Author: Su Qingtao
The intelligent driving computing platform installed on the North Benz ARCFOX Alpha S is the MDC Pro 610 customized and developed by Huawei for this car, with an up to 400+ TOPS computing power. Next, Huawei MDC will combine the industry’s general demand for intelligent driving computing power and develop MDC 810, a platform-standardized product for all car company customers based on the underlying capabilities of MDC Pro 610.
There is no clear answer as to how much computing power is needed to achieve level 4 autonomous driving, and car companies are hesitant to draw conclusions while chip manufacturers continue to adjust their figures.
A few years ago, some people said it was 300 TOPS; later, some people said it was 500 TOPS; and then, some people said it was 1000 TOPS… Judging from the Atlan AI chip with TOPS computing power, the computing power of the autonomous driving domain controller should soon be “inflated” to over 2000 TOPS.
However, regardless of whether their algorithmic level in the next few years really requires such high computing power, car companies must be actively or passively involved in this “computing power arms race”. As salespeople from a certain autonomous driving chip company put it, this is also “internal crowding out”. But there is no way around it. In the current environment, apparently only high computing power can give car companies sufficient “sense of security”.
However, at the MDC 810 (computing power of 400+ TOPS) launch event on April 18, Huawei emphasized that the autonomous driving domain controller cannot rely solely on computing power, but must also consider the engineering capabilities behind it. It is this kind of ability that gives Huawei the confidence to announce “mass production upon release” on the day of the launch event.
Currently, there are only three suppliers of autonomous driving chips with more than 200 TOPS high computing power: Nvidia, Qualcomm, and Huawei. From the information that has been publicly announced so far, Nvidia and Qualcomm’s latest autonomous driving chips have higher computing power, and Nvidia has also won the most mass production orders. However, both companies are positioned as Tier 2 and only sell chips, lacking system integration experience, while Huawei, positioned as Tier 1, has already accumulated strong machine engineering capabilities.
Undoubtedly, under the trend of software and hardware decoupling, more and more car companies will choose to bypass Tier 1 and negotiate directly with chip manufacturers for cooperation. However, at the current stage, most car companies still do not have the ability to make their own autonomous driving domain controllers – even if they have this ability, the opportunity cost is still relatively high. Therefore, they still need to entrust a Tier 1 to handle the integration work.
As a result, if the chip manufacturer can provide the domain controller themselves, the workload for the car company will be greatly reduced.
The “computing power arms race” and Nvidia’s “seating anxiety”On September 29, 2016, at GTC Europe, NVIDIA officially announced its next-generation Tegra processor, codenamed Xavier. As planned, Xavier will be manufactured on a TSMC 16 nm process, integrate as many as 7 billion transistors, and achieve 20 TOPS of computing power at 20 W power consumption. According to NVIDIA’s initial plan, Xavier was scheduled to ship on a TSMC wafer by the end of 2017.
Before that, according to publicly available information, the highest computing power among the planned autonomous driving chips was Mobileye’s EyeQ 5, with a maximum power consumption of 5 W and computing power of 12 TOPS. It seems that NVIDIA’s Xavier is aiming for the “first” position.
However, in December 2017, Mobileye, which had already been acquired by Intel, announced that the EyeQ 5 will be released in two versions: 12 TOPS and 24 TOPS. With the EyeQ 5 now capable of 24 TOPS, Xavier’s 20 TOPS is likely to lose its “first” status.
In fact, the “never-give-up” NVIDIA has already taken the initiative. On October 11, 2017, at the GPU Technology Conference in Munich, Germany, NVIDIA announced an upgraded version of Xavier: 12 nm, 9 billion transistors, and 30 TOPS, with power consumption also increased to 30 W.
At this conference, NVIDIA also launched a computing platform called Drive PX Pegasus, which claims to be “designed for L4 autonomous driving”. The Drive PX Pegasus includes two Xavier SoCs and two Turing-based GPUs, with computing power of up to 320 TOPS. Although this computing platform, with a power consumption of up to 500 W, is not suitable for mass production, it was still of great interest to many companies at that stage as it was the “king of computing power”.
Unfortunately, Pegasus, which pursued high computing power at all costs, did not stay in the “first” position for long. In October 2018, at Huawei’s Global All-Connect Conference, the MDC 600 was unveiled with a computing power of up to 352 TOPS and a lower power consumption (352 W) than the Pegasus.
What made NVIDIA particularly nervous was that the main chip in the MDC 600, the Ascend 310, had a computing power of up to 16 TOPS with power consumption of only 8 W, giving an efficiency of 2 TOPS/W, much higher than Xavier, which is also based on the 12 nm process, with an efficiency of 1 TOPS/W.In the midst of having its title of autonomous driving computational power champion snatched away by Huawei, Nvidia’s partnership with Tesla is also coming to an end. In late 2017, Tesla officially admitted to developing its own autonomous driving chips, and in August 2018, Tesla announced that its self-developed chip would soon be completed.
At Tesla’s “Autonomy Day” on April 22, 2019, the company officially announced the specifications for its self-developed FSD chip: 16 nm, 72 TOPS, 36 W. Both in terms of computational power and energy efficiency, it outperforms Xavier by a wide margin.
Indeed, considering that ASIC’s high energy efficiency comes at the expense of sacrificing some flexibility, it is “unfair” to compare its computational power and energy efficiency with that of the GPU-based Xavier. However, in the context of the race for computational power, few customers will rationally look at this issue, and even Nvidia itself has been taken hostage.
So in September 2019, with only XPeng as a mass production customer, Nvidia eagerly announced the 7nm Orin autonomous driving chip with computational power of up to 200 TOPS.
According to plan, Orin will not be mass-produced until 2023. Some speculate that Nvidia was so eager to launch Orin because it was “stimulated” by Huawei’s MDC 600 (technology research version).
However, Orin did not hold onto the top spot on the computational power rankings – it was quickly overtaken by Qualcomm.
At CES in January 2020, Qualcomm also unveiled its autonomous driving computing platform, Snapdragon Ride, which can achieve 700 TOPS of computational power under a power consumption of 130 W. At that time, it was reported that Cruise and Argo were testing with Qualcomm’s autonomous driving chip.
After CES, several Chinese car manufacturers began to engage with Qualcomm. According to insiders, Qualcomm’s autonomous driving chip has two models: SA9000 A and SA9000B, with computational power of 200 TOPS and 300 TOPS, respectively. Like Nvidia’s Orin, the SA9000 is based on a 7nm process, but its energy efficiency is as high as 5 TOPS/W (Orin’s initial version has an efficiency of 3 TOPS/W).
Qualcomm’s chips are highly competitive in the mobile phone market and the intelligent cockpit market. However, in the autonomous driving market, Qualcomm is a newcomer whose previous advantages may not be fully utilized – Qualcomm excels in CPU, but its GPU capabilities are not as strong, and in autonomous driving calculations, GPU cores play a relatively large role.Vice President Wu Xinzhou of XPeng’s autonomous driving division previously served as head of autonomous driving at Qualcomm. It is said that XPeng also had negotiations with Qualcomm, but ultimately decided to choose NVIDIA’s Orin instead of Snapdragon Ride or SA9000.
Since September 2020, companies such as Li Auto, XPeng, NIO, SAIC Zhi Zi, and SAIC R have all chosen Orin for their “embedded hardware” schemes. It seems that NVIDIA has achieved a stage victory, but in the context of hardware arms race, even more computing power is “not enough”. Therefore, NVIDIA still cannot rest easy.
There are a few anecdotes:
-
Orin was originally planned to be mass-produced in 2023, but to match the launch rhythm of Li Auto X01, Huang Renxun personally decided to advance the production time to 2022;
-
Orin’s power consumption, which was initially stated as 67 W, was later adjusted to 45 W around the time of the launch event with Li Auto in September 2020;
-
Orin’s computing power, which was stated publicly as 200 TOPS from release to cooperation with Li Auto, was upgraded to 254 TOPS at the NIO ET7 launch event in January 2021.
Originally, advancing the production time by one year was already hasty, but NVIDIA has since made a power consumption reduction and a computing power upgrade, which is very curious – if there are already samples, the power consumption reduction and computing power upgrade would require redesigning and re-fabricating (possibly starting over 2 times); if there are still no samples, the power consumption reduction and computing power upgrade should only be “reduced” or “upgraded” on the PowerPoint.
In addition, Qualcomm’s SA9000 is currently in the sample stage. At the Beijing Auto Show at the end of September 2020, the author learned at Huawei’s booth that the chip used by MDC 610 with a computing power of up to 160 TOPS had already been stocked before the US sanctions.
Also during that Auto Show, Momenta CEO Cao Xuedong said in a speech that “MDC 610 has the highest computing power among autonomously driving computing platforms that have been mass-produced.”
On April 12, 2021, NVIDIA, who attaches great importance to seating arrangements, released a new autonomous driving SoC, Atlan, which has a single SoC computing power of up to 1,000 TOPS, nearly a four-fold increase compared to Orin.
However, according to the comments in the recent report “Detailed Analysis of NVIDIA’s Latest Autonomous Driving Chip – Atlan” by Sino Auto Insights, the premise of such a high computing power for Atlan is that it does not care much about cost and power consumption.The author pointed out that one of the purposes of NVIDIA’s introduction of Atlan is to “pull up competitors to engage in a computing power arms race, create a big sensation in publicity, and force competitors to follow the computing power game until they are overwhelmed”. “Other manufacturers may not follow this computing power numbers game as it is detached from practical needs. Orin may be NVIDIA’s main product for the next few years.”
In addition, Altlan will only be available to developers in 2023, and mass production for vehicles will have to wait until 2025.
Even if the computing power is impressive, if samples cannot be provided in the short term, it cannot create any value for customers.
It was also on that day that I accidentally saw a sentence on a friend of a friend’s circle who was in charge of marketing for the Huawei MDC platform: “Don’t brag about futures, only compete with present goods for real strength.” And the automatic driving computing platform on the BJARCFOX ALPHA S, which has recently become popular due to its powerful automatic driving ability, is the “present goods” MDC Pro 610 customized and developed by Huawei for this car.
MDC Pro 610 has a total computing power of more than 400 TOPS. It seems that the computing power cannot match that of NVIDIA’s Orin with four chips, but it wins in its “availability of present goods”.
By the way, in recent days, when the self-driving car co-developed by Huawei and BJEU was running on open roads in urban areas, many industry insiders were very surprised. A friend asked for my opinion, and my answer was:
“I am not at all surprised. In recent years, there has been a bad trend in the self-driving industry, where many companies start to say ‘I plan to do it’ before they even figure out if they should do it. When they achieve 30%, they say ‘it will be done soon’, and when they reach 60%, they begin to emphasize that they ‘are already the first’;
while I found that Huawei’s style is if they say ‘I plan to do it’, they have already achieved 30%, and if they say ‘we are doing it’, they have already achieved 60%. When it comes to the press conference, they have already achieved about 80%, or even have mature products that can be delivered.
This is a habit that Huawei has developed in its long-term To B business. They often make it difficult for competitors to catch up by doing this.”
Returning to the topic, Huawei has stocked up a lot for MDC 810, but has been keeping it secret, which is also in line with its long-standing culture of “doing more and speaking less”.
Next, the Huawei MDC will adhere to the platform-based standardization research and development plan, combined with the industry’s general demand for intelligent driving computing power, and based on the underlying capabilities of MDC Pro 610, create a platform product, MDC 810, for all automakers.
Why do NVIDIA and Qualcomm only sell chips, not “boxes”?Careful friends may have noticed that Huawei rarely mentions the specific parameters of the chips in its MDC products in market promotions. Instead, it focuses on the overall computing power and power consumption of the entire MDC platform. This is due to Huawei’s business model in the autonomous driving industry – selling the “box” rather than just the chips.
Selling the “box” means that Huawei provides a full-stack solution including autonomous driving chips, operating systems, algorithms, sensor solutions, and development toolchains (including but not limited to domain controllers). This full-stack solution has passed ISO 26262 ASLD certification and will soon be installed on mass-produced vehicles in cooperation with BAIC, Changan and Huawei.
Of course, while providing the “box”, Huawei also maintains sufficient openness. For example, customers can decide how to write their algorithms and configure sensors. In addition, Huawei will not compete with automakers for control over data.
In contrast, Nvidia and Qualcomm mainly provide chips and development toolchains, and the “box” is mainly completed by Tier 1 designated by automakers. For example, in Nvidia’s cooperation with XPeng and Ideal, Desay SV Automotive is responsible for the work of the “box”.
In the trend of decoupling software and hardware, many automakers hope to write their own algorithms and design sensor solutions. Therefore, they tend to buy only chips from chip manufacturers and then cooperate with Tier 1 to make the “box”. But there are still many automakers who are eager to install autonomous driving kits on mass-produced vehicles, but they do not have time to develop their own algorithms and design collaborations. For these customers, if chip manufacturers can directly provide the “box” including algorithms, it is actually a better choice.
Even those automakers who have the ability to develop their own algorithms and only buy chips from chip makers should soon realize that “just having chips and making a domain controller is not easy”. In fact, the process from “algorithm to box” is full of various engineering challenges. If the chosen “box integrator” (Tier 1) is not strong enough, the mass production progress of automakers’ autonomous driving will be delayed. Or, even if it is delivered on time, the effect may be different from expected.
At this time, if the chip manufacturer has strong system integration capabilities and can make the “box” well, automakers can choose to buy back the hardware part, operating system and toolchain of the entire box from the chip supplier and then inject their own algorithms.
From the perspective of automakers, compared with just selling chips, the chip manufacturer selling the “box” has a major advantage: clear responsibility. When problems occur, they only need to find the box supplier, instead of being kicked around by chip manufacturers, operating system manufacturers, and domain controller suppliers.
In addition, the upcoming trend of “hardware upgrades” is objectively beneficial for chip manufacturers who can provide the entire “box”.Before talking about “hardware upgrades”, we need to talk about the currently popular “hardware embedding”. As we mentioned at the beginning of this article, there isn’t a fixed standard for how much computing power Level 4 autonomous driving needs. In fact, the standard for “how much is enough” keeps increasing.
In the latter half of 2016, Tesla launched Autopilot 2.0, which they claimed was “hardware capable of supporting Level 4”, but later found inadequate. In the summer of 2017, they added two small chips to upgrade to Autopilot 2.5, and then introduced Hardware 3.0, claiming it met the requirements for Level 4. However, Tesla is now working on Hardware 4.0, which is planned to be mass produced by the end of 2021, indicating that the computing power of Hardware 3.0 is “insufficient”.
Moreover, observant friends may have noticed that Nvidia’s Xavier was originally dubbed a “Level 4 chip”, but now is only used in the Level 2 market. Orin was also referred to as a Level 4 chip before, but perhaps after the production of Atlan, it will be classified as a “Level 2 chip”.
Therefore, we need to consider a question: in a few years, if chip manufacturers introduce low-power chips or computing platforms with computing power up to 2000 TOPS, would car companies with “hardware capable of supporting Level 4 (around 1000 TOPS of computing power)” still be able to “keep up”?
Since the chip iteration cycle is shorter than the vehicle lifespan, forward-looking carmakers such as Tesla, NIO, and Toyota also prepare for “hardware upgrades” while doing “hardware embedding”. When embedded hardware cannot meet the data processing needs of higher-level autonomous driving functions, the hardware can be upgraded or new hardware interfaces can be reserved during initial vehicle design. If direct upgrades are not possible, new hardware needs to be adopted in redesigned versions.
In response to the trend of “hardware upgrades”, companies like Huawei, who sell complete systems, obviously have a greater advantage than companies like Nvidia and Qualcomm who only sell chips. In fact, Huawei has long been prepared for this. At the Beijing Auto Show last September, Huawei exhibited MDC 610 and MDC 210, which have the same size and interfaces, ensuring that they’ll be “plug-and-play” when doing hardware upgrades in the future.
In contrast, if chip manufacturers do not sell complete systems, their progress may be hindered by the box integrator when car companies need “hardware upgrades”.NVIDIA has developed various boxes based on the Xavier chip, including DRIVE PX2 Parker, DRIVE Xavier, and DRIVE Pegasus. However, these are all positioned as “development boards” for functional verification and technical training, and are not suitable for mass production. Otherwise, there would be no need for Desay SV Automotive to develop a separate IPU03 box based on the Xavier chip.
In theory, adding a “selling box” option to NVIDIA and Qualcomm’s chip sales would be a better choice than just selling chips. However, they did not do so because of their insufficient engineering capabilities, particularly Qualcomm’s lack of experience in developing whole devices.
Engineering and technical capabilities, similar to the “craftsmanship” of Japan, may appear “monotonous,” but they have high barriers to entry.
For example, can you solve the heat dissipation problem?
For “boxes” with relatively low computing power, because the power consumption is low, natural heat dissipation is possible without the need for fans or liquid cooling. However, usually only products with power consumption below 30 W can achieve natural heat dissipation. Once the power consumption exceeds 30 W, it becomes difficult for natural heat dissipation. If the heat dissipation problem is not handled properly, the temperature will rise rapidly and may cause the product to deteriorate in terms of lifespan and stability.
In addition, no matter how advanced the process node is, an increase in computing power will inevitably lead to an increase in power consumption. While optimizing the architecture can reduce chip power consumption, it can only alleviate power consumption issues but not solve them. This means that an increase in computing power will also pose a significant challenge to the system’s heat dissipation capability. If the heat dissipation problem is not handled well, it will seriously affect the product’s reliability.
For example, the “boxes” mounted on passenger cars depend mainly on liquid cooling without fans. Even with good airtightness, there will still be a small amount of air inside the box. If the outside temperature drops while the inside temperature remains high, the water vapor inside the box may condense when it comes into contact with the cold case of the “box.” Sometimes, if the dew underneath the bottom cover is not handled well, it may drip onto the circuit board and cause it to burn out.
It is worth emphasizing again that while achieving low power consumption is a barrier when developing chips, how to solve the heat dissipation problem under high power consumption when developing “boxes” poses a high barrier. Unfortunately, this latter barrier is often overlooked in the heat of the competition of computing power.
In the traditional automotive era, because the proportion of electronic components was low, and the computing power of ECUs was generally small and not very powerful, small and medium-sized Tier 1 companies did not need to frequently deal with the heat dissipation problem. This limited their experience in dealing with the heat dissipation problem of autonomous driving domain controllers.
Another example is how to layout the lines on the circuit board. If the spacing is too large, it will occupy board space, causing insufficient space for other components, and if the spacing is too small, signal interference may occur, leading to unstable product performance.Translation:
Can the adhesive used for electronic components prevent immersion in water? Some Tier 1 companies with insufficient experience choose cheaper adhesives for cost considerations, but they can only resist water splashing, not immersion.
In terms of factory technology capabilities, barriers are often erected by spending enough time. Huawei has accumulated engineering technology capabilities such as heat dissipation and precision manufacturing in the process of making base stations.
The market pattern of large computing power automatic driving chips is stable.
Apart from Nvidia, Qualcomm, and Huawei, there may be an easily overlooked potential player in the market for large computing power automatic driving chips: Waymo.
Because they do not like the chips made by other companies, Waymo has independently developed automatic driving chips, which will be manufactured by Samsung. Since the second half of 2019, Waymo has started to explore the full-stack solution for car companies, including automatic driving operating systems, algorithms, chips, LIDAR, and millimeter-wave radar, while also operating taxi operators.
As a Tier 1 player, Waymo is extremely similar to Huawei in terms of business model. But a significant difference is that Waymo is a software company with no hardware genes, and they will hand over the integration work of their “box” to Magna.
In theory, if Waymo’s approach works, the share of Nvidia, Qualcomm, and Huawei in the automatic driving market will decrease. But reality may not follow the script designed by Waymo.
In cooperation with BAIC, Huawei’s car is equipped with three LIDARs, six millimeter-wave radars, and 13 cameras. In contrast, Waymo’s cars are equipped with five LIDARs and as many as 29 cameras, which completely ignored the feasibility of mass production and had no cost consciousness.
Considering John Krafcik’s departure, Waymo may have realized that the direct approach to L4 is unsustainable. Next, they plan to cooperate with car companies to attack the pre-installation mass production (L2 +) market, and the sensor solution will also be “downgraded.” However, for Waymo, it may be more difficult to do Tier 1 business in the L2 market than to be a robotaxi operator.
Because in the process of doing the robotaxi business, Waymo itself is the first party and can define its own needs, buy cars and components to integrate them. But if they do Tier 1, they will need to first deal with Tier 2 or even Tier 3, not only to ensure that the quality of various components is okay, but also to control costs, which seems to have far exceeded Waymo’s abilities.After sensor cutbacks, Waymo is going through a “painful period”. In addition, even though Waymo has finally started to collaborate with automakers as a “second party”, it is uncertain if the automakers will agree to it. None of the automakers participated in the $3 billion financing that Waymo secured in the first half of last year. Using this as a standard, automakers can be divided into two categories: those who have money but are unwilling to invest, and those who want to invest but do not have the money. Volkswagen-Ford, Toyota, and General Motors-Honda have all spent at least $5 billion on their own autonomous driving systems. BMW has entered into a deep alliance with Mobileye, and Daimler’s passenger cars have entered into a deep alliance with Nvidia. Hyundai has invested $1.6 billion to establish a joint venture with Aptiv. Some automakers may work with Waymo in certain markets by deploying several hundred, if not several dozen, cars for exploratory cooperation, but that is about it. Actually, only some weaker automakers, or those with less influence in the field of autonomous driving, can be considered as customers who will use Waymo’s self-driving technology for operation. But, the relationship between Waymo and these weaker automakers is not strong. The Renault-Nissan-Mitsubishi Alliance, Waymo’s first automaker customer, has a lot of uncertainty regarding its sustainability following Ghosn’s incident. Will the previously signed agreement with Waymo be fulfilled if the alliance collapses? Although FCA may work with Waymo to launch travel operations in Italy, the company joined the BMW-Intel-Mobileye-Delphi Alliance as early as 2017, and its Maserati marque has indeed planned to use the autonomous driving technology developed by the alliance. Jaguar Land Rover, which is currently supplying Waymo, seems to have no plans to become Waymo’s customer again, as it recently invested $25 million in Lyft to cooperate on developing self-driving technology.Waymo partnered with Volvo in June 2020, but according to a person who participated in the negotiation, the nature of the partnership was still “Volvo selling cars to Waymo” rather than “Volvo adopting Waymo’s autonomous driving technology”.
In fact, Waymo’s technology licensing model has been around for at least two years, but there hasn’t been much achievement.
Because car companies are unwilling to be like “Foxconn”, Waymo finds it difficult to be an operator; but the road of technology licensing is not smooth either, because proactive car companies are also unwilling to be like “HP”, “Dell”, or “Android phone manufacturers” – that is, completely dependent on suppliers for operating systems and chips.
After combing through it, the car companies that are willing to be like “HP” and “Dell” are mainly small and medium-sized car companies with weaker presence such as Mazda, Subaru, and Suzuki.
As for Chinese car companies, given that there are so many autonomous driving solution companies in the Chinese market, and top car companies have a strong determination to self-develop autonomous driving technology, Waymo can’t expect to be their Tier 1, right?
Based on our analysis above, Waymo’s technology licensing model is idealistic, but the reality is tough, and they are likely to be unable to find clients. Therefore, ultimately, Waymo will be “forced to make cars” and achieve “self-supported circulation”. But after self-supported circulation, Waymo is unlikely to compete with Nvidia, Qualcomm, and Huawei in the chip market, right?
This article is a translation by ChatGPT of a Chinese report from 42HOW. If you have any questions about it, please email bd@42how.com.