TU-TESL

• 22

Sarinx plays an important role in the field. Yesterday (December 3), selinx, an adaptive and intelligent computing company, held a 2019 selinx developer conference in Beijing. At the meeting, sarinx officially released Salil to Asian developers and Vitis to Asian developers Gamma Unify the software platform and announce Vitis for artificial intelligence and machine learning inference Gamma AI is open for free download today. Victor, President and CEO of sarinx Peng said, "the era of adaptive computing has come. As a leading enterprise of adaptive and intelligent computing, sarinx is committed to work with industrial partners in Asia and China to continuously promote industrial development and innovation in various fields. With the continuous deepening of sarinx's transformation from device to platform company and the continuous promotion of the company's three strategies, we will continue to make highly flexible adaptive computing Products and platforms enable more developers to innovate continuously. " At the beginning of 2018, Xilinx announced the launch of the company's three major strategies, namely, data center priority, accelerating the development of core markets and driving flexible contingency computing. After a year and a half of development, with its unique high-performance and flexible adaptability, Xilinx has been expanding its market and playing an increasingly important leading role in the major trend application fields of data center, artificial intelligence, 5g and other industries. Facing the field of data center, Xilinx launched the world's fastest data center accelerator card series alveo accelerator card on XDF last year. In more than one year, in order to meet the different needs of users for computing power, size, memory bandwidth and cost, four product series of U50, U200, u250 and U280 have been launched successively, which are used to greatly improve the performance of cloud and local data center servers. With the help of alveo accelerator card, customers can achieve breakthrough performance improvement with lower delay when running real-time machine learning inference, video processing, genome analysis, data analysis and other key data center applications. In addition, it is worth mentioning that the company has trained more than 7500 enterprises and academic units, over 800 accelerator projects and nearly 100 application releases in just over a year. The latest Vitis unified software platform will enable software engineers, artificial intelligence scientists and other broader developers to benefit from the advantages of Xilinx's flexible response hardware acceleration. This is the first time that Xilinx has launched a development tool platform of "integration of software and hardware design", and it is also one of the milestone products of the company's strategic transformation from devices to platform enterprises. Vitis can automatically adapt and use the Xilinx hardware architecture according to the software or algorithm code, freeing users from the complex hardware expertise. With the help of Vitis platform, both software engineers and AI scientists will benefit from the hardware advantages of Xilinx's flexible response. For hardware developers, Vitis can significantly improve work efficiency through the cooperation of software and hardware engineers on the same tool platform. In the field of automatic driving, Dan Isaacs, director of strategy and customer marketing of Xilinx automobile, said, "in the past few years, the shipment volume of Xilinx FPGA in the automobile field has reached 170 million pieces, and the shipment volume of ADAS is 70 million." This is a outstanding achievement and also the final product of Lingsi has very high quality and reliability. According to the latest news, Xilinx has launched two new products, both of which are more programmable, have higher I / O and have better performance at the automobile level. This is the latest product in the SA product line of Xilinx automobile. "This is mainly based on the needs of ADAS and autopilot, for example, now we have higher requirements for edge detectors and central pre controllers," Dan Isaac said What we need to emphasize here is the function of data collection, preprocessing and distribution of Xilinx, because more sensors, programmable capabilities and accelerators are needed in the whole system. In this regard, Xilinx has a series of scalable product lines. Here we can see that the devices of Xilinx products are covered from small to large There are potential needs of customers. Now we see the redefinition of the automobile industry. We see several very important trends in the automobile industry. The first trend is related to ADAS, that is, we are now gradually transitioning from computer vision to neural network AI. We hope to optimize our driving through the changeable logic. At present, in order to achieve automatic driving, there are two ways to go: the traditional way. At first, the traditional manufacturers first do panoramic ring detector, move from the outside to the inside of the car, and then do ADAS pre controller. Non traditional path, such as Internet companies, they directly do centralized processing module to achieve automatic driving. The products and solutions of Xilinx cover all the contents of these two paths, whether they are edge sensors or pre controllers for centralized processing. Through the products, all relevant data are collected, consistent, integrated, pre processed, and then distributed. At present, new technologies such as front camera, lidar, panoramic camera and full screen rear-view mirror are all applying products of Xilinx. In the aspect of camera, it has been transiting from machine vision to machine learning, artificial intelligence and neural network to realize automatic driving. "A lot of customers are using the functions and technologies of our automatic driving central module, whether in terms of data integration or sensor integration, they are proving once again that our products and technologies are very scalable, from very small devices to very large devices," said Dan Isaac Dan Isaacs gives an example of a centralized processing module. After the data collection, preprocessing and distribution of DAPD are completed, sarinx will send these data to the computing accelerator, and use versal to achieve acceleration, and then the data will be sent to the security processor to ensure that the cab is completely safe. Of course, it's just a kind of architecture. In fact, there are many other kinds of architectures used in self driving. That's why Xilinx must have scalability and flexibility. That's why Xilinx has adaptive and scalable products that are used in many cars. In the morning's forum, Victor Peng talked about the dual optimization of Xilinx. Here, he refers to the realization and optimization of hardware model through the optimization of machine learning and neural network. It can be said that such a way can achieve the inferential performance, low latency and high performance of the industry AI. Delay is very important for automatic driving, and the response speed needs to be very fast to truly achieve security. Generally, traditional CPU, GPU or DSP can achieve high throughput, but there is no way to achieve low delay, because the processing time will be longer, which will take longer to respond. In the automatic driving test conducted by Xiaoma Zhixing, a start-up company of automatic driving, the scheme delay of using Xilinx FPGA can be increased by 12 times. Compared with the general computing architecture, the energy consumption is only 1 / 10 of that of the general computing architecture. It is emphasized that there are still some general computing architectures that can never be provided, such as very strong timing. Sarinx FPGA can also realize on-chip configuration in operation, which is also called adaptive. According to Dan Isaac, "the safety standards of European new car test assessment are constantly changing. With these standards changing, our products and technologies can be reconfigured in operation without introducing new devices. In Mipi protocol, the data rate is increasing, we do not need to change the original device, as long as the change of programmable logic can support high data rate. " Finally, the characteristic of Xilinx devices is to use DFX, which is the performance of dynamic function exchange. At this time, it is not necessary to restart or turn off these devices to realize the function exchange between devices. In I / O or sensor, there is no need to change too much. Use part of programmable logic to change. For example, when I change the reverse gear, I start the rear camera. In this process, I don't need to turn off the whole system and restart it. The advantage of this is to save costs and increase the reliability of the car. In general, all products and technologies of Xilinx can provide AI inference with minimum power consumption and high scalability. Through programmable logic, it can help enterprises realize dynamic function exchange or remote hardware, that is, chip update. According to the requirements of sensor information I / O, it also has the ability to accelerate. At the same time, add ADAS to realize heterogeneous computing, and help vector engine, AI engine and various engines to realize heterogeneous computing.