Shulou(Shulou.com)11/24 Report--

The golden ratio is a kind of "perfect" proportional relation in mathematics, which was first systematically discussed by Euclid in the original Geometry. After that, the concept of golden ratio has been widely used in mathematics, physics, architecture, agriculture and other fields, representing the most reasonable and coordinated situation or state.

In the field of data center, the reasonable proportion of related resources is more and more concerned by users. Especially with the rise of large models and generative AI applications, the demand for computing power and survivability is increasing rapidly. More and more data center users realize that data center resources need balanced allocation and coordinated development in order to give full play to their capabilities and value.

Since the beginning of this year, the industry has repeatedly called for the construction of data centers in the AI era not to be biased, and the construction of computing power and survival power are equally important. As Liu Ximeng, deputy general manager of Chaochao information storage product line, said: "at present, the pattern of 'hundred models competing for the show' in the generative AI era is emerging, and the construction of large AI models requires not only a math base, but also a survival platform. By building the computing, all-flash and hybrid flash storage of the data center according to the 1:1:1 golden ratio, users can maximize the return on investment."

Data center can not be "biased"

Gartner predicts that 20 per cent of content will be created by AIGC by 2023 and 10 per cent of data generated by artificial intelligence by 2025. There is no denying that generative AI and large models are becoming the biggest traction force for the development of data center infrastructure. It can be predicted that the infrastructure investment brought about by AIGC and large models will continue to grow in the future.

However, judging from the current real situation, the construction of the data center is "biased" and "unbalanced". For various reasons, the phenomenon of "recalculating power and neglecting survivability" is more obvious. most users attach great importance to the deployment of computing products such as GPU, but ignore the importance of survivability construction, and lack of planning and matching for the overall resources of the data center.

As we all know, the core of the application of large models is high-quality data, which determines the performance, generalization ability and application effect of the algorithm, and the acquisition of high-quality data is closely related to survivability. Around the links of "transmission, storage, analysis, management, security" of data, survivability is an essential key factor to release the value of data.

In fact, with the development of large-scale model today, it has become an engineering problem of large-scale, high-quality data and efficient data processing. With the gradual evolution of the large model to the multimodal direction, it means that in addition to the continuous computing requirements, it brings unprecedented changes to the data storage capacity, performance, multi-protocol support, reliability, data management and so on.

For example, large model multivariate heterogeneous data collection, tagging, training, reasoning and archiving all need high efficiency to move data, which means that supporting multi-protocol fusion of heterogeneous data will be the key to solve the problem of large model data movement and processing efficiency; for example, AIGC applications will produce a large number of reasoning requirements, followed by large-scale parallel processing and complex IO, which require high data storage performance. For example, large model training often needs to call hundreds of fast GPU cards, and there are different CheckPoint, which requires more and more stable and reliable storage.

Liu Ximeng said bluntly that data storage and management undertakes two important responsibilities in the AIGC era: first, it supports the full life cycle management of massive multivariate and heterogeneous data; second, it carries various stringent requirements for AIGC data training and reasoning for performance, delay, capacity, scalability, and so on.

As far as users are concerned, in addition to paying attention to the construction of storage capacity, a realistic challenge that can not be ignored is: how should the ratio of computing power and storage resources in the data center be allocated to the best? To this end, wave information has brought its answer: from the comprehensive consideration of many factors, such as data capacity, bandwidth, access frequency and cost, the future data center will need to form a 1:1:1 golden ratio of computing power, flash memory and mixed flash in practice, in order to meet the needs of artificial intelligence applications such as AIGC, big model and so on.

How does the golden ratio come from?

Compared with the mature markets in Europe and the United States, the development of China's survival power has been lagging behind the calculation power. This can be seen from the low popularization rate of all-flash memory and the weak construction of disaster recovery protection in data centers in our country.

With the advent of the AIGC era, this lagging phenomenon is more obvious and prominent. In the face of the rapid demand for computing power of AIGC, many users "take a look at it step by step" from the very beginning, often buy computing power first, find that the existing power can not keep up with it in the process of use, and then start to configure the corresponding existing power, lack the overall planning of data center resources, and the way of construction is obviously out of date.

To some extent, by popularizing and promoting the 1:1:1 golden ratio of data center computing power, flash memory and mixed flash in our data center, we can not only enable users to better support the innovation in AIGC field at the infrastructure level, but also promote the storage capacity construction of our data center and enhance the overall resource allocation and utilization level of the data center.

But why is the golden ratio of data center resource allocation to "1 GPU node, 1 all-flash storage and 1 hybrid flash storage"? The reason why the tide message proposes the 1:1:1 golden ratio of computing power, flash memory and mixed flash mainly comes from two core reasons:

First of all, the golden ratio comes from the practice of wave information getting involved in large models earlier. As early as 2021, Wave Information released the Source 1.0 Chinese large model, when the model parameters were as high as 245.7 billion, and the amount of training text data was as high as 50TB. In these years of large model training and reasoning practice, the infrastructure products of wave information itself played a key supporting role; at the same time, wave information also deeply felt the rational allocation of computing power and storage power in the data center, which is of great importance for the development of large models.

For example, in the scenario of large model training and reasoning, the biggest challenge of data storage is how to transfer different data to CPU and GPU, so it is a great test for data processing performance and how to cooperate with GPU. "the practice of source 1.0 is the inherent advantage of wave information storage products. Few enterprises in the market can build a large-scale cluster to support the application of large models." Wave information distributed storage product line general manager Jiang Leguo said.

Secondly, as a leading Top2 enterprise storage manufacturer in China, Chaochao Information has a deep insight into the future development of flash, mixed flash and other related storage technologies. Coupled with the successful application of wave information related storage solutions in many domestic AIGC enterprises, it has accumulated a lot of practice for the overall construction of data centers in the AIGC era.

"Tide Information has the ability of full-stack technological innovation in the field of flash memory, from the controller of the underlying SSD to the software and hardware of the storage system, and then to the upper applications, to achieve disk control coordination, as well as the optimization of full data links, which is conducive to applications like AIGC to fully release the value of data." Liu Ximeng added.

In fact, considering the internal and external environment and factors of the market, the shortage of GPU in the future computing market will continue for a long time, which also makes the 1:1:1 gold ratio of computing power, flash memory and mixed flash have very strong practical significance. In the case of the shortage of computing power, under the same allocation of computing power, the value of the overall resources of infrastructure can be brought into full play through the rational allocation of computing power and storage power.

In order to further promote the promotion of the golden ratio in the data center field, Tide Information has recently officially launched a large model application storage system: AS 15000G7, to help users extricate themselves from the complex infrastructure and devote themselves to AIGC innovation.

AS 15000G7, let the golden ratio come true.

It can be said that the storage system is the key to the popularity of the golden ratio.

As we all know, in recent years, with the continuous increase of flash media capacity and the continuous decline of price, it has created excellent conditions for the development of storage power in China. There is no doubt that the rise of AIGC will further drive all-flash, hybrid flash and other storage products to accelerate innovation.

"AIGC applications have brought about an overall improvement in capacity, performance, functionality and other requirements." "Storage systems not only need to be completely combined and designed to meet the data storage needs of AIGC applications, but also avoid the complexity and inefficiency of traditional storage schemes," Jiang said. "

Therefore, Tide Information creates AS 15000G7 for AIGC application scenarios to meet users' comprehensive needs for data storage trained in large models in terms of performance, management, fusion and efficiency through extreme performance, management, fusion and efficiency, helping AIGC to land in various industries and accelerate the release of data value.

First of all, according to the characteristics of high concurrency and complex IO of large model, AS 15000G7 brings extreme performance to AIGC from the aspects of architecture, hardware, key technologies, IO path optimization and so on, and brings performance guarantee for the training of large model. Specifically, AS 15000G7 shortens the I / O path through GDS and RDMA technology, and significantly improves the speed of data access and retrieval by using intelligent metadata management; in addition, the unique intelligent network optimization technology improves the concurrency ability of network ports, reducing the delay by more than 50%, especially the delay of small file-level transmission can be reduced to millisecond.

Secondly, for large model training process management, AS 15000G7 provides transparent and controllable extreme management of the whole process. AS 15000G7 can be equipped with AIStation scheduling platform and InView data management platform to carry out intelligent operation and maintenance of AI server, network, storage and other equipment, and support multi-tenant management, resource allocation, data management and analysis of the whole process of training and reasoning. The whole process of AIGC data collection, cleaning, training, reasoning and archiving can be monitored and managed through a set of storage.

Third, aiming at the collection, tagging, training, reasoning and archiving of large model multivariate heterogeneous data, AS 15000G7 fusion architecture realizes the ultimate fusion of multi-source and heterogeneous huge data, parallel access to files, objects, big data and video, supports multi-protocol real-time access and system flat expansion, and maintains semantic consistency and lossless performance in the process of data access. Thus, the massive multi-source heterogeneous unstructured data of AI large model can be shared efficiently.

Finally, in view of the huge investment required by the large model, AS 15000G7 can help users to achieve the best allocation of data center resources in a golden proportion, improve the return on investment ratio, and bring extreme efficiency. According to different media such as flash memory, magnetic disk, magnetic tape and optical disc, AS 15000G7 is divided into three models: performance type, balanced type and capacity type, and based on automatic data layering and migration, under the premise of application security and transparency, it realizes the whole life cycle management of hot, warm and cold data, resulting in a significant reduction in TCO.

There is no doubt that the rise of AIGC marks a turning point in the development of artificial intelligence. At present, China has become a hot land for global AIGC innovation and development. According to incomplete statistics, the number of large models in China has exceeded 200. different types of enterprises are trying their best to promote the development of AIGC and large models. Nowadays, people are increasingly aware of the truth of "large model industry development, infrastructure first", and the 1:1:1 golden ratio construction concept of computing, flash and mixed flash has emerged at an opportune time, which will help enterprises exploring AIGC to reduce infrastructure complexity and thus better focus on innovation.

"AIGC is just in its infancy and will continue to bring demand for infrastructure in the future. It is expected that by 2026, the golden ratio construction model is expected to be widely used." Liu Ximeng said finally.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.