Including this year’s GTC, the author is the third time in the field to hear the speech of Mr. Jen-Hsun Huang, almost every feeling – Future Technology really fragrant. On this the GTC, so I was the biggest gain, or artificial intelligence for the car industry and affect change. Not only to enhance the computing power of the hardware, the software will create the hardware to maximize the performance, and can be the same strain, to create a platform for intelligent forever. And such a platform for intelligent autopilot and interior sides will bring great progress to.
from the graphics processing unit to the autopilot GPU
reached
scientific and technological achievements are often not easy, like a computer game quot; tech tree quot ;, need step by step to reach one by one, eventually accumulation as a landmark result. However, often led to unexpected major achievement is the development process inadvertently a small step. Automobile tires, for example, vulcanized rubber, vulcanized from canvas shoes outsole, not automobile windshield debris splash tempered glass with laminated glass gas mask is derived from the eye. Automatically calculating the driving unit chips may be traced to computer graphics GPU.
GTC Assembly name stands for GPU Technology Conference, GPU Technology Conference, while the GPU stands for Graphics Processing Unit, which is the graphics processing unit shorthand. GPU as a CPU was first assistant to exist. CPU like a very competent professors, individual combat capability is very strong, very good to him any problems he can solve. The GPU is simply a group of students with computing power, although individual combat capability is not strong, but can quickly process a large number of simple tasks. Due to the development of 3D display technology, a large number of computers required to calculate pixel content, although the calculation of a single pixel is not complicated, but if you let the CPU handle millions of pixels one by one, on the one hand overkill, but it is also time-consuming, inefficient . The ability to deal with difficult professors single task is rolling the students, but for less complex tasks, many people can not stand still, then, is the inherent advantages of GPUappear.
With the development of the times, it was found that the number of GPU processing capacity of many simple tasks, not only for the graphics card can also be used with high predictability and a large number of similar operations, and high-latency, high throughput the architecture calculation. For example autopilot involved in the integration of data and predict the behavior of one of them. GPU and then became the automobile autopilot and entertainment core areas of telematics unit.
seven times the performance of previous generation autopilot computing unit Orin was born
each of GTC General Assembly, and ultimately, the release of heavy hardware. Specializing in the new generation of SoC chip autopilot was named Orin (name derived from the name of DC Comics in Neptune, Nvidia chip named are named in the American comic book character name), and officially released at the conference.
here to explain the meaning of the SoC. Pure electric vehicle State Of Charge (charging state) is different here is the three-letter abbreviation for System-on-a-Chip is. It can be translated into the system-on-chip or system on chip. If the central processing unit (CPU) is the brain, then the SoC is the complex system that includes the brain, heart, eyes and hands.
Orin chip platform built by the 17 billion transistors, the NVIDIA team took four years to build. Orin SoC integrates a new generation of NVIDIA GPU architecture and Arm Hercules CPU core and a new depth learning and computer vision accelerators, can run 200 trillion calculations per second, almost seven times the Xavier generation chip performance on NVIDIA, Xavier can run per second 30 trillion calculations. The chip is scheduled for mass production in 2022 listed.
In fact, early in the 2018 CES show, NVIDIA released the world’s first autonomous machine processor –NVIDIA DRIVE Xavier has been ground-breaking. BAO car studio in 2015, had the honor to participate in the L4 level automatic pilot test carried out first in China to Baidu. Then use the test car platform based on the BMW 3 Series GT transformation from. Select the original 3 Series GTBecause it is has a large trunk, due to the computing system using multiple satellites traditional server-class CPU, causing the volume is too large, only the car’s trunk can barely carry. But with the core Xavier calculation unit, such as the size of only one card, the layout of the vehicle saves a lot of space. Delphi subsequent (after the independent An Bofu Inc.) L4 stage demo car and the forthcoming next year production Xiaopeng Xavier P7 are used as the core of the system, has overcome the shortcomings of earlier calculation unit volume too.
However, with the deepening of the automatic pilot study, the industry constantly updated for the autopilot cognitive difficulty. A number of special cases (Corner Case) emerge, so that the number of sensors and data calculation than the previous estimate of the value of the industry increased dramatically, which means to have greater computing power cycling requirements, coupled with the intelligent car the rapid development of voice recognition and on-board intelligent assistant function is constantly being improved, the requirements of the vehicle operator force is also increasing.
To counter this demand, Orin came into being, although the volume has not changed, but the computing power has increased dramatically. More importantly, in terms of software compatibility and security are also substantial improvements compared.
in the General Assembly speech, Huang said, Orin software systems backward compatible, Xavier previous platform R & D results can be used in Orin platform, its performance will be gain. Like, it will be compatible with previously purchased the same PS2 game than purchasing the Sony PS3, can be described as developers save a lot of development costs.
on security issues, BAO car studio this question specifically Nvidia CEO Jen-Hsun Huang, Orin whether the design fully consider the future of car prices EEA highly integrated architecture design requirements, as well as the resulting security hidden problems. Jen-Hsun Huang answer given is: triple mechanism of security.
The first 1, Orin isolation support virtual machine software, i.e., by isolating the simulator system layer and an application layer;
2, all programs are encrypted access memory;
[ 123] of 3, Orin have a sufficiently fast encryption engine, all written to memory storage and networkNetwork data are encrypted. There are two public key and a private key, the private key can ensure that the machine they’re using a unified, that is, each car is unique. In addition, the communication of the private key is private, secure communication channel, and can protect the computer security, to prevent tampering.
Jen-Hsun Huang told BAO car studio: quot; Orin’s (security) designed as data centers, rather than simply a chip. quot;
software over hardware is accelerating evolution
However, the autopilot is not a simple stack of hardware, the hardware is the foundation, the software is the soul. Autopilot research and development capabilities, from data acquisition to go through, model training, driving simulation and actual on-board computer equipped with these processes. Of which infrastructure, software applications, and artificial intelligence algorithms, NVIDIA are deeply involved in it. Software is one of the main event.
Here we have to introduce NVIDIA DRIVE. It was a autopilot NVIDIA’s industry research and development test platform. Platform includes on-board computer (DRIVE AGX) and a complete reference architecture (DRIVE Hyperion), and data center hosting analog (DRIVE Constellation ™) and the depth of the neural network (DNN) training platform (DGX ™). The platform also includes a rich software development kit (SDK), to accelerate the autonomous vehicles (AV) development. With the help of this platform, developers can develop and validate data on the number of low-cost sensors and algorithms to create modified model with high efficiency.
In this conference, NVIDIA announced that it would transport sector open source NVIDIA DRIVE autonomous vehicles depth of neural networks, the introduction of NVIDIA DRIVE model pre-trained on NGC (NVIDIA GPU Cloud NVIDIA GPU cloud). The practice is currently finished up its autopilot algorithm logic of commercial vehicles and other logistics such as vehicles and other specialty vehicles, these two areas with a normal passenger car automatic driving mode and there is a huge different logic level, NVIDIA has been a move get a response, including Daimler, Volvo and Toyota and other companies.
In addition, NVIDIA also announced the conference, will travel by bit using NVIDIA GPU technology development and other autopilot and cloud computing solutions. Bit in the data center will use NVIDIA GPU training machine learning algorithm, and using NVIDIA DRIVE L4 level for autonomous vehicles provide reasoning skills.
car reasoning software will achieve human-computer dialogue becomes fluent possible
In addition, let the news studio BAO car is more concerned about the conference on reasoning optimization software will launch the seventh TensorRT generation products. TensorRT is provided by NVIDIA for the neural network inference stage acceleration software, computing performance can be significantly improved by AI model provides optimized.
published last year in the GTC China Conference TensorRT 5 supports only CNN (Convolutional Neural Networks, convolution neural network) and 30 kinds of computing transformation, the TensorRT 7 for Transformer (Transformer model is an increasingly popular neural network architecture, this neural network structure after World war StarCraft on human use) and RNN (Recursive neural network recurrent neural network) done a lot of optimization, can with less memory efficient operation, and supports more than 1,000 calculate conversion and optimization. TensorRT 7 fusion operation can be horizontal and vertical directions, a large number RNN configuration designed for developers to automatically generate code pointwise LSTM fusion (when the length of Memories Long Short Term Memory Network) units, even across multiple time steps for integration, and low-precision automatic reasoning do as much as possible. In addition, NVIDIA introducing a core in generation TensorRT 7, RNN can be generated by any of an optimized kernel.
If the previous paragraph is not easily understood, you can skip. In short, the software further optimizes the real-time conversational AI, so the T4 GPU’sIn order to shorten the delay reasoning CPU use 1/10; think of time only 0.3 seconds. This means that the man-machine dialogue within the previous car machine, the long wait time can be shortened to 0.3 seconds, the real man-machine dialogue possible.