the car, the forward-looking car companies such as Tesla, Ideal, Toyota are preparing for “hardware upgrade” while doing “hardware.” The pre-buried hardware cannot meet the higher level of automatic driving function to upgrade the hardware, which requires the corresponding hardware interface at the beginning of the vehicle design. If you can’t upgrade directly, you need to get new hardware in the reservoir version.

When cope with the “hardware upgrade” trend, Huawei’s “selling box” companies, obviously more advantages over the company with only WVIDA and Qualcomm. In fact, Huawei has long prepared for this. At the Beijing Auto Show last September, Huawei exhibited MDC 610 is exactly the same as the size and interface of MDC 210, which is to “plug and play” when it is guaranteed to do hardware upgrades in the future.

In contrast, if the chip manufacturer does not do a box, the progress may be dragged by the box integrator when the “hardware upgrade” is required in the subsequent car enterprise.

Ying Weida has done a box of Drive PX2 PARKER, Drive Xavier, Drive Pegasus, etc. Based on Xavier chip, but is located in “Development Board” (for functional verification and technical training), not suitable for mass production, otherwise it will not Let Des Saiwei make an IPU03 box based on Xaiver chip.

In theory, for the British Weida and Qualcomm, in addition to the sale of the chip, it will add a “selling box” option, which is better than only sell chips. However, why did they do this?

The answer is: insufficient engineering capacity. In particular, Qualcomm has never done the whole machine.

Engineering skills, similar to Japan’s “craftsman spirit”, the surface looks “monotonous boring”, there is a high barrier.

For example, can you get a refined problem?

The “box” smaller than the power consumption is low, it can naturally heat, neither the fan does not require liquid cold; but usually only the product consumes less than 30W can naturally heat, once the power consumption exceeds 30W, It is difficult to naturally heat dissipation. If the heat dissipation problem is not treated, the temperature will rise sharply, hot, leading to the life and stability of the product will be affected.

Also, regardless of how high the process is high, the improvement of power will increase the power consumption, as for the optimization architecture to reduce chip power consumption, often only make power consumption problems to be relieved, but can’t cure, this means, force force The improvement will inevitably make a big challenge for the heat dissipation capacity of the system; if the heat dissipation problem is not good, it will seriously affect the reliability of the product.

For example, the “box” equipped on the passenger car mainly leverages liquid cooling, no fan, even if the airtightness is done well, there will be a small amount of air in the box, if the temperature is lower than the outside, the temperature ratio High, when water vapor inside, it is easy to liquefy;

Sometimes, the bottom cover below the dew, if the processing is not good, it will drop it on the board, causing the circuit board to be burned.

Need to emphasize: When doing a chip, it is a very high barrier when the chip is established; but when the chip is intended to be “box”, how to handle the heat dissipation of high-power consumption, it is a high barrier. Unfortunately, under the boom of the force competition, the latter barrier is often ignored.

Traditional car age, due to the low proportion of electronic components, and the ECU is generally small, the power is not too high, so small and medium tier 1 does not have to process heat dissipation problems, which causing them to drive automatic driving. Insufficient experience when the heat dissipation of the domain controller.

For example, how does the line on the board? The spacing will take up the board area, causing the space being insufficient; the spacing is too small, the signal interference will be more serious, which in turn leads to extremely unstable product performance. For example, can the glue use of the electronic component not to prevent water? Some experienced Tier 1 for the cost of the cost, the price is low, the result, only protective water, but can not immerse water.

Barriers in factory technical capabilities often have enough time to street. And Huawei has accumulated engineering technology capabilities such as heat dissipation and precision manufacturing during the process of doing base stations.

The market pattern of the large-minder automatic driving chip is stable

In addition to the British Weida, Qualcomm and Huawei, the large – mindant force automatic chip market may have a potential player that is easy to ignore: Waymo.

Because “can’t see” other companies, Waymo has been self-developed automatically, the chip will be distributed by Samsung. Since the second half of 2019, Waymo has begun to explore the RobTaxi operator while doing Tier 1., providing car companies to include automatic driving operating systems, algorithms, chips, laser radar, millimeter wave radar. .

Waymo appeared as Tier 1, in business model, is very similar to Huawei. But there is a major difference that Waymo is a company born company, completely unable to do hardware genes, their “box” integration, will be handed over to Magna.

In theory, if Waymo’s road can go, Ying Weida, Gao Tong, Huawei’s share of the driving market will decrease. But reality may not walk towards Waymo designs.

Huawei cooperated with Beiqi, a 3 laser radar +6 millimeter wave radar +13 cameras, compared with 5 laser radar on Waymo, and the camera is up to 29, which is complete There is no feasibility of mass production, there is no cost awareness.

From John Krafcik’s undergrass, Waymo may have realized that the way directly to do L4 is not passing. Next, they plan to work with the car to attack the front (L2 +) market, the sensor program will also “decrease”. However, for Waymo, Tier 1 for the L2 market may be more difficult than doing Robotaxi operators.

Because when doing the Robotaxi business, Waymo yourself is Party A, you can define your needs, buy your car and components to integrate; but if you do tier 1, they still need to get Tier 2 or even tier 3, not only ensuring each The quality of the parts is no problem, but also controls costs, which seems to be far exceeding Waymo’s ability.

In addition, Waymo’s original L4 algorithm architecture is made based on 5 laser radar. Now do L2 +, the sensor has been reduced, can your algorithm are not adjusted?

From the extravagant, it is extravagant, and Waymo will experience a quite long “pain period” after the sensor is reduced.

Worse, although Waymo finally reassured to give car companies to “Party B”, the vehicle enterprises may not be bought.

In the first half of last year, Waymo did not participate in investment in the first half of the year. That is to say, there is no car company with Waymo’s potential “technical supplier” depth bundle.

Perhaps, in order to invest Waymo as the standard, you can divide the car manufacturers two: money but don’t want to cast; I want to vote but no money.

Let’s take a look at “rich” head car companies:

Volkswagen – Ford, Toyota, General – Honda’s company / alliance, has been a fund for self-developing automatic driving.

BMW has aligned with Mobileye, Daimler’s passenger cars have allocated Ying Weida deeply, only with Waymo in the truck business (but there is still no self-developing automatic driving truck technology).

Modern has been funded by $ 1.6 billion with Anbofffu …

Do not exclude these car companies to take hundreds of people in the local market, even dozens of cars do some exploratory cooperation with Waymo, but only this is only.

In fact, Kenmo’s customers (carrying Waymo’s automatic driving technology to operate), but also limited to some vulnerable vehicles, at least some car companies in the automatic driving area.

But Waymo is not strong with these “vulnerable car companies”.

Such as Waymo’s first car enterprise customer Renault Nissan Mitsubishi. After Gorn Incident, the relationship between Renault Nissan Mitsubishi Alliance continues, and there is a big question. If the affiliate is disintegrating, can the agreement with the Waymo sign as usual?

FCA, although it may work with Waymo to carry out travel operations in Italy, the company has joined BMW, Intel-Mobileye-Delphi, which is also intended to use the alliance development technology.

The Jaguar Land Tiger, which is currently doing suppliers to Waymo, as if not intended to make Waymo customers. Because the company has invested 25 million US dollars in the second largest travel company of the United States, 2020, and cooperated to develop automatic driving technology. Waymo has also collaborated with Volvo in June 2020, but according to a person who has participated in cooperation negotiations, in fact, the nature of the cooperation between the two parties is still “Waymo”, not “Volvo uses Waymo. Automatic driving technology.

In fact, Waymo’s technical authorization mode shouted for at least two years, but there is nothing to get.

Because the car companies are not willing to do “Foxconn”, Waymo feels that it is not easy to do operators; but the technology authorized this road is not good. Because it is not willing to do “HP” “Dell” or ” Android phone manufacturer “- that is, the operating system and chip are fully dependent on the supplier.

After combing, it was found that the car company is willing to do “HP” “Dell”, it should be mainly Mazda, Subaru, Suzuki, a weaker small and medium-sized car.

As for China’s car companies, in view of the Chinese market, there is so many automatic driving plan companies, and the headless companies are determined to self-develop automatic driving technology, Waymo doesn’t expect to be their Tier 1?

In conjunction with the above analysis, Waymo’s technical authorization model, ideal, but realistic bone, it is very likely to find customers. Therefore, in the end, Waymo will “forced the car” to achieve “internal circulation”. However, Waymo after the internal cycle is not much likely to compete with the Ying Weida, Gao Tong, Huawei will compete in the chip market?

Leave a Reply

Your email address will not be published. Required fields are marked *