Hello everyone, I'm Brother Hu. After a period of sorting out, I also did some research on how to find suitable application scenarios for TX1/TX2 products that seem to be outdated. Now, because many of the materials are relatively old, and some expressions are also somewhat inconsistent, so here we will focus on introducing and comparing Jetson TX1 and JetsonTX2.
Table of contents
1. Why edge computing and terminal intelligence will become a key direction in the future
2. Introduction of Jetson TX1 /TX2
2.1 Advantages of the Jetson platform
2.3 The predecessor of Jetson tx1
1. Why edge computing and terminal intelligence will become a key direction in the future
-
The first point is the issue of bandwidth. According to Nvidia statistics, by 2020, there will be 1 billion cameras in the world, and massive amounts of data will be collected every day, which will be an unbearable pressure for both the Internet and storage. The best solution is to extract the required information at the front end, near the camera, or inside the camera. The best way is to filter out junk information and keep only valid data.
-
The second point is latency. This is more important for robotics and automation than artificial intelligence. The reaction time of the control robot is often in milliseconds. If the time is very long, such as transmitting the things seen in front to the cloud through images, and then returning one more time, the whole process is often more than 200 milliseconds. This is not available for industrial production lines.
-
The third point is privacy. Taking medical care as an example, patients do not want their information to be disseminated and known to everyone. Generally speaking, in the construction of a medical network, only private cloud or local storage is considered. So in this case, how to carry out remote medical assistance? If information is placed on the Internet, it can then be hacked and the entire system can collapse. In this case, artificial intelligence can be used to extract some intermediate results, encrypt and transmit them. This will ensure privacy.
-
The fourth point is the question of feasibility. We are transmitting a 1080P image. After repeated compression, the transmission also requires a bandwidth of 8Mbps. If I want to see the operation of the front-end drone, or the operation of the robot, it must have a bandwidth of 8Mbps. However, in currently inhabited areas around the world, the network bandwidth of more than 50% of the areas is below 8Mbps. In most places where there are no people, there is not even a 3G network. What should I do when the most basic 8Mbps bandwidth cannot be reached? It is possible to make the terminal robot intelligent, let it handle some simple things by itself, get back useful information, and then process it.
It can be seen that the intelligence of the terminal (or front end) is very important. The Jetson series also has many classic use cases at the application level. For example, Cisco's video and teleconferencing system (face recognition, voice recognition), Farah's factory automation (intelligent recognition, sorting of parts), Toyota's service robots, and so on.
2. Introduction of Jetson TX1 /TX2
2.1 Advantages of the Jetson platform
Jetson is an open platform with complete equipment and lower barriers to entry for developers. For enterprises, startups, and ordinary developers involved in AI business, anyone can use it to develop artificial intelligence solutions for terminal applications. TX1/TX2 are mainly deployed on terminal applications, including intelligent factory robots, commercial drones and smart cameras, etc. The Jetson series is also aimed at promoting the intelligence of terminals.
2.2 What is Jetson tx1
tx1 is an embedded computer based on NVIDIA TegraX1 platform. It is an artificial intelligence supercomputing module launched by NVIDIA, with 256CUDA cores and 4GB LPDDR4 memory. tx1 supports up to 6 camera inputs and a variety of communication interfaces, such as USB, Ethernet, HDMI, etc., and can be widely used in artificial intelligence image recognition, robotics, automatic driving, drones, VR/AR and other fields.
2.3 The predecessor of Jetson tx1
tx1 is one of the embedded computing module series products launched by NVIDIA. Its predecessor can be traced back to the TegraK1 chip launched in 2014. This chip adopts NVIDIA's unique Kepler architecture and uses a 28nm process. It contains at least 192 CUDA cores for security. The computing power of Jetson TK1 (Tegra K1) is: 326 GFLOPS (this is maximum and operating with built-in fan). Later, NVIDIA launched the JetsonTK1 development board based on the TegraK1 chip, which is an embedded computer with built-in TegraK1 SoC, 2GLPDDR3 RAM and 16GeMMC storage, as well as multiple interfaces and expansion structures for easy development and application. JetsonTK1 development board has been widely used in robot vision, automatic driving, medical diagnosis and other fields.
2.4 Evolution of tx1
tx1 is an evolutionary version based on JetsonTK1, which uses NVIDIA's self-protected TegraX1 chip, which uses a 20nm process and has a higher energy efficiency ratio and stronger computing performance. tx1 is not only a supercomputing module, but also a complete intelligent development platform, including TX1 module, JetsonCarrier board, Linux operating system and development tools, etc. It can not only meet complex computing needs, but also easily connect various sensors and peripheral devices to create an intelligent computing ecosystem. The single-precision floating-point computing capability of the Jetson TX1 GPU module has been improved to 1 Teraflops. It is also said that the floating-point algorithm is actually only 472 GFLOPS (FP16)
2.5 Jetson tx2 upgrade
The TX2 offers twice the performance of the tx1, which means it can run at more than twice the power efficiency and consume less than 7.5 watts. Such performance allows TX2 to run larger and deeper neural networks on terminal applications, making terminal devices more intelligent, and at the same time achieve higher performance in a shorter period of time when performing tasks such as image classification, navigation and speech recognition. precision. The effective computing power is: the computing power of single-precision floating-point numbers reaches 1.26TFlops. Why is the performance 2 times higher than TX1? In fact, we are talking about the comprehensive power efficiency. Here is a test comparison:
Jetson TX2 performs GoogLeNet inference at up to 33.2 images/s/watt, nearly twice as efficient as Jetson TX1, and nearly 20 times more efficient than Intel Xeon. Jetson TX2 is the same size as TX1, and the interface is also the same. You can directly replace Jetson TX1 with Jetson TX2 to upgrade.
3. Jetson TX1 /TX2 comparison
In terms of processor, TX2 is upgraded from Tegra X1 of TX1 to Tegra Parker processor. This processor is manufactured by 16nm process and designed with 6 cores. The CPU part is composed of 2 Denver + 4 A57 cores. The GPU adopts Pascal architecture. 256 CUDA, the floating point performance is 1.3TeraFLOPS, which is about 50% higher than the GPU performance of the old Tegra X1. In addition, it is equipped with various interfaces such as CAN, UART, SPI, I2C, I2S, GPIOs, etc., with powerful performance and small size, which is very suitable for application scenarios with low energy consumption and high computing performance.
Therefore, the difference between TX1 and TX2 is actually reflected in the following aspects:
-
Storage RAM: The RAM of TX1 is only 4G, while TX2 is upgraded to 8G
-
Storage EMMC: The EMMC of TX1 is only 16G, while that of TX2 is increased to 32G
-
CPU: TX1 is a 4-core A57, and TX2 has 2 additional Denver2 (dual-core) @ 2GHz in addition to 4-core A57, so their CPU computing power is basically the same
-
GPU: TX1 uses a 256-core Maxwell GPU, while TX2 uses a 256-core NVIDIA Pascal™ GPU, so the floating-point computing power is increased by about 50%, but it is actually less than 50%
-
Peripheral interface CAN: TX1 has no CAN controller, while TX2 has 2 CAN interfaces
-
Peripheral interface PCIE: TX1 only has PCIE 1X4+1X1 interface, while TX2 can be equipped with PCIE 2X1+1X2
Consulting and technical service contact information: [email protected]