Huang said lower computing costs have allowed AI to emerge and launch the Blackwell Ultra chip next year

CBN

2024-06-02 23:02Published on the official account of Shanghai Yicai

After hosting the head of the supply chain manufacturer in Taiwan, China on May 30, Nvidia CEO Jensen Huang stepped onto the stage of the National Taiwan University Gymnasium on the evening of June 2, dressed in a classic black leather suit, and shared how NVIDIA is promoting the evolution of AI and how AI is transforming the industry in a two-hour speech.

According to local media reports in Taiwan, it was raining in Taipei that day, and Liu Yangwei, chairman of Hon Hai Group, Lin Baili, chairman of Quanta Computer, Shi Chongtang, chairman of Asus, and other leading figures in the field of science and technology braved the rain to go to the scene, and there were many people outside the venue.

Huang talked about Nvidia's new developments, showing off the latest Blackwell chips and saying that Nvidia will launch the Blackwell Ultra AI chip in 2025, and the next-generation AI platform will be named Rubin.

Huang also reviewed how NVIDIA is driving down compute costs and AI evolution through GPU-parallel computing.

Huang said the performance of the computer industry's engines running on central processing units (CPUs) has scaled much slower, but the amount of computation we need is still growing exponentially. If the data that needs to be processed continues to grow exponentially, but the performance doesn't, we're experiencing compute bloat.

"There's a better way to take the load off the CPU and do the job well with a dedicated processor for acceleration. Computer graphics can be used in parallel computing, which can be accelerated by combining GPUs (graphics processing units) and CPUs. We can make it 100 times faster. That's what data centers do, Huang said, and $500 million of a billion-dollar data center is GPU, and it suddenly becomes an AI factory. Cost and energy savings can also be achieved by accelerating computing.

Huang said that every time an application is accelerated, the cost of computing goes down, and a 100-fold increase in speed can save 96 percent, 97 percent, 98 percent of the cost. Over the past decade, the marginal computational cost of a particular algorithm has been reduced by a factor of 1 million. "Now we can train large language models with all the data on the internet. Artificial intelligence is possible because we believe that as computing becomes cheaper, there will be people who will find good uses. ”

Huang reviewed how NVIDIA is driving the creation of large language models. He said that Nvidia changed the architecture of GPUs after 2012 to use Tensor Cores (tensor computing unit), and 10 years ago, Nvidia invented NVLink (a bus and its communication protocol). Nvidia integrates all the new technologies into a single computer.

"Nobody understood it at the time, and I didn't think anyone was going to buy it, but when we announced it at GTC, OpenAI, a small company in San Francisco, saw it, and they asked me to send them one." Huang said that in 2016, Nvidia delivered its first DGX supercomputer to OpenAI and has continued to expand the supercomputer's capabilities to train large amounts of data. So, in November 2022, based on the training of thousands of NVIDIA GPUs, OpenAI launched ChatGPT and gained millions of users in 5 days.

"What started as a supercomputer evolved into a data center, and we were able to use generative AI to generate tokens (units of text) for almost everything. In the 90s of the 19th century, Tesla invented the alternator, and we invented the 'artificial intelligence generator'. Both of these things have huge market opportunities and can penetrate into various industries, which is why a new industrial revolution will be triggered. Huang said that the models are expected to get bigger and bigger, and we will need larger and larger GPUs.

In addition, Huang showcased NVIDIA's Earth-2 digital twin Earth platform, which is used for weather observations and blends artificial intelligence, physical simulations, and observational data based on NVIDIA's CorrDiff generative AI model technology to predict the impact of tomorrow's world from today's data. Huang said Earth-2 will support hyperlocal predictions, which can simulate airflow around buildings.

Huang's trip to Taiwan has attracted much attention. Huang arrived in Taiwan on May 26 for a trip to attend COMPUTEX Taipei. On May 29, Huang Jenxun, Lin Baili, and TSMC founder Zhang Zhongmou appeared at the Taipei Night Market.

On the evening of May 30, Huang Jensen hosted a banquet for supply chain manufacturers in Taipei, including Liu Yangwei, Lin Baili, Shi Chongtang, Weiying Chairman Hong Lining, Inventec Chairman Ye Licheng, Wistron General Manager Lin Jianxun and other technology company leaders attended, and almost all of NVIDIA's supply chain partners in Taiwan were invited to participate in the dinner. The restaurant that set up the dinner was a traditional Taiwanese restaurant, and local media reported that the dinner set up seven tables, each table cost about NT$4,653, and the tech tycoons on the scene were worth more than 100 billion US dollars.

Nvidia has also recently become a star among U.S. technology stocks. Nvidia's stock price fluctuated after the release of its earnings report for the first quarter of fiscal 2025, in which revenue rose 262% year-on-year to $26 billion and net profit rose 620% year-on-year to $14.88 billion. On May 28, Nvidia's stock price exceeded $1,000 and closed at $1,148.25 per share on May 29. After a two-day decline, Nvidia closed at $1,096.33 per share on May 31, with a market capitalization of $2.7 trillion and a distance of $0.25 trillion from Apple's market capitalization. Nvidia's share price has doubled this year.

As Nvidia's stock price rose during the year, Huang's worth rose. On May 29, the Forbes real-time rich list showed that Huang's net wealth reached $1,006, exceeding $100 billion for the first time, and the Bloomberg Global Billionaires real-time list showed that Huang's net worth reached $100 billion, ranking 15th. On June 2, in the Bloomberg Global Billionaires Real-Time Rankings, Huang's net worth fell back to $9.63 billion, ranking 16th.

(This article is from Yicai)

View original image 402K

•