The development cycle for mass production of autonomous driving systems is about 2 to 3 years, so computing platform manufacturers always provide chip samples 2 to 3 years in advance.

Chip mass production began only after the development of the entire system was completed.

As a result, in fact, the chip pattern for autonomous driving after 2023 is already basically determined today.

Autonomous vehicle chips are expensive to develop and require at least 7 nm and 5 nm processes to manufacture them due to high performance and low power consumption requirements.

This level of process requires relatively high shipment volume:

On the one hand, because TSMC has almost a monopoly on the foundry of high-performance chips below 7 nm, there is a shortage of production capacity.

If the order volume is too low, chip manufacturers will wait for the schedule in TSMC's order.

This schedule is one and a half to three years long.

Chip manufacturers will definitely lose customers during this time.

On the other hand, chips below 7 nm are expensive to develop, starting at $1 billion.

If not enough shipments are amortized, the unit price of a chip will be very high, which in turn will affect sales.

Among the major players in the current market, Tesla and Apple have closed systems that integrate software and hardware, and do not sell chips separately to the outside world.

Huawei provides the MDC computing platform, but its chips are not sold separately.

Currently, major global independent chip manufacturers that can provide high-performance autonomous driving chips and have a presence in the market include Mobileye, Nvidia, Renesas, and Qualcomm.

The corresponding chip products are as follows:

Because these chips involve multiple versions, the comparison here is a list of top products.

Mobileye EyeQ6: Embracing Intel, Chasing High Performance

At the end of 2019, more than 54 million Mobileye EyeQ chips were shipped worldwide.

In September 2020, Mobileye revealed that more than 60 million EyeQ chips have been shipped worldwide.

These 60 million films are the sum of EyeQ2, EyeQ3, and EyeQ4, of which the portion added in 2020 was mainly EyeQ4.

Currently, the EyeQ5 has not been shipped in batches.

EyeQ5 provides the highest level of computing power of 24 TOPS. Compared with several other companies, this level of computing power is quite inferior.

The EyeQ6 is Mobileye's real high-performance high-end.

EyeQ6 is expected to be mass-produced in 2024/2025 and is divided into three versions: high, medium, and low.

Mobileye began designing the EyeQ5 in 2016 and chose MIPS's I6500 as the architecture.

On top of the I6500 architecture, MIPS introduced the I6500-F, which specifically targets automotive regulations, while the subsequent I7200 targets the wireless market.

Therefore, in the next generation of chips, Mobileye abandoned the MIPS architecture and decided to use Intel's Atom core [1].

Atom is the evergreen of the Intel processor family, and the typical in-vehicle platform is Apollo Lake.

In June 2016, Intel switched from Apolllo Lake to Goldmont architecture and used it extensively on Tesla, BMW, Cadillac, Red Flag, Hyundai, Volvo, and Chery.

Among them, BMW has adopted the most, almost all series.

Tesla Model 3 also uses Apolllo Lake.

The latest Atom series, the Elkhart Lake series or X6000e, launched in September 2020, uses the Tremont architecture.

Compared to the previous generation architecture, the Tremont architecture mainly adds L2 cache, the process is increased from 14 nm to 10 nm, the operating frequency is slightly increased by about 200 MHz, and the maximum turbo frequency can reach 3.0 GHz.

As with the previous generation, the Tremont architecture also has at most 4 cores.

Overall, Mobileye's chip updates are slower.

Coupled with the recent blow to Intel's CPU core business from Apple, Microsoft, and AMD, the company's market capitalization has declined significantly.

EyeQ6 will not be mass-produced until 2024, and it also appears to be a bit behind in the competition among various companies.

Renesas R-CAR V3U: Strong Japanese manufacturer, flexible and cost-effective

Renesas is the world's second largest automotive semiconductor manufacturer, the world's largest automotive MCU manufacturer, and the largest semiconductor manufacturer in Japan other than Sony (Sony's main business is mainly image sensors).

In terms of high-performance in-vehicle computing, Renesas' current top product is the R-CAR H3, which is mainly used in the cockpit field.

Initially, the R-CAR H3 also considered autonomous driving applications, but the R-CAR H3 was designed in 2013.

It's hard to predict that today's customer demand for AI computing power and CPU computing power is so strong.

The R-CAR H3 does not have a built-in AI accelerator, and the CPU power is only 40K, which clearly does not meet the requirements for autonomous driving system development.

Currently, it is mainly used in cockpit mass production, such as the 2021 Great Wall H6. The R-CAR M3 is also used in the cockpit of VW's Chinese models.

Renesas began enhancing the design of high-power chips in 2017.

The first vision SoC, the R-CAR V3H, was launched in 2019.

The chip's AI computing power is 4 TOPS. Bosch's next-generation vision system incorporates V3H, and also includes some Japanese fully automated parking systems.

In 2018, Renesas began designing an enhanced V3U of the V3H, and the design was basically completed by 2020.

At present, the outside world can apply for a sample of the V3U, which is a bit faster than the other three.

The mass production of the V3U is expected in early 2023, and Toyota and Honda also participated in the design of this chip.

There are very close groups between Japanese car companies and suppliers. I think Toyota and Honda will probably use V3U for autonomous driving systems.

The V3U internal frame is shown above: it uses an 8-core A76 design.

Instead of stacking 12 A72s like Tesla, Renesas used ARM's Corelink CCI-500, or consistent cache interconnection.

The video processing pipeline of the V3U is shown above. As you can see, the V3U has many hard core computer vision modules, including stereo binocular parallax, dense optical flow, CNN, DOF, STV, ACF, etc.

In terms of computer vision functions, support includes modules such as image formatting, target tracking, lane detection, free space depth, scene labeling, semantic segmentation, and detection classification.

In order to save costs, reduce power consumption, and also focus on the needs of automotive applications, Renesas did not use a GPU that was too expensive, but only added a low-power GPU, namely:

Powerful Technologies' PowerVR GE7400, 1 shader cluster plus 32 ALU cores, computing power is only 38.4 GFLOPS @600MHz.

Considering cost factors, Renesas did not use the trendy 7 nm process, but 12 nm process, and upgraded from the original Renesas R-CAR H3 16 nm FinFET process to 12 nm FFC process, with very little one-time expenses.

However, when it comes to AI performance, second only to those 5nm chips, Renesas claims the V3U has achieved an impressive energy efficiency ratio of 13.8 TOPS/W, which is six times that of the top EyeQ6 [2].

V3U is also a series of products. Multiple versions can be provided to meet the needs of different levels of autonomous driving. This is done to further increase shipment volume and reduce costs.

The V3U product line uses a modular design, and the A76 can be 2, 4, or 8 cores.

You don't need a GPU, and you can easily add or remove peripherals, which is very flexible.

Among the top four autonomous vehicle chip manufacturers, Mobileye, Renesas, Nvidia, and Qualcomm, only Renesas's main business is automotive semiconductors, so it places the highest priority on vehicle safety. The V3U's planning target is ASIL-D.

Nvidia Orin: Extreme performance, favored by new cars

Nvidia released the Orin chip at the end of 2019:

Mass production is expected in 2022 or 2023, and samples will be available in early 2021.

Public information about Orin remained until it was released at the end of 2019.

The software surrounding Orin is said to be extremely complex to work. The hardware is fully ready, and mass production may not be possible until the end of 2023.

Orin has top performance, but can be very expensive.

Naturally, L4 autonomous driving is also very expensive. Reducing the main chip by a few hundred dollars is also a drop in the bucket for a system that costs tens of thousands of dollars.

Most manufacturers invest in L4 to set a flag and create a high-tech image.

Large-scale mass production is very difficult. The supporting V2X, high-precision maps, and high-precision positioning are all immature, and regulations also need to be revised.

Therefore, manufacturers in the early stages of development are not sensitive to costs. In other words, the car manufacturer did not expect to cut costs on the main chip.

Like the R-CAR V3U, the Nvidia Orin is also a series of products.

The latter's lower-end product probably has only 2 to 4 A78 cores, 20 to 40 TOPS of AI computing power, and probably no Ampere GPUs or a few cores.

Qualcomm Snapdragon Ride: Attack on Titan of Mobile Chips

There is very little public information about Qualcomm's Snapdragon Ride.

Qualcomm's core business is still on mobile devices, so Qualcomm's strategy is to make the most of R&D results in the mobile phone field.

According to this strategy, Qualcomm's latest Snapdragon 888 (or SM 8350) chip will be closest to the Snapdragon Ride SoC.

Qualcomm's Ride platform is similar to Nvidia and is based on a SoC+AI accelerator separation method.

Qualcomm claims that the 888 chip will be manufactured using Samsung's 5 nm 5 LPE process, and it was decided two and a half years ago.

However, at present, Samsung's 5nm has not been used by any manufacturer, while TSMC's 5nm has been verified by Apple A14.

Regarding the key indicator of transistor density, Samsung's 8 nm is about the same as TSMC's 12 nm.

Samsung's 5 nm is about the same as TSMC's 10 nm, significantly lower than TSMC's enhanced 7 nm.

However, TSMC's 5nm production capacity has been overtaken by Apple; Qualcomm can only find Samsung.

On the 888 chip:

Arm's Cortex-A78 and Cortex-X1 are both based on the previous generation Cortex-A77.

But the two Arm processors were designed with different goals:

The Cortex-A78 focuses on providing higher performance per watt while being smaller, while the Cortex-X1 pursues maximum performance.

Cortex-X1 is Arm's first commercial product for the “CXC Project”.

In terms of performance, the Cortex-X1 will be 30% better than the Cortex-A77.

Compared to the Cortex-A78, the integer arithmetic performance of the Cortex-X1 was increased by 23%.

The Cortex-X1 also has twice the machine learning capabilities of the Cortex-A78.

Cortex-X1 is the equivalent of a “supersized core.” It has the same architectural design as the Cortex-A78, but has expanded almost everywhere.

ARM defines Cortex-X1 as a “customizable” mobile platform, and chip vendors can make requirements from ARM according to budget and demand.

ARM then adjusts the specification design of each module of Cortex-X1 according to different application scenarios.

Even though the S888 is very powerful, the transistor density is far less than TSMC's 5 nm, and not as high as TSMC's 7 nm because of Samsung's 5 nm process.

Therefore, the S888's single-core performance still lags behind Apple's previous generation A13, and the gap with TSMC's 5nm A14 is even more obvious. The A14 has a 41% higher single-core score than the S888.

The GPU side can further highlight Samsung's backwardness in process.

According to GFXBench Aztec testing:

• A14 peaks at 102.24 frames per second

• A13 reaches 91.62 frames

• S888 only has 86.00 frames

• Huawei's Kirin 9000 is 82.74 frames.

In terms of AI performance, the S888 scored very high. Using UL Procyon to test AI inference, the AI inference was 32228.

Huawei's Kirin 9000 is 12596, and the S888 is almost three times larger than Kirin.

The theoretical value of the S888 is 26 TOPS, which is also higher than the 21 TOPS of the Apple A14.

The Ride platform is used in the field of autonomous driving, so Qualcomm can cut out the X60 5G modem on the S888 and leave more room for NPU. The AI computing power is estimated to reach 30-40 TOPS.

Considering costs and vehicle regulations, Qualcomm won't add much AI computing power, because Qualcomm still has an accelerator, which is similar to the Nvidia A100.

Huawei MDC: The light of domestic goods, where will it go under the blockade

Huawei's autonomous driving computing platform is managed by the MDC Product Division under Vehicle BU.

The AI coprocessor used in MDC is a Shengteng series chip, while the CPU comes from Huawei's Taishan server division, the Kunpeng series chip.

The full name of MDC is Mobile Data Center, Mobile Data Center.

MDC's members are partly from the central hardware department of Huawei, whose main business is developing ARM servers before moving to the field of autonomous driving.

The chip portion of the MDC is still provided by Hisilicon.

MDC currently focuses on two products:

• One is the MDC 210 for L2+

• Another MDC 610, mainly used on L4

The CPU part of the MDC 210 is unknown; the AI processor is an Ascend 310.

The CPU of the MDC 610 is probably the Kunpeng 916, and the AI processor is the Shengteng 610.

The Kunpeng 916, whose internal code name is Hi1616 in Hisilicon, is a 2017 product.

It uses a 32-core ARM A72 parallel design with a minimum power consumption of 75 watts and standard TDP power consumption of 85 watts, which is the benchmark for the Intel Xeon series of server CPUs.

The parameters and internal frame of the HUAWEI Kunpeng 916 are shown above:

The 16nm process is used, which means that SMIC is capable of contract manufacturing.

The more advanced product in the Kunpeng series is the 920, Hi1620, internal code of Hisilicon. It uses a 16-96 core design, an architecture developed by Huawei itself, an ARM v8.2 instruction set, and a 7 nm process.

The Kunpeng 930 plans to use a 5 nm process.

As mentioned above, Huawei MDC's AI processors are mainly Ascend 310 and 610.

According to Huawei's roadmap, the official plan was to launch Shengteng 320, 610, and 920 in 2020, but there is no news until now.

The Shengteng 310 is manufactured using TSMC's 12 nm FFC process and was launched in 2018, so the performance is average, with only 16 TOPS computing power.

Judging from Huawei's official introduction, the Ascend 920 and 610 are both positioned for deep learning training on servers, not for in-vehicle applications.

These two processors have an obvious Cowos multi-memory chip package design. This package is also expensive and not suitable for cost-sensitive areas.

Who is the strongest chip?

Overall review: Among the five major manufacturers, Renesas focuses on high cost performance, and has been supported by automakers at the beginning of the design.

Among Japanese car companies, apart from Nissan, which has a relatively high degree of internationalization, other manufacturers will undoubtedly prefer Renesas' V3U.

Renesas has accumulated a lot in terms of vehicle safety regulations, which German manufacturers are also very concerned about.

Therefore, Renesas, which originated in the automotive semiconductor field, is more favored by Japanese and German manufacturers.

Mobileye has shipped more than 60 million tablets and has a huge user base. American, Korean, and domestic independent brands all tend to favor Mobileye, but currently the launch of the EyeQ series products is too slow.

This is also the reason why many new manufacturers, such as Ideal and NIO, have abandoned the EyeQ platform.

Nvidia has first-class performance. As for the price, in Patriarch Huang's words, “the more you buy, the more you save.”

Emerging car builders pursue high performance. NIO, Ideal, and Xiaopeng also have cash reserves of tens of billions of yuan. Among them, Nvidia is very popular.

The Qualcomm Snapdragon Ride platform is similar to Renesas, focusing on cost performance, and Qualcomm's original factory support is relatively strong.

Currently, Great Wall and a well-known leading car builder have chosen the Ride platform.

The biggest obstacle for Huawei is chip production capacity.

Currently, SMIC's 14nm process is not mature. Judging from financial data, SMIC's 14nm business accounts for only 1% of its revenue.

Currently, SMIC is also being sanctioned by the US, and it is very difficult to increase processes and production capacity.

Even if the blockade is lifted, Huawei will not sell chips separately to the outside world.

Regardless of which platform car companies choose to use, they need sufficient support from the original chip manufacturer.

In this regard, Renesas' senior original engineers are all in Japan, and support is poor.

Nvidia has limited human resources, and it is said that its support is not very friendly.

Qualcomm has experienced years of hard work on mobile devices and is very adaptable to supporting dozens of manufacturers.

Combined with Mobileye's new pace of promotion, I think there is hope that Qualcomm and Renesas will win in the end.