news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

AI for DB is quietly becoming a hot track. One of the most notable features is the vector database/vector retrieval technology that was very popular last year, which is increasingly sought after by AI big models.

AI for DB focuses on AI serving databases. From the user's pain point, traditional database infrastructure is not designed for large AI models, nor is it designed to meet today's vector retrieval.

For example, enterprises need to build huge data sets to implement large model applications. Only high-quality, high-density data can train models to achieve more accurate results. However, acquiring and managing such large amounts of data usually requires a lot of resources, including storage resources, computing power, and data processing capabilities. At the same time, integrating data sources with various formats, qualities, granularity, and heterogeneity will also complicate the model training process. This is one of the reasons why enterprises are still cautiously optimistic about generative AI.

Titanium Media has noticed that since last year, in the overseas market, leading database/data warehouse companies, and even large model companies, have been actively taking the approach of launching products, or making acquisitions and cooperation to seize the market opportunities of AI databases. For example, cloud data warehouse company Snowflake announced that it will cooperate with NVIDIA to tailor AI models for enterprises; Databricks acquired Tabular, the company behind Apache Iceberg, for US$1 billion; OpenAI acquired Sockset, a database company with vector search technology, for US$500 million...

However, from the current perspective, the combination of AI and databases is not only related to vector retrieval. In the past few years, autonomous databases, database self-monitoring and self-diagnosis, and the introduction of low-code + AI to text2SQL are all directions that enterprise customers in the industry are still exploring and have not yet reached a conclusion.

However, some industry insiders have warned that the combination of AI and databases is still a very new technological direction and there may be technological gaps.

Regardless of the trend of AI for DB, the first question that arises is: why now? And what are the new solutions?

Simplify data application and use

Take Oracle as an example.

In the past two months, Oracle has successively announced the AI ​​reshaping of its two core database management systems, Oracle Database and MySQL HeatWave database. The change in the name of the database alone has an obvious direction: the former was directly changed from Database 23c to Database 23ai, and the latter was upgraded to HeatWave GenAI. The changes in versions of different eras from "i" Internet, "g" grid, "c" cloud, "ai" artificial intelligence, and "GenAI" generative AI can reflect Oracle's keen insight into the detonating points of customer demands in different eras. Among them, Database 23ai is an upgrade of the above-mentioned vector database and more than 300 major functions.

Developers can "talk" to the Oracle database in natural language, call on the capabilities of generative AI, generate SQL and execute the final results, thereby achieving the purpose of communicating with the database.

Specifically, 23ai's Vector Search function enables LLM (large language model) to query private business data using a natural language interface and helps LLM provide more accurate and relevant results. Customers can use the Vector Search function to securely search documents, images, and other unstructured data in conjunction with private business data without moving or copying the data. This means that AI algorithms can be introduced to where the data is located without having to migrate the data to where the AI ​​algorithm is located, enabling real-time operation of AI in Oracle databases, greatly improving the effectiveness, efficiency, and security of AI.

HeatWave GenAI mainly includes in-database LLM, automated in-database vector storage, scalable vector processing, and the ability to conduct natural language contextual conversations based on unstructured content. Using HeatWave GenAI, developers can use built-in embedding models to create vector storage for enterprise unstructured content with a single SQL command. Users can perform natural language searches in a single step using in-database or external LLM. Data does not have to leave the database, and because HeatWave has a large scale and ultra-high performance, users do not need to provision GPUs. As a result, developers can reduce application complexity, improve performance, strengthen data security, and reduce costs.

It is not difficult to see that Oracle's idea is to provide a unified operating platform for AI and data, which is in sharp contrast to other database products.

For example, the in-database LLM feature enables users to perform the tasks required to develop models and applications without having to export data to a potentially unsecure environment or import potentially unsecure LLMs into their data environment. Since there is no need to export or import, there is no cost typically associated with exporting large amounts of data or importing large amounts of LLMs; in-database vector storage eliminates the need to move data to a separate vector database or have AI expertise.

As for the vector database that the industry is concerned about, Titanium Media has previously analyzed that if database manufacturers do not develop vector databases independently, they will basically advocate supporting native vector word embedding and vector search engines.

At present, 23ai is also proving itself through its products:Vector retrieval should be a built-in capability of the database rather than an independent product.If both types of data are managed by a single database, searching for a combination of business and semantic data will be easier, faster, and more accurate. The solution that supports this path is a database that can manage all data in a high-performance and very economical way. In the view of Wu Chengyang, vice president of Oracle and managing director of China, "all data should be placed in one place. This makes it much easier to ask questions and inquire."

"Today, most people take the data from the database to AI and then take it out again, which often involves data security issues, management authority issues, etc. Oracle's approach is to bring AI to the database and embed the vector database into the entire database. Not only vectors, but also a fusion database that can integrate multiple types of data such as text, graphs, JSON, etc. Only Oracle can do this."Wu Chengyang said.

Li Jia, senior director of Oracle China's technical consulting department, shared a case with Titanium Media: a corporate customer migrated from an open source vector database to an Oracle Fusion Database. The driving factors behind this are threefold:First, in terms of application architecture, the original application architecture involves different technology stacks, and has high management complexity and low efficiency. Second, there are performance issues when expanding data and architecture. Third, it is impossible to integrate with existing business data, and the efficiency of the overall retrieval process is often low.In Li Jia's opinion, more and more customers are making this choice and it is no longer an isolated case.

"Some customers put label information in MongoDB, permission information and identity information in MySQL, knowledge graphs in graph databases, and then store vector data such as documents in vector databases, which makes application integration more difficult," said Li Jia.

Wu Chengyang pointed out that migration itself is not complicated. The key is that customers need to compare and feel which technical solution (fusion or other) is more suitable for them. Customers think that data is very important, but except for professional DBAs, customers are often indifferent to databases. Today's databases are not about fashionable technical terms, but about how the database should be used based on the customer's experience.

To this end, Oracle has also proposed a modern data platform including "4 Any", namely Anytime, Anywhere, Any Data, Anyone, with the goal of simplifying data management, development and generation.

AI for DB enters the next competition cycle

Overall, Oracle's AI strategy is formulated around the actual scenarios in which enterprises use AI, creating an end-to-end generative AI matrix covering the entire technology stack. This includes AI infrastructure construction support based on Oracle Cloud Infrastructure (OCI), database products such as Oracle Database, Oracle Autonomous Database, and MySQL HeatWave that provide data for AI, and SaaS applications such as ERP, HCM, and CX that have built-in generative AI functions.

In its recent fiscal year financial report, Oracle released an important piece of information: In the fourth quarter alone, Oracle signed more than 30 AI sales contracts with a total value of more than US$12.5 billion, including an important cooperation to expand the Microsoft Azure platform to OCI to support OpenAI's needs for computing power such as reasoning.

The competition for big models is very fierce now. The iteration speed of big model products has been significantly accelerated recently, which puts high demands on the speed of model training. The more GPUs, the larger the data set, the larger the corpus, the stronger the infrastructure capabilities provided, and the shorter the training time, the faster the new product update speed can be.

"At present, Oracle's largest computing cluster can reach 30,000 cards, and the scale may be even larger in the future." Ji Xiaofeng, senior director of Oracle's China Technical Consulting Department, pointed out that OCI has been committed to providing advanced AI and HPC infrastructure since day one.We have specifically optimized the network and built a lossless network system to make the scalability of the entire GPU even more powerful.

OCI Supercluster can achieve the collaboration of multiple GPUs, and Oracle is about to release a high-performance file system to better meet customers' training needs. With the new OCI Compute bare metal instance, ultra-low latency RDMA network and high-performance storage, the speed of OCI Supercluster will be significantly accelerated. OCI will launch a model using NVIDIA B200 to maximize the help enterprises cope with the growing demand for AI models.

It is worth noting that in 2022, Oracle and NVIDIA announced a long-term partnership aimed at introducing NVIDIA's complete accelerated computing stack into OCI. Today, OCI has become NVIDIA's hyperscale cloud technology provider, providing large-scale AI computing services NVIDIA DGX Cloud.

Ji Xiaofeng explained: "Although we now have the MoE model, a lot of computing power is still required in the inference stage. The cooperation between Oracle and NVIDIA is different from previous cooperation between partners. In the implementation of some core services, the product departments of both sides have in-depth cooperation."

In a sense, Oracle is no longer just a database company. In recent years, Oracle's investment in OCI, SaaS and other aspects has made it a cloud computing company like Microsoft and Google. Therefore, to understand Oracle's investment logic at the database level, we cannot simply copy the limitations of database technology products, and we cannot judge Oracle's more opening paths in the Chinese market from the perspective of domestic substitution.

Currently, the public cloud version of 23ai has been launched, and the local version is expected to be launched in the second half of this year. This means that the threshold for Chinese corporate customers to use 23ai will be greatly reduced.

In the past few years, Oracle has been constantly emphasizing that in serving the "dual circulation" expansion logic of China's overseas expansion and multinational companies' business in China, Oracle's cooperation with Chinese corporate customers is also refreshing its understanding of user demands.

(This article was first published on Titanium Media APP. Author: Yang Li, Editor: Gai Hongda)