Li Ming, Senior Technical Manager of NVIDIA China
Nowadays, with the boom of AI, NVIDIA, which has powerful computing power, has embarked on the fast track of development, and Slogan, the company, has become "the leader of AI computing".
With its deep cultivation in the GPU field, NVIDIA has launched many products in the field of deep learning and artificial intelligence, including various supercomputing platforms and data processing accelerators, and Jetson platform is one of NVIDIA's many products.
Specifically, Jetson is an embedded AI supercomputing platform launched by NVIDIA, which can be deployed on many terminals (possibly cameras, UAVs, robots and unmanned cars, etc.) to provide AI computing capabilities. And "embedded" can solve the problems of bandwidth shortage and delay that these terminals need to have artificial intelligence computing ability.
On March 8 this year, Jetson TX2, a new generation product of the Jetson family, was officially released, which also marks a step forward in the layout of NVIDIA in the field of terminal artificial intelligence.
Jetson TX2
It is understood that Jetson platform has two generations of TK1 and TX1 products before, and has accumulated many customers and application cases, including Cisco Teleconferencing System (Face Recognition, Intelligent Recognition), Farah's factory automation (parts sorting), Toyota's service robots and so on. In China, there are also many users of Jetson platform. Among them, Haikangwei, a well-known security company, has adopted TX1 in its camera products. In addition, the Delivery Unmanned Vehicle in Jingdong is also a customer of TX1. It can be seen that terminal artificial intelligence has penetrated into many aspects of human daily life.
In fact, before TX2 was released, Li Ming, senior technology manager of NVIDIA China, elaborated on the performance details of the new product. At that time, compared with the previous generation, TX2's GPU and CPU were upgraded, memory increased, storage increased, support Wifi and Bluetooth, encoding and decoding support H.265, the size of the same compact; suite has USB interface, SD card interface, HDMI interface, etc., connected to the mouse, keyboard and display, is actually a computer, can carry out people. AI development or daily use.
Jetson TX1 and Jetson TX2 Performance comparison
According to NVIDIA, TX2 provides twice as much performance as the previous version, which means it can operate at more than two times the power and is less than 7.5 watts. This performance enables TX2 to run a larger and deeper neural network in terminal applications, makes terminal devices more intelligent, and achieves higher accuracy in a shorter time when performing tasks such as image classification, navigation and speech recognition.
In addition, in order to facilitate the development of developers based on TX2 platform, NVIDIA has also launched a software development kit of JetPack 3.0, which is equipped with a complete set of development tools and greatly reduces the access threshold for developers.
Because of TX2 performance improvements, many Jetson customers are currently migrating to the platform. Li Ming said that at present, TX2 is better in both upward and downward compatibility, so there is no transfer cost for users who used TX1 to transfer to TX2.
So far, the various advantages of TX2 have been very clear: whether it is doubled performance growth, reduced development threshold, or no platform transfer costs. NVIDIA also has high expectations for the product's market performance.
In order to make the product more convincing, NVIDIA officially held another product interpretation activity more than a month after TX2 was released. It explained some Demo in detail. It also invited relevant leaders and senior technical experts of TX2's representative customer Haikangwei to share their stories about their affiliation with Jetson platform and their relationship with Jetson platform. Technical considerations.
Before demo, Li Ming mentioned NVIDIA's "AI City" concept. In NVIDIA's eyes, the world generates a large number of video streams every moment, which will come from more than 1 billion cameras worldwide in 2020. Faced with huge amounts of data, only deep learning related technology can be used to understand and analyze it, which directly transforms the information at the pixel level into a kind of semantics, or a kind of understanding of the scene. These scenarios include, but are not limited to, intelligent transportation, and there will be more application scenarios in the future, which will eventually cover the whole city and form the so-called "AI City".
How to do it? Li Ming showed two videos of Demo.
When 4K video stream is transmitted to the workstation equipped with Jetson TX2, the AI terminal can interpret the video information in real time, identify vehicles, pedestrians, road signs and other information, and even recognize the color and brand type of vehicles, as well as the gender, age and hands of pedestrians. Whether to mention items and other information.
The whole data processing process follows: video stream input-video decoding-using artificial intelligence means to identify targets (such as license plate, face, etc.) and frame-code, complete local processing-store to the cloud or display on the monitoring screen.
The other is a video of American police patrolling the parking lot. The camera terminal equipped with TX2 can monitor and record all kinds of details of vehicles and population. If suspicious vehicles and population are found in the parking lot, the system will automatically compare with the suspicious model built in-house, and the police will soon be able to lock in the problematic pair. Elephant. Such an efficient means of processing video information is the gospel in the field of security and public security.
Li Ming said that the above "AI City" related Demo source code and Pipline will be packaged into Jetpack 3.0 for users to refer to.
This time, NVIDIA invited Wang Penghe, Director of High Performance Computing Department of Haikang Video Research Institute, and Jiang Chao, a senior technical expert, to be Jetson TX2 platform.
Haikang Video is the world's leading solution provider in the field of security. Last year, they and NVIDIA jointly released a full range of security products based on in-depth learning technology from the back end to the front end. Its binocular intelligent face camera "Deep Eye" (front-end) and video cloud structured server "Falcon" (back-end) all adopt NVIDIA technology solutions.
Wang Peng, Director of High Performance Computing Department of Haikang Visual Academy
Wang Peng said that the reason why he chose to cooperate with NVIDIA before was that he valued its ability in GPU and in-depth learning, and that NVIDIA's ability in this area has improved rapidly in recent years.
Wang Peng said, "Our security development goals in recent years have shifted from"visible"to"visible". In recent years, the enthusiasm of artificial intelligence has also given us a new direction, that is, to understand. "Understanding" actually coincides with NVIDIA's "AI City" concept, so the requirements for computing power and massive video information processing capacity are also higher, so Jetson platform is a very suitable terminal intelligent choice for Haikangwei.
In practice, Wang Peng said that after adopting a set of technical solutions of NVIDIA, the computing performance of "blade" of 1U intelligent processing server, which we are proud of, has reached 16T, and its power consumption in 1U space is only 300 watts. The general purpose server should achieve about 14T performance, and its power consumption should reach more than 8000 watts. Therefore, whether in space or power, using this high-performance chip has great benefits for the cost implementation of the whole back-end deployment of Haikongwei.
"NVIDIA's high-performance in-depth learning engine and end-to-end solutions enable us to land research results on products at the fastest speed and in the shortest time," Wang concluded.
Jiang Chao, a senior technical expert, brought his indoor robot to the scene. In his opinion, when a robot enters a family or business situation, one of the problems it has to solve is "how to walk". This problem leads to three core sub-questions: Does the robot know where it is in space? Does the robot know where the road is? Can the robot receive the user's instructions?
Senior technical expert Jiang Chao (from Orion Star)
Therefore, Jiang Chao believes that vSLAM (visual Simultaneous Localization and Mapping) is actually the bottom and core part of the whole robot technology.
Finally, when the robot walks up, all kinds of sensors will generate relevant information, and control instructions will be sent to the robot continuously, which means that a very complex control network will be formed. In real business scenarios, robots must be able to work right now.
There is a contradiction here - if we adopt "cloud computing" technology or transfer data back to calculate and return, in the current network situation, it can not meet the real-time requirements.
So these complex calculations have to be done on the terminal. "We have to solve problems on the terminal, nearest to the user, so we need a very powerful computing platform," Jiang Chao said.
NVIDIA's TX1 provides good computing power, but computing power is always insufficient. When Jiang Chao applied it, he found some limitations. "So when I see TX2, I'm very excited, because the overall performance of TX2 is twice as much as that of TX1, so there's room for us to do more." The so-called "more things" include applications such as "semantic maps".
Jiang Chao also mentioned that TX2 can support not only the hardware itself, but also the very good software - VisionWorks (included in Jetpack 3.0). This software actually implements a lot of bottom operations of machine vision on OpenVX standard, and also provides a programming framework. "We can use these things to do it." Make rapid development.
In addition, TX2 can also be used for SLAM back-end optimization, because some optimization tools running on the CPU will have very limited performance.