Article first: Shen Yao’s observation of science and technology

There are no doubt that there are two important trends in the global software industry: on the one hand, “software eats everything now,” and every company is becoming a software company; On the other hand, “open source is also devouring everything”, and more and more companies are embracing open source and giving back to open source.

Intel is not exceptional also, in the past few years Intel started opened by “to the PC as the center” to “data-centric” business transformation, and in the “data-centric” business transformation, Intel is put forward, including process and encapsulation, XPU architecture, memory and storage, interconnection, security, software, six technical backbone.

At present, Intel has formed a strong software stack ecosystem covering firmware IP and BIOS, hardware drivers, operating systems, virtualization and cloud computing, underlying system libraries, middleware and framework layers, runtime libraries and applications. In addition, Intel has been embracing open source for more than 20 years, especially in the Linux Kernel/KVM two communities, Intel code contribution is always the first.

In this sense, software is deeply embedded in Intel’s DNA. Then, under the wave of digital intelligence transformation, how does Intel release its own software and open source power to promote the acceleration of new architecture, new technology and new business?

The software is preferred

We know that at Intel’s “Architecture Day 2020” event held in August this year, Intel announced a series of latest technological advances, including Tiger Lake SoC, XE GPU, Superfin transistor process, FPGA roadmap, OneAPI framework, and in particular the emergence of XE GPU. It is also officially announced that Intel has completed the full coverage of four computing types of chips: Scalar (CPU), Vector (GPU), Matrix (ASIC) and Spatial (FPGA).

Xie Xiaoqing, vice president of Intel’s architecture, graphics and software group and general manager of China region, said that with the strong rise of heterogeneous computing, more and more diversified application scenarios, as well as the continuous improvement of users’ pursuit of application experience, all of these have put forward unprecedented challenges and requirements for software. To this end, Intel also “keep pace with The Times” put forward a new software strategy, hoping to better play the value advantage of software, specifically:

One is, software first. From the point of view of computing, Intel in the past few decades on the CPU to build software ecosystem is a valuable “wealth”, which makes Intel in programming language, system libraries, and has a strong ecological support tool chain, but Xe GPU, after all, just “scene” in the future graphics, multimedia, computing requires more powerful software support, In this regard, Intel can take advantage of the existing software ecosystem as a foundation to lay a better starting point and foundation for the next development of XE GPU.

Second, it is easy to expand. Currently, there are many segments in the GPU market. Different markets require different power consumption and different performance. For Intel, the future also wants to be able to maximize the support of all GPU developers, so the scalability of the software is becoming critical.

Third, new computing loads and user scenarios. As traditional software and solutions struggle to cope with the increasing diversity of workloads and application scenarios, Intel Software’s future focus and direction will be to develop new computing loads and user scenarios. This is the trend of the future and the value of Intel differentiation.

Not only that, but OneAPI undoubtedly plays a key role in the implementation of Intel’s new software strategy. Earlier this year, Intel released a new specification for OneAPI. In December, it released the Gold version of OneAPI, which is based on Data Parallel C++ (DPC++), a language built on C++ and the Khronos SYCL standard. In addition to the DPC++ compiler toolchain based on LLVM/Clang, OneAPI also includes a number of libraries, such as OneDNN for deep learning, OneMKL as a math kernel library, OneDAL for analysis, OneTBB for threading, And software components such as OneVPL for video processing.

According to Xie, OneAPI offers three aspects of value to developers. One is a very friendly programming environment, allowing developers to choose their hardware platform freely. The compilers and systems provided by Intel are highly optimized, so developers can maximize hardware capacity and support hardware acceleration for heterogeneous computing in an optimal way. In addition, its development mode is very fast and efficient, so that the maintenance cost of the source code can be minimized.

“The slogan of OneAPI is’ No Transistor Left Behind ‘. A new Gold version, due out in December this year, will first support the CPU and GPU. In the future, OneAPI will also support other AI hardware acceleration and computing chips such as FPGAs. This allows developers to develop software more quickly and efficiently.” She said.

Thus, under the impetus of the “software” priority strategy, at the same time with the aid of the oneAPI such brand-new development tools, Intel is making the future much calculation software architecture is blossoming into a “new normal”, believe that will also be able to better support the broader industry ecosystem and more developers, under the unified software architecture can assign, Realize the “new evolution” of application development and innovation.

Open source and open

In addition to strengthening the value of software, in the face of the rising tide of artificial intelligence driven by data, algorithms and computing power, Intel has also increased investment in open source. The most symbolic event is that Intel announced the establishment of Big Data Analysis and Artificial Intelligence Innovation Institute in China in June last year. It hopes to accelerate the innovation and application of unified big data analysis and artificial intelligence technology.

In the view of Dai Jinquan, president of Intel’s Big Data Technology Global CTO and Big Data Analysis and Artificial Intelligence Innovation Institute, AI has become the new driving force and driving force in China’s industrial upgrading and enterprise digital transformation. However, it is actually a very complex and refined project for AI to be applied in a wider range of industries.

In this process, developers often face two major challenges: on the one hand, the amount of data is huge, and developers are faced with how to apply AI algorithms to complex data; For example, in the stage of data annotation and preparation, data annotation and preparation are very time-consuming, accounting for about 50% of the overall development time. Therefore, it is often said in the industry that “there is no intelligence without manual labor”. Therefore, how to apply AI directly to the production environment of big data, and be able to seamlessly scale on a large scale to achieve the best effect of AI application is the “top priority”.

In the face of this industry problem, Intel also opened an open source software platform of big data AI, Analytics Zoo. Analytics Zoo is built on the underlying accelerated library optimization framework of OneAPI, which can parallelize data. Developers using TensorFlow, PyTorch, Spark, Flink and Ray can easily deploy the AI without any need for optimization. Analytics Zoo will automatically complete the cluster scheduling and distributed computing, making the entire AI development process “flow smoothly”.

Dai JinQuan said Analytics Zoo has the software capability of end-to-end big data +AI. The bottom layer provides the ability of data pipeline, which can help developers directly and seamlessly run the AI model on its distributed big data. The middle layer provides the workflow of machine learning, which can automate many manual jobs and tasks. At the top, a lot of different application scenarios have been built, including recommendation system, time series analysis, computer vision and natural language processing, etc., which also means that Big Data +AI can enter a new era of high automation and intelligence in a new way of one-stop development.

Objectively speaking, Analytics Zoo not only realizes the application developers of big data +AI in the enabling industry, but also comprehensively improves the development efficiency and landing effect of AI in the industry. “We are always committed to bringing cutting-edge innovative technologies to developers. From BigDL in 2016 to Analytics Zoo two years ago, the goal is to connect end-to-end big data and AI. On this basis, we have successively added new technologies such as AutoML. This will allow more developers to easily build real-world application scenarios.” Dai Jinquan told me.

It is not difficult to see that after Intel gets through the underlying architecture through OneAPI, it also provides a brand new way of industrial AI implementation through the open source software platform such as Analytics Zoo. It can be said that it will have a huge effect and value for both industry developers and AI to enter into thousands of industries. Really make big data +AI “omnipotent”.

Assigned to innovation

Today, with the implementation of software-first and open source strategies, Intel’s path of industry-enabling innovation is becoming wider and more powerful.

For example, Intel’s OneAPI and the newly released GPU-SG1 play a large role in the overall Android Cloud game framework, and this solution is expected to be the first to complete production on Tencent Cloud. On the CPU side, Intel uses container technology to put Android apps into lightweight containers, which can take full advantage of Intel Xeon’s computing power and provide killer applications in the cloud, according to the company. On this basis, Intel Server GPU’s powerful graphics rendering ability and multimedia coding and decoding ability can help users directly render the graphical interface of cloud games in the cloud with independent GPU, and release it directly to the terminal after coding it in the way of streaming media.

Tencent first cloud gaming platform, deputy general manager Allen Fang, on the scheme gave high evaluation, in to the support of strong extensibility and discrete graphics processor, Intel provides a high density, low latency and low power consumption and low TCO solution, a have two SG1 graphics xeon server instance can run more than 100 games, Including popular games like The Legend and King of Glory.

For another example, based on Analytics Zoo, Goldwind Huineng also uses deep learning and machine learning methods, combined with fan level weather forecast, wind track simulation and other weather forecast data, to build a new intelligent scheme for power forecast by multi-model combination.

Through Analytics Zoo, Kim was able to seamlessly integrate Spark, TensorFlow, Keras and their software and framework into the same pipeline, so as to better integrate data storage, data processing and the pipeline of training reasoning into a unified infrastructure. Reduce the cost of hardware management and system operation and maintenance.

In addition, Analytics Zoo not only provides a unified end-to-end distributed solution for the solution, but also helps users improve the efficiency and scalability of the development and deployment of the system, especially in the analysis of time-series data. The test data showed that the accuracy of the final wind power increased from about 59% at the beginning to about 79%, which not only improved the accuracy of wind power prediction, but also greatly improved the efficiency of end-to-end training.

In general, the rise of heterogeneous computing, combined with the ubiquitous mass of data and the wave of intelligence generated by evolving algorithms using deep learning, has generated far more practical application effects and commercial impact than ever before.

In this process, Intel, through continuous innovation, unified platform provided by OneAPI and open source of Analytics Zoo, undoubtedly maximized to help enterprises quickly build a set of end-to-end data analysis and AI solutions. It completely solves all the problems from the underlying architecture, data collection, data preprocessing, training and model deployment. It not only realizes the rapid response to the customized needs of different industries and scenarios, but also makes the road of “inclusive AI” of the industry truly become a reality from the dream, and its value is “not only for the present, but also for the future.”