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

The more chips you buy, the more you save? With only one computer, you can fine-tune the model at home? The meta-universe, which used to be time-consuming and costly, can now quickly generate digital twin models and quickly build virtual worlds with generative AI.

On August 8, Leather Master Huang Renxun delivered a NVIDIA keynote speech at SIGGRAPH with a series of updates from Nvidia. SIGGRAPH is the annual top conference of computer graphics and an important organization in the field of computer graphics and interactive technology.

Since its founding in 1993, Nvidia has taken root in the field of computer graphics, driving the development of graphics processor (GPU) technology, including real-time ray tracing technology breakthroughs, multi-core and multi-core processor innovation, and accelerated algorithm training and reasoning in the fields of artificial intelligence and deep learning. It was also five years later that Leather Master Huang Renxun once again stepped onto the stage of SIGGRAPH.

So what surprises did the Leather Master bring to all kinds of spectators in this speech? Draw the point!

1. GH200+Grace Hopper: reducing the cost of large language model reasoning in May this year, Nvidia released the GH200 system, which is designed to handle large-scale generative artificial intelligence (Generative AI) workloads. GH200 fully connects 256NVIDIA Grace Hopper ™Superchip into a single GPU, which can support models dealing with large-scale recommendation systems, generative artificial intelligence and graph analysis.

In this speech, Huang Renxun introduced in detail his "precious" GH200. In his view, the future cutting-edge model will be trained and applied in large-scale systems. There will be a single Grace Hopper on each node. This architecture has been the way of computing for the past 60 years. Now it will become the mainstream in accelerated computing and AI computing. In the future, it will enable cutting-edge models to be better trained and applied.

In other words, this kind of system can be applied universally, and the frontier model in the future will become the front end of various applications. Every application, every database, when interacting with a computer, may first interact with a large language model. These models will understand the user's intentions, desires, and specific situations, and present the information in the best way. This model will conduct intelligent query and search, and may even be used in areas such as music generation. "in the future of computing, accelerated computing and AI computing will become mainstream." Huang Renxun looked forward to it with confidence.

In terms of energy efficiency and cost-effectiveness, Huang Renxun repeatedly stressed: "the more you buy, the more you save." This is not a slip of the tongue, accelerating the calculation of Moore's law equivalent to 20 times the energy efficiency and cost-effectiveness of generative AI applications and the current CPU expansion. Moore's law is a view put forward by Gordon Moore, one of the founders of Intel. Its core content is that the number of transistors that can be held on an integrated circuit doubles about every 18 to 24 months. In other words, the performance of the processor doubles roughly every two years, while the price falls to half what it used to be. So, what is the concept of the 20-fold Moore law?

Huang Renxun gives a vivid example: specifically, to build a data center, GPU using general computing costs $100 million, while Grace Hopper using accelerated computing costs only $8 million. And the energy consumption using general calculation is megawatts, while using Grace Hopper requires only 262kW (262000 watts), which reduces the energy consumption 20 times. In terms of cost, the cost of using Grace Hopper is 12 times lower than that of general computing. The use of accelerated computing can significantly reduce energy consumption and significantly reduce costs. In other words, the more GH200 you buy, the more cost savings you will achieve.

Second, the Omniverse of attack: is the meta-universe within reach? Before generative AI can automatically generate digital twin model, the construction of digital twin model is a relatively complex process, which involves the drive of multi-dimensional virtual model and fused data. Whether it is the process of multi-dimensional virtual model construction, data acquisition and sensor integration, model fusion and calibration, interactive simulation and optimization, the previous digital twin model construction process depends more on the experience and manual operation of professionals. it is necessary to integrate and calibrate the actual data with the virtual model to realize the functions of monitoring, simulation, prediction and optimization.

Such a production process, not to mention individuals, even many "deep-pocketed" enterprises are also prohibitive. However, the update of Nvidia Omniverse further lowers the threshold for digital twinning, and developers, enterprises and industries have been able to use the OpenUSD framework and generative AI to optimize and improve 3D processes, that is, developers are free to create their own virtual assistants and digital people on Omniverse, and enterprises can automatically generate digital twin ads for products on this platform.

Huang Renxun showed off the Omniverse-based partnership between WPP, the world's largest advertising company, and BYD, an electric car maker. OmniVerse Cloud allows BYD to create real-time digital twins using high-fidelity data to achieve physically accurate simulations. WPP artists can collaborate seamlessly using tools such as AutoDesk, Adobe and SideFX in the same environment, enabling BYD to connect designers and developers in different time and space through OmniVerse Cloud, thus quickly building a virtual world with physical accuracy.

To feel the cool digital twin blockbuster generated, the car color and scene can be changed at will, one second or black, the next second can turn red, the last picture is still in the prairie, the next frame directly appears in the snow!

Huang Renxun further explained that Open USD technology enables WPP to create a super digital twin that integrates all possible changes into a single asset. Deployed on the Universe Class GDNA network, the digital twin implements a fully interactive 3D configurator that delivers a high-fidelity real-time 3D experience to devices in more than 100 regions around the world. The solution can also generate personalized content for global marketing campaigns, and USD models are located in a 3D environment and can be created using real-world scanned data or generative AI tools.

At the same time, the Machinima application has also been updated to help users better build realistic virtual images. The introduction of NVIDIA Omniverse Avatar Cloud Engine (ACE), as well as new Omniverse connectors and applications, allows users to easily build and customize virtual assistants and digital people.

At present, Omniverse users can enhance the interaction in USD (Universal Scene Description) workflow through OmniLive and introduce the real-time accuracy of the real world into the 3D virtual world.

3. AI Workbench: can Fine-tune,AI stack BUFF for you at home? With only one computer, you can test and fine-tune the model at home? To sum up, in the words of Leather Master Huang Renxun, "everyone can generate AI."

This is no longer a fantasy. Huang Renxun announced at the scene that Nvidia has released a new unified workspace called NVIDIA AI Workbench. The platform provides developers with a unified, easy-to-use workspace that enables them to quickly create, test, and fine-tune generative AI models on personal computers or workstations, and then extend these models to almost any data center, public cloud, or DGX Cloud.

At the same time, Huang Renxun announced that NVIDIA, together with startup Hugging Face, will provide generative AI supercomputing to millions of developers to help them build advanced AI applications such as large language models. Developers will be able to train and adapt advanced AI models using NVIDIA DGX Cloud AI supercomputing within the Hugging Face platform.

Specifically, how will AI Workbench and Hugging Face help beat workers free their hands?

NVIDIA's NVIDIA RTX 6000 workstation GPU, based on the next-generation Ada Lovelace architecture, provides designers and creators with powerful tools with 2x-4x performance improvements. This GPU enables designers and engineers to drive advanced simulation-based workflows to build and validate more complex designs. Artists can push the narrative to a new height, create more fascinating content, and build an immersive virtual environment. Scientists, researchers, and medical professionals can use supercomputing power on workstations to accelerate the development of life-saving drugs and programs with 2-4 times the performance of the previous generation of RTX A6000.

NVIDIA's RTX 6000 Ada Generation GPU uses Ada architecture AI and programmable shader technology to provide an ideal platform for neural graphics and advanced virtual world simulations for creating metaspheric content and tools using NVIDIA Omniverse Enterprise.

In addition to RTX 6000, NVIDIA also launched three new workstations GPU:RTX 5000, RTX 4500 and RTX 4000, each with different specifications. The RTX 5000 uses an AD102 chip, which is a reduced version of RTX 6000 and provides 12800 CUDA cores and 32GB GPU memory. The RTX 4500 uses an AD104 chip with 24GB GPU memory. The RTX 4000 is an entry-level workstation GPU with 20GB GPU memory and a 6144 CUDA core. These new GPU will provide designers, creators and engineers with a wider choice to meet the needs of different areas. In other words, with AI Workbench and Hugging Face, workers can stack BUFF and use the platform to complete more elaborate and complex content.

There is no doubt that the progress of NVIDIA in generative artificial intelligence, digital twin modeling, the launch of AI Workbench and cooperation with Hugging Face are promoting the future of artificial intelligence, virtual reality and content creation, lowering the threshold of generative AI, so that "will not use the low cost of AI to use AI, will use AI more skillfully and more cost-saving", to provide "accelerators" to many industries. Leather guru Huang Renxun is constantly pushing the boundaries of technology in the fields of artificial intelligence, graphics and simulation, bringing new surprises.

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.