On March 21, local time, NVIDIA's headquarters in Santa Clara, California, USA, opened the 2023 annual technology summit GTC for software developers. NVIDIA founder and CEO Jensen Huang disclosed new artificial intelligence and chip technology, and said that the "iPhone moment" of artificial intelligence has arrived.
Nvidia's GPU chips are providing computing power for most of the AI systems used around the world. In his keynote speech at GTC, Jensen Huang explained that from AI training to deployment, from semiconductors to software libraries, from systems to cloud services, technological breakthroughs will make AI within reach.
The DGX supercomputer is the modern AI factory
"If accelerated computing is compared to a warp speed engine, then AI is the source of power, and the extraordinary capabilities of generative AI create a sense of urgency for companies, and we have to reimagine products and their business models." Huang said, "Moore's Law is slowing down, and the emergence of accelerated computing and AI is timely. ”
NVIDIA's supercomputer DGX has become the "new factory" of modern AI. Huang said that the deployment of large language models (LLM) similar to ChatGPT is an important new inference workload, and in order to support large language model inference, NVIDIA released a new GPU, H100NVL with dual-GPU NVLink.
According to Huang, the H100 based on NVIDIA's Hopper architecture is equipped with a Transformer engine that can handle GPT-like models, and compared with the HGX A100 for GPT-3 processing, the standard server equipped with four pairs of H100 and dual GPU NVLink can be up to 10 times faster.
"H100 can reduce the processing cost of large speech models by an order of magnitude." Huang said. Based on the H100 chip, NVIDIA built the latest DGX supercomputer, equipped with 8 H100 GPUs, connecting them into a huge GPU, providing a "blueprint" for the construction of AI infrastructure, and the new DGX supercomputer has been fully put into production.
Over the past 10 years, cloud computing has grown at a rate of 20% per year and has rapidly grown into a trillion-dollar industry. In order to quickly bring DGX's capabilities to startups and other types of enterprises, NVIDIA also released DGX cloud services, through cooperation with Microsoft Cloud Azure, Google Cloud, Oracle Cloud infrastructure, large-scale supercomputer systems can be equipped with up to 32,000 NVIDIA chips, from the "browser" to bring DGX's AI supercomputer capabilities to each company.
Huang said that NVIDIA will bring the ecosystem to cloud service providers, expand the scale and capabilities of NVIDIA, and empower more enterprises to build generative AI capabilities.
However, such a cloud service may not be cheap. According to reports, the monthly rent of DGX cloud services is $37,000.
Huang Renxun told the first financial reporter in a media interview on March 22, Beijing time: "We cooperate with cloud service providers in Europe and the United States to provide the capabilities of NVIDIA's DGX system AI supercomputer. In China, we have specially customized Ampere and Hopper chips. However, these will provide landing capabilities through Chinese cloud providers, such as Alibaba, Tencent and Baidu, and I fully believe that they have the ability to provide top-level system services, and Chinese startups will definitely have the opportunity to develop their own big language models. ”
The Ampere and Hopper chips available in China that Huang refers to are the A800 and H800, which are currently used by most Chinese developers when developing large language models.
AI's iPhone time has arrived
At present, in addition to US giants such as Microsoft and Google investing heavily in AI large-scale language models, Chinese Internet giants and technology start-ups have also invested in the research and development of GPT-like models. Nvidia dominates the field of artificial intelligence chips, and the large demand for chips has pushed Nvidia's stock price up more than 77% this year. Nvidia's market capitalization soared to $650 billion, about five times that of Intel.
"The iPhone moment of artificial intelligence has begun." In his keynote speech at GTC, Huang said that he believes that the impact of artificial intelligence on society may be the same as Apple's iPhone opening up the smartphone market.
Hans Mosesmann, chip semiconductor analyst at brokerage Rosenblatt Securities, said Nvidia's latest product release was "years ahead of its competitors." "NVIDIA's leadership in AI software is not only a milestone, it is also accelerating." He said.
Huang also mentioned in the interview the three stages of the AI boom and how NVIDIA is preparing for the development of AI in the process.
He said that 12 years ago, Nvidia realized that deep learning would change the way software was deployed, so Nvidia reinvented the modern computer, and it is updating with each generation. "This is the first phase of AI development, building the infrastructure." He said, "The second stage is the learning perception of AI, such as machine vision, automated use cases, and so on; The third phase is the AIGC computer-generated content phase that we are going through, where AI is the co-creator and involved in all the work. ”
Huang believes that AI can help humans create first drafts, establish preliminary designs, help humans open "brain holes", stimulate human creativity, and improve production efficiency. He said that NVIDIA prepared in several ways, first of all, building infrastructure and deploying software; The second is to put this computing power in the cloud and cooperate with cloud service providers to make infrastructure more quickly shared; The third is to prepare the supply chain to meet the global demand for AI in the context of the explosive growth of AIGC's demand.
Nvidia has also launched a service called AI Foundations, to help companies train their custom AI models, which several stock image database vendors already plan to use.
Quantum computing and lithography innovations
In addition, NVIDIA also recently announced a collaboration with quantum computing researchers to accelerate software development projects, and a collaboration with chip industry giant TSMC to accelerate chip development.
On the chip manufacturing process, NVIDIA announced a new software technology for lithography machines, cuLitho, which will use NVIDIA's chips to accelerate the steps between software-based chip designs and the physical fabrication of lithography masks for printing that design on silicon. Traditional computing chips take two weeks to complete, but Nvidia's chips and software at the same time can handle these tasks overnight and reduce the power used from 35 megawatts to 5 megawatts.
NVIDIA said that cuLitho runs on GPUs, and its performance is 40 times higher than the current lithography process, which can provide acceleration for large-scale computing workloads that currently consume tens of billions of CPU hours per year.
With this technology, 500 NVIDIA DGX H100 systems can do what would otherwise take 40,000 CPUs, and they can run all the processes of the computational lithography process at the same time, helping to reduce power consumption and environmental impact. In the long run, cuLitho will lead to better design rules, higher density and yield, and AI-driven lithography.
Nvidia said it is working with ASML, Synopsys and TSMC to bring the technology to market. Huang expects TSMC to begin preparing production of the technology in June.
"The chip industry is the foundation of almost every other industry in the world," he said. Lithography is approaching the physical limit, and the launch of NVIDIA cuLitho and our partnership with TSMC, ASML, and Synopsys will enable fabs to increase yields, reduce carbon footprints, and lay the foundation for processes at 2nm and beyond." ”
For the first time, NVIDIA has positioned itself as a key player in the quantum computing space, and the mixing of quantum and classical computers is a trend. To that end, NVIDIA launched NVIDIA DGX Quantum, the world's first GPU-accelerated quantum computing system, enabling researchers to build powerful applications that combine quantum computing with advanced classical computing to advance calibration, control, quantum error correction and hybrid algorithms.
Nvidia also launched CUDA Quantum, a platform for building quantum algorithms using the popular classical computer coding languages C++ and Python, which will help run algorithms on quantum and classical computers.
"CUDA Quantum will enable domain scientists to seamlessly integrate quantum into their applications and gain access to new disruptive computing technologies." Tim Costa, head of high-performance computing (HPC) and quantum at NVIDIA, said, "One difference is that while CUDA is proprietary, CUDA Quantum is open source and was developed with input from many quantum computing companies. ”