"Xinzhi Driver Press: This article comes from the future travel service provider Xin Yuexing, Joint Foundation & CEO Xu Chao, Joint Foundation & CTO Li Linfeng's Technical Details. New Enciprome Row current business line includes new energy vehicles and L3 unmanned driving integration Program. In April this year, Xin Yue Walk released an independent development of WiseadCu driverless computing control unit. In this article, the author disassembles the Teslaap 2.0 / 2.5 operation unit and combines the previous international advanced unmanned driving operation. The platform analysis report of the control unit is shared.
First look at a set of pictures, this is the standard configuration of almost all host plants and unmanned emerging teams:
Miscellaneous trunk
So in usually, what will this? The industry's usual standard is the industrial control computer (GPU computing card, CAN card), UPS power supply / regulator power supply, switch, low voltage power distributor, GNSS / IMU module, vehicle control unit (usually dspace microautobox), heat dissipation and cooling mechanism Wait. If some specific sensors are used, there will be a separate industrial computer, a fusion, an interface, and a power module. Plus the interconnected cable, high and low-voltage cable, and HMI and debugged interface, which is not as a small businesses.
It is because of this, Xin Yizhi will design his Wiseadcu, just as Tesla's autopilot arithmetic control unit, AUDI's ZFAS, etc. Automatic driving from laboratory to mass production, inevitably reducing volume, cost and power consumption while lifting reliability. Of course, NVIDIA provides the Drive Series, and NXP also has Bluebox, but these are just "development tools", while mainstream parts suppliers announced their arithmetic control unit or domain controller, for the host factory, is just a black box The existence, and it is likely to bundle its chassis electronic related products. Xin Yizhi's Wiseadcu is an open hardware and software cooperation model to "empower" for the host factory and unmanned team, so that it fully masters the hardware and software planning and definition of unmanned driving control system.
First, NVIDIA Drive PX2
First analysis is Drive PX2 AutoChauffeur development board, which is the latest generation of the NVIDIA Drive Series:
NVIDIA's Drive Series
Although Xavier-based next-generation SOC has been made public, the conservative estimation of Q2 will be opened by Q2, which is the opportunity to see the car grade, based on the NAVER's product rhythm.
NVIDIA's Drive PX2 version
In DRIVE PX2, AutoChauffe is a defined version of the L3 version, the configuration of its computing part is the PASCAL architectural independent operation unit of dual PARKER SOC plus double MXM3.1 interface, and lower autocruiser is a single Parker SoC, more The high level of Full Autonomy is composed of two autochauffeur. AutoCHAUFFEUR and AutoCruiser have a functional security control unit that Infineon's Tricore Aurix TC297 as ASIL-D.
Drive PX2 AutoChauffeur Difference
In fact, AutoChauffeur also has at least two different versions that differ from the interconnection mode between PARKER, one of which uses Altera's Cyclone V FPGA + ARM SOC and Ethernet, and the other uses PLX PEX8724 non-transparent PCIe swap chips and Ethernet interconnects.
Drive px2 autochaufffeur block diagram
Because there is no FPGA version, there is no logic of FPGA, you can only share the case of the PEX version, simply, use the PEX version, you can hang the independent Pascal GPU on any Parker So, using PCIe 4X interconnected.
front
back
12-way FHD camera, using MAixm custom GMSL-CSI2 conversion chip on PX2, which is a non-public material.
The standard Drive PX2 is relatively disclosed, although we have the hardware of Drive PX2, but because of the confidentiality agreement and relevant legal terms, the above analysis is substantially analyzed, but you can find the source of information disclosed in the Internet.
Sentimental Test Tesla (Tesla) is working with AMD to build its own unmanned chip, of course, to create their own chips should also be the inevitable choice of Tesla. But as a person who has worked in Agricultural Enterprise (AMD), combined with the actual application and understanding of the DRIVE series platform, in particular the DRIVEWORKS-related framework, the accumulation of farmers in the automatic driving field is more than 3. ,5 years. OpenCl is still very pale in front of the M-shaped DRIVEWORKS framework and mass-based Cuda-based algorithm. In fact, farmers have long-term accumulation in the field of industrial and enterprises. I have previously taken a GEODE series acquired by Natsemi after Cyrix. There are also farmers' own AM186 ~ AM486 series until today is still available today. Just in front of the survival, facing the toothpaste factor and the two-squeezed enterprises, a business that is not too dare to touch, and how much energy and courage go to the unmanned driving area?
Say, in most of the current sensors, there must be certain algorithm processing capabilities, and can do results. If it is not extreme cost optimization and control requirements, is it necessary to be severely dependent on CUDA, OpenCL, and various CV acceleration?
Second, Tesla Autopilot 2.0
Make some background paving, everyone is most concerned about Tesla's AutoPilot operation unit analysis, and there will be a description of Drive PX2, because you have seen many domestic and foreign articles and analysis, everyone will Tesla AutoPilot hardware It is directly called "NVIDIA DRIVE PX2", which is inaccurate. At least AutoPilot 2.0 and 2.5 are not a simple copy of any version of Drive PX2.
In short, autopilot 2.0 is basically half of autochauffeur, or AutoCruiser plus a PASCAL independent operation unit while adding GNSS receiving chips. And AutoPilot 2.5 is AutoChauffeur removes a PASCAL independent operation unit, and the other is turned from the MXM card into onboard, adds a set of integrated systems based on an intelral / navigation board for an Intel chip and NXP MCU while using TBOX and GNSS. Receive the integration on the board.
Autopilot 2.0 appearance, double fan
Fan under the cooling
Autopilot2.0 motherboard front
Autopilot2.0 motherboard back
And Drive PX2 AutoChauffeur compared, autopilot 2.0 removes the PEX PCIE exchange chip, adds the Ublox NEO-M8L GNSS receiving chip, replacing the Maxim's GMSL camera interface with Ti's FlatLink III chip, becomes 6-way LVDS from 12-channel GMSL Keep two GMSL video outputs and all the way HUD output. Increased all the way 22W mono Class D amplifier and a three-dimensional CODEC, 9 channel CAN bus, where there is no welding receiver. The three Ethernet is retained. There is no welding all the way, and the two channels do not use Broadr Reach, but the standard AVB, and the heat sink is added to double the fan wind cooling.
Simply listed the model of the main chip:
1, NVIDIA "PARKER" P94W97.01P TA795SA-A2, PARKER SOC master, built-in 256 CUDA unit, 4 core A57 64-bit ARM and 2 core Danver 64-bit ARM
2, four Samsung DRAM K4F8E3S4HBMHCJ
3, NVIDIA GP106-505-KC's MXM card, 4GB GDDR memory, reserved 4 pads, up to 8GB, belonging to the GP106 series to remove the display section, 1280 CUDA operation units
4, Infineon Tricore Autrix TC297TX-128 AUTRIX TC297TX-128 Asil-D MCU5, Ublox Neo-M8L GNSS Receive Module 6, Toshiba EMMC and Spansion Nor Flash7, Marvell 88EA1512 AVB / Ethernet Transceiver 8, Marvell 88EA6321 7 Port AVB Switching Chip 9, Maxim Max9260 GMSL display output
10, Ti DS90UB964 LVDS camera input
Combined with the actual application and performance evaluation of DriveWorks, the hardware architecture such as Autopilot 2.0 can complete a few levels of automatic driving? Here, do some analysis and decomposition.
AP2.0 interface analysis
First, the CPU part, 4xa57 + 2xdenver + AURIX, its calculation capacity is sufficient to handle automatic driving-related HMI, bus communication, resulting output sensor fusion, GNSS data processing, and vehicle motion control, about half of redundancy.
The GPU is 256 + 1280 CUDA units, with 6 cameras, four as a lounge and SFM, two front views for lane lines, pedestrians / target detection, traffic identification detection, and forward space. Testing, according to the code provided by DriveWorks, its ability is not enough to meet the requirements of simultaneous operations.
SFM is put on one side, the lane line, the target, and traffic identification detection is about 90% of the GPU, and the GPU, number, and identification of these objectives, and may be further decomposed on the scene. Not tracking and identifying all the goals, this is an important difference between Tesla production autopilot and NVidia Drive development platform, but in any case, 1536 CUDAs even for the GP106-505 optimized GP106-505, we judge whether to rely on AutoPilot's computing power It is still difficult to achieve complete L3.
In fact, if it is divided by Tesla and Mobileye breaks up, the latter has handled most of the front visual related part, and AP2.0 should complete L3.
The onboard NEO-M8L GNSS reception chip is a single-frequency multimode 2.5m CEP automatic driving accuracy module, even with IMU, it should not reach the accuracy and reliability of high-precision positioning navigation requirements.
Third, Tesla Autopilot 2.5
For Tesla AutoPilot 2.5, Xinyue is always called AP3 or AP nextgen, when we get this unit, we don't know its specific model, but according to architecture and morphology, it should be used on Model 3. Foreign includes Tesla official forum called Autopilot 2.5.
Autopilot 2.5, water-cooled Autopilot 2.5 shell decomposition, we analyze that AP2.5 is used for Model 3, because this unit is in front of a total of two plates, one and ap2.0, but there is a Parker The automatic driving control panel, and the other is an Intel-based card with SPC5748G MCU with SPC5748G MCU, with Telit Modem and LG's BT / WLAN module and the 8-channel Class SB-i digital amplifier. From MODEL 3, it canceled the instrument, and the information entertainment system and instrument were done, then SPC5748G is the MCU that handles functionally safe, that is, the guarantee of the reliability of the instrument system. And Intel's SOC is to do an entertainment navigation or some data processing. The interior of the unit of the two circuit boards is not physically connected, but the water-cooled heat dissipation portion is integrated, and integrated into the same metal casing, no radical heat portion, the body portion of the body is not more than AP2.0 too much. AP2.5 Multi-Angle View The motherboard of the motherboard and information entertainment part of the automatic driving part is analyzed one by one.
AP2.5 automatic driving motherboard front
AP2.5 automatic driving motherboard back
Interface part, because AP2.5 (perhaps because we are not a Release version) labeled all interface features, so don't want AP2.0 to speculate. According to the interface, the motherboard supports the following interface:
1, GPS antenna
2, Rear Cam post camera
3, Selfie built-in driver camera
4, main CAM front main camera
5, Repeater two forwarding
6, b-pillar b column two-way camera
7, Fisheye Narrow two narrow fish eye camera
8, a series of IO, CAN and power
9, single Ethernet
10, USB
11, MCU external interface
The main chips are as follows:
1, two NVIDIA "PARKER" PC5S58H.s
8P TA795SA-A2, PARKER SOC master, built-in 256 CUDA unit, 4 core A57 64-bit ARM and 2 core Danver 64-bit ARM
2, six Samsung Dram K4F8E3S4HBMHCJ, four of them give A, two to B
3, NVIDIA GP106-510-KC onboard chip, 4GB GDDR memory, should be an improved model of AP2.0
4, Infineon Tricore Autrix TC297TX-128 ASIL-D MCU
5, Ublox NEO-M8L GNSS Receive Module
6, a Toshiba Emmc gives two A and Spansion Nor Flash, AB is 7, Marvell88AE1512 AVB / Ethernet transceiver 8, Marvell 88EA6321 7-port AVB switched chip 9, three Ti DS90UB964 LVDS camera input 10, a Ti DS90UB954 LVDS camera input
11, the same TAS5421 mono digital amplifier like AP2.0 and TLV320AIC3104 stereo CODEC;
12, FTDI 1647-C serial port transfer USB debug chip
13, several CAN transceivers
Cancel the obvious display output section, is it handed over to the information entertainment motherboard?
AP2.5 information entertainment motherboard front
AP2.5 info entertainment motherboard back
Interface part:
1, FTDI USB debugging
2, WLAN antenna
3, BT antenna
4, LTE USB debugging
5, CAM IN camera
6, Cam Out camera out
7, BMP debugging port
8, BMP LPC debugging port, talk about LPC (Low Pin Count), basically think that the INTEL's chip is the X86 series.
9, PMIC debug interface
10, GSM switch
11, Broadreach network port
12, CAN / POWER interface
13, Ethernet network port
14, AUDIO interface
15, USB interface
The main chips are as follows:
1, a module with Intel confidential in 4G memory
2, there is a Nor Flash and an EMMC near the module, you can understand BIOS / EFI and hard drives
3, NXP SPC5748GSMMJ6
4, LG INNOTEK B216C BT / WLAN module
5, Marvell 88ea6321 7-port AVB swap chip
6, Ti DS90UB949 Serialization and Ti DS90UB940 Square Chip Each
7, there is also an unknown WQ1214CS chip
8, two St TDA7802 Class-SB-i digital power amplifier, each containing a four-way 28 watt standard to 72 watt peaks. A total of 8 channels
9, a tja1059, a tja1043
10, and a 4G module slot similar to MiniPCIE
From the audio configuration and SPC5748G and interface configuration, this is basically that the information entertainment system unit has not ran, but why there is no obvious display interface above? Is it like a set of Tegra 3 as a set of Tegra 3 like a set of Tegra 3 in the LCD screen of Model 3?
As mentioned earlier, AP2.0 is PARKER + PASCAL, and AP2.5 is dual PARKER + PASCAL in the automatic driving part. It is actually the rest before our understanding, the Parker itself has the rest, and 256 CUDA What changes can increase? According to our analysis, this can actually bring a lot of benefits, with an independent PARKER to handle tasks that are not high in association, avoid competition with the main operational tasks, reduce the overhead of switching and resource scheduling. This should be an improvement that Tesla made from the actual application angle.
Model S's display unit, T301 should be the Version of TEGRA 3
Back to AP2.5, just mentioned that there is a modem slot on the information system board, this MODEM is long?
AP2.5 MODEM module
This module uses Telit's LE940B6-NA car regulatory module, and has an 88EA1512 network chip. From the shell structure, this module can support hot swap. In this, Tesla's engineers should be considered. First, the USB can be hotly swapped, but it is worse than the Ethernet. And use AVB Ethernet, you can flexibly configure accessibility to the gateway and other car modules.
In the previous generation of information entertainment systems, it should be a USB link.
Communication module of the previous generation of information entertainment systems, integrated LEA-6R GPS and a very special GYRO
Continue to return to the AP2.5 information, there is a SIM card and TF card slot on the motherboard. The SIM card should be a card from ICCID. There is only one partition in the TF card, and some logs have been stored. There is no map and other data.
Of course, the analysis is still continuing, we have some interesting "cross-boundary tools", such as ...
Repair mobile phone must have a good product
Write here, I want to say, no matter what is analyzed for PX2 or AP2.0 or 2.5, there will be some biased. After all, it is not a primitive design, but this kind of gidding analysis, how much can be obtained in conjunction with the industry experience in the chip industry. Less useful information.
Tesla's several generations of products are quite good from the PCB process, all of which are 12-layer blind holes, many connectors and chips are custom models, such as MXM, Tesla design, with locks and brackets, The connection is quite stable, and such a connector can also be seen on the ZFAS generation prototype.
From the perspective of the interface, Tesla has a large number of AVBs and LVDs (including FPDLINK3 and GMSL, but also reserved CAN, LIN, etc., as the automatic driving control unit of many large-scale mass production, regardless of it The actual implementation effect is actually practiced, and this series of control units are the perfect art of art from design or manufacturing processes.
Fourth, from Wiseadcu to discuss the design of unmanned driving domain
The future driver's domain controller, its design concept should be the aforementioned "high reliable, high performance, low cost, low power consumption, small volume" high integrated hardware and software platform, while chip companies, algorithm teams, parts enterprises And the divide existed between the main plant needs further filling.
First, from the various automatic driving and driver's system architectures, environmental perceptions and fusion develop have been very rapid, but everyone ignores the difference between the motion control and line control of the vehicle itself. More than once in a very professional conference and organization, I heard the "vehicle kinetic model and simulation are unnecessary" "environment sense blending gives the guiding curve, and the line control can be implemented" such as "actually Just drone in the laboratory stage.
As the industrialized landing team of Jigda Automotive Engineering, Xin Yizhi, we deeply appreciate that simply puts the vehicle as a centrifuge rigid body is a very simple understanding. The body attitude and motion status are one level. The environment sensor is a level. The ability of the vehicle itself is also a very important level. If the "capabilities" of the actuator are not intended to consider the status of the actuator in decision conditions, it is impossible to complete accuracy. Reliable automatic driving system, it is more unable to meet the two automatic driving and driver's basic needs of "safe and comfort".
Xin Yizhi Westeadcu-L4V1
Second, the hardware and hardware architecture of the future driverless domain control should be fully planned, which can be divided into two main factions from the system architecture, which is domain control algorithm and domain control weak algorithm. It distins very similar distributed operations and virtualized centralized calculations.
Combined with new pleasion, you can design the history and understanding of Wiseadcu, do some sharing.
The domain control of the strong algorithm can reduce the complexity and cost of the front-end sensor, more efficiently use the domain control operational resources and capabilities, which makes the fusion of various raw data more reasonable, and can make the physical architecture clearer. Its shortcomings or challenges are also very obvious:
1, understanding, integration, optimization, and resource rational distribution and dispatch, very testing system architects, and the implementation of engineer team;
2, compared to the ASIC to process specific calculations and data, use CPU and GPGPU to make operations, not necessarily the optimal situation;
3. Raw Data is more demanding on bandwidth and real-time requirements;
4, the increase in system complexity increases the chances and influences of error, and the malfunction of the original single sensor may enhance the overall system failure.
On the domain control of the Audi ZFAS (non-first generation) weak algorithm of the Mobileye Eyeq3 chip, in short, everyone is doing their respective things. According to the unified agreed and agreement, the results of the target detection are in the original data. The integration is extremely low, there is no pre-rear connection, depending on the unified timestamp to do instant operations forming decision-making results, some sensors itself relies on the body attitude and related state, requires multiple system redundancy. Considering such a problem, consider designing an extensible architecture, such as Xinyue Wiseadcu joins the FPGA unit, and the front-end data processing and data access to the sensor can be flexible, and the vehicle status and attitude have a unified protocol and Data interface. At the same time, it is compatible with the processes of the original data and the result data. In addition to providing detection result data, the front-end sensor provides an original data, and WiseAdcu takes out some specific information in accordance with the original data, and does not need to involve very professional algorithms. Almost all of the data that can be collected by the status of the "domain control", form a reasonable "annular-vehicle-weight-fusion-decision-execution" large closed loop. Xin Yue District Wiseadcu hardware block diagram Wiseadcu hardware architecture is based on the above concept. Xinyue Zhikong CTO Li Linfeng calls its Wuhan Haimai Technology Core Auto Electronics team averaged hardware and driver engineers to participate, and Xinyue engineers Complete the conversion of the concept to the actual product. From the first edition of the case, there are only more than 30 Minor Bugs, which is very difficult for a more than 600-multi-device, more than 9000 pads of complex systems. Network architecture made based on Xin Yue Walk WiseadCu Domain Control
From some subdivision:
Network architecture, AVB-based TSN will be an important architecture, but it is guaranteed because of the TSN to give up CAN / Lin, for some maturity and simple scenes, CAN / Lin regardless of cost or reliability, it is guaranteed.
Xinyu Zhiking Wiseadof System Architecture
The operating system, the Ubuntu Linux and its kernels of the Drive Series Development Tools need to do in-depth transformation and optimization, and believe that Tesla has done and has been doing this job. First, it is not possible to completely negate Linux reliability. Many people talk about Linux, think that it is worse than certain commercial real-time systems, from the macro level, the Linux system kernel has developed to 4.x version, the code of its code is indeed More than certain specific commercial real-time systems, but from a microscopic perspective, a simple kernel, if you don't consider driving, file systems, and protocol stacks, it is actually not too much.
Information security, the safety of the sensor itself, the safety of the transport layer, and the system security are needed to consider. As an old driver involved in the network security and embedded system in 1997, he has heard a very classic "Fortress is often from the inside. Break. This is a systematic engineering, and it can make a series of long-term big interlaces alone.
Vehicle control, the functional security of Asil-D, Dano Xi City, Dan, once as the old employees of Fujitsu / Spans, have made a very detailed explanation of Asil-d and Lockstep related knowledge, the old driver is also Wiseadcu Core expert and think tank. Coupled with Jida Chassis and Motion Control Country Key Laboratory, and the practice of vehicle movement control, Wiseadcu follow-up and the depth integration of various mainstream vehicle simulation software are also in line with future mainstream automatic driving domain control trend. WISEADCU uses the ASIL-D MCU hardware with the Autosar specification system architecture, and implements a variety of simulation software interfaces. Its main purpose is to replace Microautobox and other interface hardware in the scale application phase.
In general, the future is not a driving domain control, there must be a good overall architecture design, a complete and meticulous block diagram, and a team of actual experience, the pragmatic work of the awe, open mentality, cross-border exchange, and strength team can design and land Depth cross-border products and maintain its continuous continuation and evolution.
Original link: https://www.eeboard.com/news/teslaap2-02-5/
Search for the panel network, pay attention, daily update development board, intelligent hardware, open source hardware, activity and other information can make you master. Recommended attention!
[WeChat scanning picture can be paid directly] "
Our other product:
