"Indeed this year, there was an obvious explosive growth in demand for AI servers. As the industry leader, Sugon prepared a complete set of solutions aiming at AI, but artificial intelligence (computing is focused on training and reasoning), is actually one of the application directions among many of the HPCs. Sugon, derived from the Chinese Academy of Sciences, is more concerned about the technical challenges. ", said by Ji Qing, chief scientist of Sugon HPC products division, MIT postdoctoral researcher, when receiving the exclusive interview from Iyiou.
AI computing needs not only faster, but more accurate
The rise of AI leads to soaring demands for HPC equipment. Aiming at the needs for AI, Sugon prepared a complete set of solutions from product, platform and solutions. Helping users not only solve speed issues by HPC, Sugon also provides data service. On a well -documented and well-founded basis, Sugon help users improve execution accuracy and promote high efficient value acquired by AI.
From product level, Sugon set up dedicated development team, which is focusing on developing HPC products according to the training and reasoning demands for deep learning neural network. At present, Sugon has already produced XMachine series of HPC products massively. From the one hand, in face of the intensive computing requirements at training end, Sugon provides HPC servers with integrating GPU and MIC accelerator. One out of every eight GPU servers in the world comes from Sugon (Top500, 16 NOV). From the other hand, reasoning has more strict demands for responding speed of high concurrency. Besides traditional CPU servers, Sugon has also launched the world's first AI reasoning dedicated server based on the Cambrian AI special chip. The efficiency ratio is much higher than the general processor, and the efficiency ratio is about 30:1, and the peak value computing power of the single node exceeds 100TFLops.
In particular, what is worth mentioning, Sugon has also strengthened big push in server cooling technology. From cold plate cooling server to full immersion cooling server and to phase change immersion cooling server, Sugon has achieved good marketing focus and performance. It is reported that the technology has been deployed in large-scale in electric power, finance, meteorology and many other core business units, which has been highly praised by customers. Sugon’s liquid cooling solution can be optional when applying in AI training and reasoning server, which can effectively reduce energy consumption and noise, and it can dramatically increases system stability.
As for software, in response to current situations of the inexperience of most AI users, Sugon integrates the commonly used publicly available TensorFlow, Caffee neural network frameworks and ImageNet public data sets into a more user-friendly "Sothis AI" management platform, which is more convenient for users having offline training, online reasoning and other related operations. It is easy to use and rapid to deploy.
In terms of cloud platform, Sugon came up with ABC² integrated solution. ABC² can manage all hardware downward, while it can drive detailed industry applications based on scenarios upward. ABC² integrated solution has been deeply customized in 10 vertical industries. Especially in four vertical directions: security protection, traffic control, land resources, smart campus, it has a very mature proposal. There are abundant landing cases in many parts of the country, and even overseas.
Based on current solutions, Sugon can be the integrated solution provider integrating upstream and downstream resources and services and provide the end users “one-stop-service” convenient solutions. From the industry perspective, Sugon unites professionals from all walks of life and proposes solutions throughout from upstream to downstream. For example, Sugon can provide partners with cameras, algorithms, high speed network, etc. Sugon can also work together with upstream and downstream resources to complete the deployment of complete set of AI solutions.
Different from simple emphasis on fast speed, Sugon still pursues computing accurately. To be accurate, more adequate data, adequate new data, adequate complete data are required. Furthermore, to be more accurate, more advanced and efficient algorithms are also needed. Such as, more deep network layers or further system optimization.
In order to achieve the computing accuracy, Sugon made efforts in many aspects. Each has its own advantages in computing platform, but Sugon provides a better cost-effective Cambrian chip proposal. Sugon’s deeply optimized algorithm set XSharp, comparing to the open source base library market, has a significant advantage. In terms of medical image recognition, Dawn's XSharp algorithm set is more than 200 times more accurate than common open source libraries. At the data level, basically all open source libraries can be integrated into Sugon’s SothisAI platform.
HPC not only used in AI, its application ecology more broadly and largely
HPC is a very powerful tool, just like a super screwdriver, which can be applied in many scenarios. AI is only one of the application directions. In addition, it can also be applied in computing materials, biology, precision medicine, meteorology (atmospheric ocean, circulation), petroleum exploration, petroleum processing, remote sensing, data simulation, wind power and many other fields.
Smart cities and IoT promoted strongly by the state are the important application scenarios of HPC. In 2009, Sugon registered city cloud trademark. Now there are nearly 40 city clouds landed in China. After the city cloud landed, Sugon found a lot of new and unique local needs in various places. In this way, personalized solutions are provided according to local conditions. For example, the city cloud in Inner Mongolia and Hebei is very concerned about the environment, while the city cloud in Shanghai is more concerned on Finance. Wuxi cloud center is to strengthen cooperation with various local educational institutions. The smart city based on HPC and local personalized demand is also applied to the construction and service of many systems in the city, such as environmental protection, finance and education.
Comparing to the information interconnection in smart cities, the ultimate target of IoT is to realize the interconnection of all things and no less than the smart city in terms of connecting nodes and data needs. To some extent, although edge computing can relieve some of the pressure on large sized computing center, future applications based on IoT are still inseparable from the support of HPC. In the face of future application of extensive IoT, the current performance of HPC as a whole is not adequate; there is quite much potential developing room. Sugon is constantly innovating on HPC so as to adapt future new demands of all things connected to the calculation of new requirements.
Sugon concerns more on Technology Innovation
Sugon, as a spinoff from the Computing Institute of CAS, has the passion for technology in blood. In the face of the emerging technological frontiers, Sugon is more willing to pay more attention and enthusiasm in the research and development work. At the same time, the customers who have an idea about the architecture of new technologies are more willing to choose Sugon as their partner. On the one hand, Sugon has a unique advantage in the accumulation of technology; on the other hand, Sugon has deep insight into customers’ needs. Taking the recently introduced M-Pro product as an example, only Sugon and Facebook have it in the world, and its technology foresight has now drawn another Internet giant into it. Another example of innovation is that Sugon is the first to introduce a refactored cloud server that can be used to define software as a storage or server.
As the industry leader in HPC in China, Sugon is also actively calling for more partners to join the team to promote the development of HPC while innovating and developing itself. HPC industrial chain is very long, and it involves quite a lot of technical aspects. It has both the direction of the hardware, as well as application algorithms, programming, compiler and other applications and technologies, and it also has the specific application of the physical model. It involves a lot of steps, each step needs to be improved, and it needs to draw more attention from the industry. The progress of the entire industry also requires the joint efforts of all colleagues in all aspects.