close
tech

Left brain + right brain on Tesla’s intelligent road

99c17ca0-d544-11e9-ba9a-aaaa00151f43.jpg

Editor’s Note: This article is from the WeChat public account “Chip Cookie Puff” (ID: YummyCookiePuff), 36 is authorized to release.”The best time to plant a tree is ten years ago, followed by the present.” Tesla released a new version of the multimedia controller MCU2 (MCU, Media Controller Unit) based on Intel Atom (Apollo Lake) chip last March.In April of this year, Autopilot hardware HW3 (FSD, Full Self Driving) of its own SOC (System On Chip) was released to support the computing power needed for fully automated driving in the future.The multimedia controller MCU and the Autopilot ECU form the brain of Tesla. The latest version of the two hardware began to be used in the Model 3 produced in late April and the Model S/X produced in late May.In Model 3, Tesla divided the two boards of MCU and Autopilot ECU into two layers (as shown below), making the two parts of the original entertainment domain and autopilot domain like the left brain andThe right brain has the same relationship, physically combined to form a vehicle brain component.As shown above, the upper layer of the board is a media controller, using the Intel Atom E3950’s 4-core 2.0Ghz x86 architecture processor, running Tesla’s Linux-based operating system to control the large-screen infotainment function (Tesla)The Model 3 Linux code has not been open sourced according to the GPL open source agreement.The lower board is Autopilot ECU, and the HW3 version consists of two Fiss chips developed by Tesla. The two SOCs are powered by two independent power supplies, sharing sensor data of all cameras input ultrasonic radar, parallel calculation and mutual verification.For security redundancy.The FSD chip is manufactured using Samsung’s 14nm technology. As shown below, one SOC integrates: three 4-core (12 cores) ARM Cotex-A72 architecture 2.2GHz CPU, one 1GHz GPU, two 2GHz NNPs (neural)Network processor).Each NNP has a matrix of 96*96 MACs, 32MB SRAM, and the processing capacity of each NNP is 96*96*2(OPs)*2(GHz)=36.864TOPS, then the single chip is 72 TOPS, including twoThe chip’s Autopilot ECU is 144TOPS.Tesla analysis applies a typical visual perceptual algorithm using CNN convolutional neural networks, where more than 99% of the computational operations are addition and multiplication (see the figure below).By providing specific matrix operation hardware acceleration instructions (DMA read/write, convolution/deconvolution, dot multiplication, etc.) and loading the model into the SPU of the NPU, the NPU can greatly optimize the application of the neural network model to the original GPU.The processor is 21 times faster (on solving specific problems).The CPU and GPU computing power in the SOC will be used to complete the automatic control of the automatic driving, including multi-sensor signal fusion, high-precision positioning, route planning, control execution command delivery and execution result verification.Looking back over the past eight years, Tesla has performed multiple iterations on the intelligent left and right brain (see below).It can be seen that Tesla’s investment in the MCU media controller is less than that of the Autopilot ECU. It is nearly 6 years after the first version of the hardware. After NVIDIA abandoned the Drive CX platform for the smart cockpit, it turned to Intel.s solution.In terms of Autopilot ECU, it is very radical: from the initial perception, Mobileye EyeQ3 is responsible for only simple rule-based control, evolution to self-developed machine-based learning Autopilot software, to application-level hardware optimization and future planning.Fully automatic driving.From the perspective of the end user experience, Tesla has been promoting its functions in advanced assisted driving and automatic driving. Compared with domestic manufacturers, it does not have the convenience of intelligent voice interaction in the car, music navigation service and other applications.There are so many excellent places.In the field of in-vehicle infotainment, thanks to the rapid development of domestic mobile Internet, the smart cabin experience currently developed by domestic manufacturers will be ahead of Tesla; in the field of automatic driving, the domestic manufacturers are currently developing relative to Tesla’s hardware computing power.There will be a gap of one to two years to catch up.At the operating system level, the entertainment domain application is suitable for multi-tasking time-sharing systems (like IOS, Android, windows and other operating systems on mobile phones and PCs), and the auto-driving domain function is suitable for real-time operating systems (analogous fighter control systems).Problems in different fields need to be optimized for their characteristics.In terms of the operating system of the entertainment domain, Tesla chose the Linux route in 2012 when the car Android has not yet developed. However, in the current AndroidAutomotive development, various CP/SP manufacturers have provided Android car version SDK or car application.In a fast-integrated environment, Android Automotive will be the mainstream choice for entertainment domain operating systems.The original text was first published on the WeChat public account: the cookie puff search did not get lost.Newly set up “Smart Car and Internet Technology” group chat, welcome to add group owner WeChat 6080901 to add to the group (into the group please note the company name)..