Heterogeneous computing provides surging power for the digital economy

On May 11, at the "2022 Alibaba Cloud Visual Computing Private Sharing Conference", Zhang Xintao, an expert of Alibaba Cloud elastic computing products, shared the theme entitled "Alibaba Cloud Elastic Computing Application Practice in Visual Computing". The following content is based on his speech.

The metauniverse, digital twins and virtual reality mean that we are evolving from one-way interaction to all-factor interaction. This evolution is to achieve realism through immersive iterative real-time interaction, and realism brings about the improvement of human cognitive efficiency.

This judgment stems from the famous pyramid model of human learning and cognition put forward by an educator in the last century. In the model, the most efficient learning method of human is Feynman learning method, but it is not suitable for all people. For ordinary people, it is more beneficial to improve learning efficiency and cognitive efficiency through simulation practice. The most efficient way in simulation practice is real-time interaction of all elements. For example, pilots will have a long time of hard training in the flight simulation cabin before they actually fly to the sky. This approach is very close to the learning efficiency of flying in a real plane and is more safe and economical.

The improvement of cognitive style is essentially the improvement of productivity. As an important concept of human society, productivity makes it necessary for us to seriously understand the metacosmic industry. What opportunities and challenges does it bring?

The Metaverse seven-layer model is a model that describes the metauniverse industry relatively completely at present, from the user experience level to the creator economy, and finally to the infrastructure level. It includes all relevant concepts, performance levels and technologies, involving many industries.

Mapping these concepts and technologies to the common fields of cloud computing will ultimately have a great impact on cloud computing in five aspects: computing, application, communication, human-computer interface and business ecology. This is also an unprecedented challenge, even a leap challenge for cloud computing.

Take smart phones as an example. Compared with PCs 20 years ago, smart phones today have made a huge leap in all aspects. This is due to the upgrade of the interactive mode of smart phones, and the emergence of the operating system that supports multi-touch control. Finally, a rich application ecosystem has grown on this platform, which in turn has promoted the continuous improvement of computing performance and communication performance. At present, the performance and function of mobile phones even surpass that of laptops. Mobile phones provide sensors and interaction modes that are not available in laptops.

The interaction modes faced by the universe may be real-time 3D, gesture interaction, voice interaction, etc. These interaction modes ultimately bring more challenges to real-time computing and interactive communication. For example, there was only touch screen before, but now there are more sensors to recognize gestures, scenes, and space. These require real-time computing, and the communication requirements are also more demanding, requiring greater bandwidth and lower delay.

At the same time, the business ecosystem faced by MetaUniverse will be more complex than any other IT system in the past, so another problem we need to face is how to operate this whole set of business systems.

Therefore, in summary, the three main problems that cloud computing needs to face to adapt to the metauniverse business are computing, communication and ecology. It requires 10 times+computing performance and computing scale, 10 times+communication and low latency, and 10 times+ecological efficiency and openness.

In terms of computing performance and scale, it is necessary to have 10 times+graphics computing to render higher fidelity 3D models, and to realize real-time ray tracing of more scenes in order to achieve a true description of the real world; In AI, we need stronger computing power to train and run AI models, and stronger computing power to process streaming media; In terms of bandwidth and latency, greater bandwidth and lower latency are needed to be responsible for pushing rendered high fidelity images to the client, as well as low latency bandwidth to solve the training and reasoning of large-scale AI models, and high-performance bandwidth is also needed to solve the communication between humans and virtual humans through subtle expressions.

In addition, the bigger challenge comes from ecology. The technological ecology of the universe is very complex. Previously, the case by case industrial cooperation mode could not expand the industrial scale to 10 times or even 100 times. Therefore, we need new ways of ecological collaboration to make collaboration more efficient and open.

In response to the above three problems, Alibaba Cloud has done a lot of work on technical performance indicators and computing scale in the past decade. For example, the fourth generation DPCA architecture was released at the Yunqi Conference last year, and the large-scale networking of eRDMA was realized through hardware acceleration. The delay was as low as 5 microseconds, which is at the leading level in the industry. It has the ability to recognize human subtle expressions on the cloud.

Alibaba Cloud's GPU computing instances for AI and Meta Universe have also made great progress, and the performance of GPU instances has been improved 100 times in the past few years. Last year, a 7-generation GPU instance was released, with significant improvements in image computing and AI computing. What's more, Alibaba Cloud has achieved the accumulation of very large computing scale. Through commercial operation, it has achieved the largest GPU cluster in Asia, with a total computing power of more than 1700P, and is still maintaining rapid growth.

In addition to technology, business ecology is also a huge challenge for us. To prosper the industry, we must improve the productivity of the industry and straighten out the production relations of the industry. Together with partners and customers, we have built Alibaba Cloud's visual product system through ecological operation, and have achieved a lot of accumulation in products.

As we all know, heterogeneous computing is a very difficult computing system, and our production and research team has also made great efforts here. By means of software, the common problems in the use of heterogeneous computing in various industries have been solved, and a more efficient, high-performance, cost-effective and easy-to-use heterogeneous computing product has been finally realized.

More importantly, we have built visual computing software services with the leading partners in all walks of life, and have landed many products on Alibaba Cloud.

We divide software services into three layers to solve the problems of industrial ecology.

Common layer: two of the more important products are the cloud XR platform and the cloud real-time 3D editor platform, which mainly solve the problems of 3D content and application hosting and running environment, and the problems of 3D content and application development and production.

Industry level: I hope to build a huge industry editor matrix with editor partners from all walks of life to solve the problem of their own editors. The editor helps to more effectively implement the enterprise application of MetaUniverse, let everyone participate in it, and improve the productivity and creativity of the whole industry.

Ecological layer: The ecological layer mainly solves the problems of content production and content trading. It is responsible for providing materials for the first two layers, so it also needs to focus on construction.

In the near future, these products will be launched on Alibaba Cloud. They will work with industry editors and general editors to solve problems in all walks of life.

The most common application scenarios of the current metauniverse are shown in the figure above. After normalizing the scenario requirements, they are finally concentrated in five levels. The requirements of these five levels can be one-to-one corresponding to the previous product system.

For example, the important link of automatic driving before the road test is to do real-time simulation of 3D environment, which can generally be completed through the industry editor. With the industry editor, it is no longer necessary to use the general 3D editor to build houses, roads, etc., but to use the simulation platform to quickly complete the environment construction on a white map, which greatly improves the efficiency. It can even provide different weather, wind speed, lighting conditions, rain and snow conditions to simulate the real driving environment to detect the operation of the automatic driving program.

However, not all industries have such editors. Therefore, we provide two relatively common products to solve similar problems in various industries.

The first is the cloud XR platform. It is a 3D real-time rendering platform built on Alibaba Cloud based on the CloudXR suite, which solves issues related to the installation, release, and hosting of 3D applications. Customers can install their own developed meta universe applications and 3D applications on the platform, and then publish them to specific audiences. Users can access and use these 3D applications through various terminals, and currently can support most common 3D engines and VR/AR terminal devices.

We have also adapted and landed in various industries with the aim of improving the creativity and productivity of various industries. In addition to the XR operating platform, we also need more content. How are the contents of these 3D applications generated?

You can rely on the cloud real-time 3D editor platform.

Cloud real-time 3D editor is the first cloud native 3D editor implemented by Alibaba Cloud together with international top 3D engine partners. Compared with previous development on PC workstation, cloud real-time 3D editor is more flexible and powerful.

Its flexibility is that it can be accessed through any device. Previously, it can only use workstations or game books; But now, any terminal such as Pad and mobile phone can be connected to do 3D app design and game design.

Its strength lies in that the previous development workstation was limited to existing computing resources, such as CPU resources, memory resources and GPU resources. But in the cloud editor, it can use the unlimited computing power on the cloud to solve the time-consuming computing problems and greatly improve the research and development efficiency. Compared with the baking process that took more than ten hours to complete before, it only took more than ten minutes on the cloud.

In addition, after these applications are developed, they can also be adapted on the cloud. Because Alibaba Cloud has GPU instances of various systems, such as x86 and ARM, which can simulate various terminal environments to do product adaptation, which is more efficient than previous adapters. Our platform is also about to go online. Welcome to test and try it out.