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

At the GTC conference, Lao Huang made an earth-shaking sacrifice of the H100 NVLINK specially built for ChatGPT, and said that Nvidia is the TSMC in the industry.

Nvidia, win!

At the just-concluded GTC conference, relying on a screen full of "generative AI" and holding an H100 NVLINK chip that supports ChatGPT's computing power and speeds up 10 times, Lao Huang almost wrote these words on his face-"I am the winner."

ChatGPT,Microsoft 365 AzureMAX stable Diffusion,DALL-E,Midjourney. Nvidia can get a piece of all the most popular and popular AI products nowadays.

The global popularity of ChatGPT at the beginning of this year caused Nvidia's share price to soar, directly increasing its market capitalization by more than $70 billion. Currently, Nvidia has a market capitalization of $640 billion.

Now that AI's iPhone moment has come, the fourth technological revolution is about to begin, and Nvidia, with the A100 and H100, may be the biggest winner.

At the GTC conference, Lao Huang announced Nvidia's remarkable progress in GPU, acceleration library, computing lithography and cloud platform, and even made a bold statement that Nvidia wants to be the TSMC of AI circle!

Now it has been speculated that today's lectures were generated using the AIGC model on H100.

ChatGPT dedicated GPU has come to the most important launch at this conference, which is the NVIDIA H100 NVLINK built for ChatGPT.

Because of the huge demand for computing power, according to the reasoning of LLM such as ChatGPT, Nvidia launched a new Hopper GPU, which is equipped with double GPU NVLINK PCIE H100 and 94GB memory.

In fact, the history of deep learning has been closely related to Nvidia since 2012.

Lao Huang said that in 2012, Hinton, a veteran of deep learning, and students Alex Kerchevsky and Ilya Suskever used GeForce GTX 580s to train AlexNet.

Subsequently, AlexNet won the first place in the ImageNet image classification competition, which became the singularity of the deep learning explosion.

Ten years later, Ilya Suskever in OpenAI also used Nvidia's DGX to train GPT3 and GPT3.5 behind ChatGPT.

Lao Huang said proudly that the only GPU on the cloud that can actually handle ChatGPT is HGX A100.

But compared with the A100, a server with four pairs of H100 and dual GPU NVLINK is 10 times faster! Because H100 can reduce the processing cost of LLM by an order of magnitude.

With the wave of opportunity set off by generative AI, AI is at a turning point, which makes the reasoning workload grow step by step.

In the past, designing a cloud data center to handle generative AI was a huge challenge.

On the one hand, ideally it is best to use an accelerator to make the data center resilient; on the other hand, no accelerator can optimally handle the diversity of algorithms, models, data types, and sizes. Nvidia's One Architecture platform has both acceleration and flexibility.

Today, Nvidia announced the launch of a new reasoning platform. Each configuration is optimized for a certain type of workload.

For example, for AI video workload, Nvidia launched L4, which is optimized in video decoding and transcoding, video content review, and video call functions.

One 8-GPU L4 server will replace more than a hundred two-socket CPU servers for processing AI video.

At the same time, Nvidia also launched the L40 for generative AI such as Omniverse, graphics rendering, and text-to-image / video. Its performance is 10 times that of Nvidia's most popular cloud reasoning GPU T4.

At present, the powerful power of Runway's Gen-1 and Gen-2 generative AI model is with the help of Nvidia's GPU.

In addition, Nvidia also launched a new super chip Grace-Hopper, which is suitable for recommendation systems and vector databases.

Challenge chip limit, computational lithography speed up 40 times in the chip field, Nvidia in conjunction with TSMC, ASML and Synopsys, it took 4 years to complete a major breakthrough in computational lithography technology-NVIDIA cuLitho computational lithography library.

After reaching the limit of the 2nm process, lithography is the breakthrough point.

Computational lithography simulates the behavior of light interacting with photoresist after passing through an optical element. by applying the inverse physical algorithm, we can predict the pattern on the mask in order to generate the final pattern on the wafer.

In the field of chip design and manufacturing, computational lithography is the largest computing workload, which consumes tens of billions of CPU hours every year. By contrast, the new algorithm created by Nvidia allows increasingly complex computational lithography workflows to be executed in parallel on GPU.

To sum up, cuLitho can not only increase computing speed by 40 times, but also reduce power consumption by as much as 9 times.

For example, the Nvidia H100 requires 89 mask plates.

It takes two weeks for each mask to be processed with CPU. If you run cuLitho on GPU, it only takes 8 hours to process a mask.

TSMC can also use 4000 Hopper GPU out of 500 DGX H100 systems to do work that previously required as many as 40000 CPU-based servers, and the power will be reduced from 35MW to 5MW.

It is worth noting that the cuLitho accelerator library is also compatible with GPU for Ampere and Volta architectures, but Hopper is the fastest solution.

Lao Huang said that because lithography technology is already at the limit of physics, wafer factories can increase production and prepare for 2nm and future development.

The iPhone moment of AI these months, ChatGPT is on the verge of setting off a fourth scientific and technological revolution. The saying that "we are in the iPhone moment of AI" has also spread widely.

At the GTC conference, Lao Huang also excitedly repeated this sentence three times.

When the time comes for iPhone, startups such as OpenAI are racing to build disruptive products and business models, while established companies such as Google and Microsoft are looking for ways to deal with it.

Their various actions are all caused by the sense of urgency of formulating AI strategy caused by generative AI in the world.

Nvidia accelerated computing began with the DGX AI supercomputer, which is the engine behind the current breakthroughs in large language models.

On GTC, Lao Huang proudly said that I personally gave the world's first DGX to OpenAI.

Since then, half of the Fortune 100 companies have installed DGXAI supercomputers.

The DGX is equipped with eight H100 GPU modules, while the H100 is equipped with a Transformer engine that can handle amazing models like ChatGPT.

Eight H100 modules are connected to each other through NVLINK Switch to achieve comprehensive non-blocking communication. Eight H100s work together, just like a giant GPU.

To Lao Huang's excitement, Microsoft announced that Azure would release a private preview to its H100 AI supercomputer.

And said, "the DGX supercomputer is a modern AI factory. We are in the iPhone moment of AI. "

Bringing out ChatGPT with one hand over the past decade, the combination of acceleration and scale-up has improved the performance of a variety of applications millions of times.

The most impressive example is the introduction of the AlexNet Deep Learning Framework in 2012.

At that time, Alex Krizhevsky, Ilya Suskever, and Hinton used 14 million graphs on the GeForce GTX 580to complete the training, which could handle 262 terabytes of floating-point operations.

Ten years later, Transformer came out.

Ilya Suskever trained GPT-3 to predict the next word, requiring 1 million times more floating-point operations than training the AlexNet model.

As a result, AI--ChatGPT, which shocked the world, was created.

Sum up with a sentence from Lao Huang:

This means that a new computing platform has been born and the "iPhone moment" of AI has come. Accelerated computing and AI technology have come into reality.

Acceleration library is the core of accelerated computing. These acceleration libraries connect a variety of applications, and then connect to a variety of industries to form a network in the network.

After 30 years of development, thousands of applications have been accelerated by Nvidia's library, covering almost every field of science and industry.

Currently, all Nvidia GPU are compatible with CUDA.

The existing 300 acceleration libraries and 400 AI models cover a wide range of fields such as quantum computing, data processing, machine learning and so on.

At this GTC conference, Nvidia announced that it had updated 100 of them.

The Nvidia Quantum platform consists of libraries and systems that allow researchers to advance quantum programming models, system architectures and algorithms.

CuQuantum is an acceleration library for quantum circuit simulation, among which IBM, Baidu and other companies have integrated this acceleration library into their simulation framework.

Open Quantum CUDA is Nvidia's hybrid GPU-Quantum programming model.

Nvidia also announced the launch of a quantum control link, which was developed in partnership with Quantum Machines. It can connect the Nvidia GPU to a quantum computer and correct errors very quickly.

There is also the launch of a new RAFT library to speed up indexing, data loading, and nearest neighbor search.

In addition, Nvidia announced DGX Quantum, built with DGX and leveraging the latest open source CUDA Quantum, a new platform that provides a revolutionary high-performance and low-latency architecture for researchers engaged in quantum computing.

Nvidia has also introduced NVIDIA Triton Management Service software that automatically extends and orchestrates Triton reasoning cases throughout the data center. It is suitable for multi-GPU and multi-node reasoning like GPT-3 big language model.

CV-CUDA for computer vision and VPF for video processing are Nvidia's new cloud scale acceleration library.

Lao Huang announced that CV-CUDA Beta has optimized pre-processing and post-processing to achieve higher cloud throughput and reduce cost and energy consumption by 1/4.

At present, Microsoft handles visual search and Runway generates AI video processing, all using CV-CUDA and VRF libraries.

In addition, Nvidia accelerated computing has helped to achieve a milestone in genomics. The use of Nvidia-powered equipment to reduce the cost of sequencing the entire genome to $100 is another milestone.

The Nvidia NVIDIA Parabrics Accelerator Library can be used for end-to-end genome analysis in the cloud or in instrumentation, and is suitable for a variety of public cloud and genomics platforms.

ChatGPT is running, Nvidia is making money now, ChatGPT, Stable Diffusion, DALL-E and Midjourney have awakened the world's awareness of generative AI.

Only two months after its launch, Deep-Fried Chicken ChatGPT has more than 100 million monthly active users, making it the fastest-growing app in history.

It can be said that it is a computer. It can not only generate text, write poems, rewrite research papers, solve mathematical problems, and even program.

Many breakthroughs have created today's generative AI.

Transformer can learn context and meaning from data relationships and dependencies in a massively parallel manner. This enables LLMs to use huge amounts of data to learn and perform downstream tasks without clear training.

In addition, the physics-inspired diffusion model can generate images through unsupervised learning.

Lao Huang concluded that in just more than a decade, we have gone from identifying cats to generating cats in spacesuits that walk on the moon.

Now it can be said that generative AI is a new kind of computer, a computer that can be programmed in human language.

Previously, ordering computers to solve problems was the sole prerogative of programmers, but now everyone can be programmers.

Like Bill Gates, Lao Huang has a similar definition: generative AI is a new computing platform similar to PC, the Internet, mobile devices and the cloud.

Through Debuild, we can design and deploy Web applications directly as long as we are clear about what we want.

It is clear that generative AI will reshape almost all industries.

In order to be the "TSMC" of AI, professional companies need to use their own proprietary data to build customized models.

Then, Lao Huang proudly announced that the industry needs a contract factory similar to TSMC to build a custom large language model, and Nvidia is this "TSMC"!

At the conference, Nvidia announced the launch of NVIDIA AI Foundations cloud services, allowing customers to customize LLM and generative AI.

This cloud service includes language, visual and biological modeling services.

Among them, Nemo is used to build a custom language text-to-text generation model.

Picasso is a visual language model that can be used to train custom models, including images, videos, and 3D applications.

As long as a text prompt and an API call to metadata are sent to Picasso, Picasso uses the model on DGX Cloud to send the generated footage back to the application.

What's even more powerful is that by importing these materials into NVIDIA Omniverse, you can build realistic meta-universe applications and digital twin simulations.

In addition, Nvidia is working with Shutterstock to develop an Edify-3D generation model.

At the same time, the partnership between Nvidia and Adobe continues to expand, incorporating generative AI into the daily workflows of marketers and creative people, with particular attention to artist copyright protection.

The third area is biology.

Today, the value of the drug R & D industry has reached nearly 2 trillion yuan, and the R & D investment is as high as 250 billion US dollars.

NVIDIA Clara is a medical and health application framework for imaging, instrumentation, genomic analysis and drug development.

Recently, the hot direction of the biosphere is the use of generative AI to discover disease targets and design new molecular or protein drugs.

Accordingly, BIONEMO allows users to use proprietary data to create, fine-tune and provide custom models, including AlphaFold, ESMFold, OpenFold and other protein prediction models.

Finally, Lao Huang concluded that NVIDIA AI Foundations is a cloud service and contract factory for building custom language models and generative AI.

Old Huang Yun service, which rents for $36999 a month, Nvidia this time, also launched a cloud service.

It is keenly aware of customers' need for easier and faster access to NVIDIA AI, so it launched NVIDIA DGX Cloud.

DGX Cloud works with Microsoft Azure, Google GCP and Oracle OCI. As long as a browser, NVIDIA DGX AI supercomputer, you can instantly access every company!

On this cloud, you can run the NVIDIA AI Enterprise acceleration library suite to directly solve the end-to-end development and deployment of AI.

Moreover, the cloud provides not only NVIDIA AI, but also some of the major cloud service providers in the world.

And Nvidia's first NVIDIA DGX Cloud was Oracle Cloud Infrastructure (OCI).

In OCI, the two kings, NVIDIA CX-7 and BlueField-3, immediately combine to create a powerful over-the-counter.

According to reports, enterprises can now rent DGX Cloud, starting at US $36999 a month.

Finally, of course, it is the reserved program of the annual GTC conference-Omniverse. Lao Huang announced the update of meta-cosmos platform Omniverse.

Now, Microsoft and NVIDIA are preparing to bring Omniverse to hundreds of millions of Microsoft 365 and Azure users.

In addition, it is also reported that in order to allow the H100 to be exported to China in compliance, Lao Huang specially adjusted a "H800" based on the previous experience of the A800, reducing the data transfer rate between chips to about 50% of the H100.

To sum up, Lao Huang has made it quite clear at this conference that Nvidia is going to be a TSMC in the field of AI, providing contract manufacturing like a fab, and on this basis, let other companies in the industry train algorithms.

Can this business model be successful?

Reference:

Https://www.NVIDIA.com/gtc/keynote/

This article comes from the official account of Wechat: Xin Zhiyuan (ID:AI_era)

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.