Nvidia Huang Renxun: Big models are like brains, nothing can be done without applications

Interface News Reporter | Cui Peng

On October 24th, at the Lenovo TechWorld conference, NVIDIA founder, president, and CEO Huang Renxun delivered a speech and had a conversation with Lenovo Group chairman and CEO Yang Yuanqing.

This year, Huang Renxun has talked about the opportunities brought by artificial intelligence on various occasions. At this conference, Huang Renxun mentioned many technical issues that generative artificial intelligence will face in the vertical industry.

This includes various issues such as how to vectorize the database, how to put semantic features into the database, how to create applications based on generative artificial intelligence, how to deploy, install, and run operating systems, and how to coordinate these workloads and run them in the data center.

Based on these challenges, Huang Renxun and Yang Yuanqing announced on-site that Nvidia will collaborate with Lenovo to launch a hybrid artificial intelligence plan. After cooperation, Lenovo will provide enterprise level artificial intelligence solutions based on NVIDIAMGX architecture and Lenovo's hybrid artificial intelligence services to the outside world.

Huang Renxun stated that Lenovo's partnership with NVIDIA has been ongoing for 25 years, with cooperation in workstations, gaming PCs, and high-performance computers, and now expanding to the fields of AI and intelligent infrastructure.

In his speech, Huang Renxun mentioned the concept of an "AI factory", which is a specialized computing infrastructure paired with Nvidia's supercomputers specifically designed to optimize artificial intelligence.

This AI factory has only one purpose, which is to process a large amount of raw data that is about to enter it and output intelligence, "said Huang Renxun.

For the data issues involved in this concept, he believes that currently, whether it is traditional voice, image and text data, or various unstructured data (such as chemicals or sports), any information can be digitized.

After adding AI support, computers can not only understand this data, but also generate more data. In Huang Renxun's view, there are great opportunities for technology companies here.

Because the vast majority of data in the world is embedded within enterprises, kept confidential and strictly regulated, it cannot be transferred to cloud storage. And what NVIDIA needs to do is create an infrastructure and solution stack, so that these enterprises can confidently use artificial intelligence technology.

Huang Renxun believes that there are several parts of work involved here, first of all, the pre training of models. These basic models exist on NVIDIA's cloud, and NVIDIA needs to optimize them. The faster this process is processed, the lower the cost for the enterprise.

Secondly, there is the question of "how do we view artificial intelligence models". Huang Renxun believes that it is essentially similar to the brain, and unless it is turned into an application, "big models will do nothing (can only think).

He mentioned the concept of retrieval augmentation generation (RAG) here, which is another way to create chat robots. The functionality of chat robots can be enhanced through user data, and Nvidia has the ability to quantify this data to form a semantic database. Users can talk to this database and ask questions, and it can respond in a way that users understand.

In NVIDIA's view, the way to build applications in the future will no longer be traditional coding or SQL language, but rather integrate AI like an integration team, and these AI will perform various tasks on behalf of users.

In addition, over the past four years, Nvidia has also collaborated with VMware to transform the operating system originally designed for virtualization into an operating system for artificial intelligence.

According to Huang Renxun's introduction, this system runs on NVIDIAAIEnterprise for data processing and deep learning. From training, fine-tuning, to model inference and deployment, the entire stack runs on Lenovo's server.