Analysis of BMW's autonomous driving computing architecture: Why is EyeQ5 so late?

Mobileye or Intel's EyeQ5 was first announced in May 2016, claiming that engineering samples would be available in 2018 and mass production in 2019. In December 2017, Intel acquired Mobileye and announced that mass production would begin in 2020.

Nvidia's Xavier was revealed in September 2016, and in March 2017, it announced a partnership with Bosch to develop an autonomous driving controller box. It was officially announced at the CES conference in January 2018, and mass production began in June 2018. Obviously, Xavier's engineering samples should have been available by the end of 2016.

In 2019, Intel announced that EyeQ5 would be put into mass production in March 2021, and the mass-produced car Xiaopeng P7 using Xavier has already been launched. Even if everything goes well, mass-produced cars using EyeQ5 will not be launched until 2022. The first car to use EyeQ5 is BMW's iNext.

Nvidia's main partners are Bosch, ZF, Volvo, Mercedes-Benz, Audi, and Toyota. Intel's main partners are BMW and NIO.

In mid-2019, a coalition of 11 companies including Aptiv, Audi, Baidu, BMW, Continental, Daimler, Fiat Chrysler, Here, Infineon, Intel and Volkswagen published a white paper (Safety First For Automated Driving, SaFAD).

In February 2020, Toyota, Denso, Valeo, Mazda and Nissan of Japan also held a joint meeting with the SaFAD Alliance. The SaFAD Alliance is leading an ISO standard, ISO Draft Technology Report (DTR) 4804, in an attempt to create an industry standard for autonomous driving that will replace ISO26262.

The chairman of the ISO 4804 standard is Simon Fürst, the principal expert in BMW’s autonomous driving. At the end of April, EE Times interviewed Simon Fürst, https://www.eetimes.com/unveiled-bmws-scalable-av-architecture/, and revealed a lot of information about BMW’s autonomous driving computing system.

The above picture shows the BMW intelligent driving hardware system roadmap. In fact, this roadmap was formulated as early as the beginning of 2017, when BMW decided to cooperate with Intel (Mobileye). BMW's logic is to maximize the reuse of research results, reduce R&D costs, and adopt modular design so that it can be flexibly added or deleted. 

According to this picture, we can see that the basic module of BMW is L1 level using EyeQ4 plus Infineon's Aurix MCU. L2 uses Intel's denverton dual-core CPU plus EyeQ5 plus two Infineon's Aurix MCUs. Generally speaking, L3 systems do not have Fallback, but BMW still added a set of Fallback (completely independent, including power supply and wire control execution system). BMW's L3 Fallback system is the L2 system. The same is true for the L4 system. The main system of L3 is a CV version of EyeQ5 and an open version of high-end EyeQ5, plus two Intel denverton 8-core CPUs. Infineon's Aurix MCU is of course still there.

In terms of sensors , a forward-facing laser radar is added , which should be the second generation of Valeo Scala. L4 uses a 24-core Xeon processor to replace the two 8-core processors of L3, and adds an open version of the high-end EyeQ5. In terms of sensors, side and rear laser radars are added. The L2 system is called mPAD, L3 is hPAD, and L4 is uPAD. BMW publicly displayed these controller boxes at the end of 2018. Both L3 and L4 systems are water-cooled.

The picture above shows the mPAD that Aptiv built for BMW

The above picture shows the internal disassembly of mPAD, with Intel CPU on the back.

However, according to the article interviewing Simon Fürst, it is quite different from the roadmap provided by BMW. The platform's baseline uses Infineon's Aurix and Renesas' R-CAR SoCs to optimize its application in stereo front cameras. According to the article, the basic system of BMW's intelligent driving is a forward stereo binocular system composed of Renesas ' H3 and V3H. In the roadmap, Renesas' H3 and V3H are only the automatic parking part. 

As we all know, BMW seems to have given up stereo binocular and switched to trinocular system since 2018. This is very confusing. There are several possibilities. The first is that the interview article is completely wrong and the roadmap is right. The second is that the interview article is right and the roadmap is wrong, but the error is not much. BMW combines EyeQ5's trinocular and Renesas's stereo binocular, and BMW has gone back to stereo binocular. The third is that both the interview article and the roadmap are wrong, and BMW is not telling the truth.

I think the second possibility is more likely. Although BMW is not as advanced as Mercedes-Benz in stereo and binocular technology, it has also accumulated a lot of experience. However, the first possibility also exists. Mobileye has always emphasized that monocular can also perform stereo and binocular depth calculations and can completely replace binocular. The possibility of using binocular in the Mobileye system is very low.

BMW L3/L4 Intelligent Driving Software Architecture

BMW's L3/L4 main system calculates the path, and the Fallback system monitors the main system. When it is learned that the path calculated by the main system will cause an accident or collision, the Fallback system will switch to the main system. The main system uses artificial intelligence's non-deterministic algorithm, and the Fallback system uses a classic deterministic algorithm to ensure safety. Simon Fürst believes that ASIL B level is sufficient in most places, and D level is only used in very few places. For perception, BMW believes that the key is not to detect the target, but to predict the target's movement trajectory and calculate the operative space based on the trajectory.

These are exactly what LiDAR and binocular cameras are best at (optical flow method). For trajectory planning, BMW mainly relies on LiDAR to estimate road curvature. Maps produced by LiDAR can also provide curvature, as is the case with Cadillac's Super Cruise. LiDAR high-precision maps also help with positioning. Simon Fürst did not mention Mobileye's visual positioning REM. Simon Fürst disagrees with sensor fusion. He believes that sensor fusion is currently only in the scientific research stage and is far from the practical stage. This is consistent with Tesla's view that Tesla is a pure visual system without sensor fusion. The industry still needs to build a fundamental understanding on how different algorithms should apply to different sensor modalities. Sensor fusion currently has little improvement in performance, and sometimes may cause more missed reports, which reduces safety, but greatly increases cost and complexity.

BMW Intelligent Driving Safety Level

Putting aside the question of whether it is binocular or not, let's look at Intel's CPU. Denverton is the edge computing Atom C3XXX series CPU launched by Intel in the second half of 2016. It uses a 14nm process and is codenamed "Denverton" and "Denverton-NS". The C3XXX series is divided into three target markets, one is server and cloud storage, one is network and enterprise storage, and the last one is IoT. The C3XXX series does not have products for vehicles. Intel's vehicle-mounted Atom is the A39X0 series. However, the maximum operating temperature of the IoT series is 85 degrees, which can barely be considered as a vehicle specification. Therefore, BMW uses water cooling instead of natural cooling.

Since Tesla disdained the car regulations, traditional car manufacturers have also wavered. EyeQ5 has not passed the functional safety certification. Xavier has applied for functional safety certification for more than two years, but the official can only say Target. So it is not surprising to use the C3XXX series. Obviously, the dual-core Denverton is C3308, and the octa-core is C3708.

C3XXX series platform internal framework diagram

The reason why BMW uses C3XXX series chips for non-vehicle applications is mainly because it values ​​the Ethernet MAC and switching functions of C3XXX. The target market of C3XXX series is mainly CPE.

The picture shows the LAN controller part of the C3XXX series. The dual-core has 4 2.5G AMCs, and the 8-core has 4 10G MACs. In the L3/L4 computing system, in addition to the main chip, the most expensive is the Ethernet switch chip. If there are many ports and 10G level, and the performance is strong, the price of the Ethernet switch chip may exceed the main chip such as EyeQ5. If the C3XXX series CPU is used well and there are not many sensors, the expensive Ethernet switch chip can be omitted.

Of course, the Phy physical layer is still indispensable. Intel has a complete product line, including Ethernet switching and physical layer chips, but none of them are automotive-grade. As for the 24-core Xeon processor, the selection range is very narrow. Most 24-core Xeon processors have a power consumption of more than 160 watts, and the standard retail price is more than $4,500. BMW can only choose the Xeon Gold 6262V, which has the lowest power consumption among the 24-core Xeons, only 135 watts, and a standard retail price of $2,900. However, Intel's Atom P5962B for 5G servers, which was just launched in April 2020, is also OK. This is the first Atom processor based on the 10nm process and the new Tremont architecture. The maximum operating temperature is 85 degrees, barely reaching the lowest level of automotive regulations. Power consumption parameters are not provided, and it is expected to be around 80 watts. R-CAR V3H was launched in February 2018 and mass-produced in the third quarter of 2019. It contains 4 A53 cores and one Cortex-R7@0.8GHz core, and also reaches ASIL B level.

Renesas uses three types of accelerators. One is the IMP-X5 accelerator based on the pipeline engine, which has pipeline calculations for fixed functions, similar to a GPU. There is also a computer vision engine CVE, which uses a programmable computer vision engine to minimize floating-point operations. Another is a hardcore accelerator, including stereo disparity and optical flow for binocular purposes. There is also a clusterer. The last is a CNN accelerator similar to a multi-core DSP.