Under the big wind of large models and generative AI, NVIDIA vigorously advocates and takes the lead in entering the high-end market of traditional data centers-accelerated computing data centers. Since entering the era of AI and big data, the concept of "accelerated computing" has emerged in the industry. Simply understood, accelerated computing means using acceleration chips such as GPUs, FPGAs, and various DPUs and related software technologies to "accelerate" traditional data center computing to meet the needs of specific computing scenarios such as AI, big data, and edge computing.
Since the explosion of large-scale models and generative AI in early 2023, accelerated computing has quickly captured the industry's "headlines" overnight, and NVIDIA has also rushed to a trillion-dollar market value in this wave of market enthusiasm. The eyeballs of the industry have earned "a lot of money". In order to further consolidate accelerated computing, NVIDIA launched a full set of generative AI data center solutions at COMPUTEX 2023 not long ago, especially the world's first Spectrum-X Ethernet network architecture specially built for generative AI, intending to create a A brand-new AI cloud high-performance data center market.
The newly launched NVIDIA Spectrum-X network platform is a reshaping of the existing data center network. On the basis of Ethernet, NVIDIA BlueField-3 DPU is used to accelerate the network and applications on the host side, coupled with the high performance of NVIDIA Spectrum-4 Ethernet switches, and corresponding software, can realize GPU-GPU efficient communication in the cluster, so that hundreds of servers can be connected into a large cluster, forming a super-scale with thousands or even tens of thousands of GPU cards Data center to meet the requirements of large model training, this is the generative AI data center.
Previously, NVIDIA has launched the NVLink Switch technology, which has been developed to the fourth generation. It can realize cross-server GPU-GPU communication in the cluster, and can connect up to 32 servers and 256 GPU cards. The latest DGX GH200 released at COMPUTEX 2023 The supercomputer is the first "giant GPU" built with NVLink Switch technology - 256 GH200 Grace Hopper chips and 144TB shared memory are connected to build a shared memory platform with 256 GPUs.
NVIDIA is building a generative AI cloud supercomputer called Israel-1 in Israel. Israel-1 consists of 256 HGX-based Dell servers, each with 8 GPUs, that is, a total of 2048 GPU cards and 2560 With a BlueFiled-3 DPU and more than 80 Spectrum-4 switches, it can reach a peak AI performance of 8 EFlop/s, and will be one of the world's top-ranked AI supercomputers.
As early as 2020, NVIDIA completed the acquisition of Mellanox Technology, an Israeli provider of high-performance connectivity solutions for servers and storage. senior talent. Of course, Mellanox is a firm promoter of InfiniBand technology, and InfiniBand is a network technology that uses RDMA technology to directly access remote GPU memory. NVIDIA Quantum-2 InfiniBand platform is another network platform of NVIDIA, used to achieve more high-end GPU interconnection network.
It can be seen from the configuration of Israel-1 that Israel-1 is to verify the feasibility of building a high-performance Ethernet-based generative AI data center to the server industry and data center industry.
Compared with traditional data centers based on existing Ethernet technology solutions, Israel-1 provides a brand-new design architecture. By using BlueFiled-3 DPU to accelerate network and key applications on the host side, Spectrum-4 replaces traditional switches to realize network High-speed data exchange and business performance isolation in the center, LinkX helps to ensure end-to-end data integrity, NVIDIA full-stack AI software provides an efficient and easy-to-use operating platform, meeting the performance pursuit of the generative AI data center on the Ethernet platform .
Simply understand, NVIDIA is creating a brand new data center market, a future-oriented data center market, rather than attacking the traditional data center market, that is, the high-end AI factory and the luxury version of the generative AI data center, which is equivalent to The highest configuration of Android and the highest configuration of Apple in smart phones have completely subverted the traditional mobile phone market. Whether NVIDIA can create brilliance in this market? It also depends on the extent to which generative AI can truly become an industry, rather than eventually calm down like traditional AI.
Following Goldman Sachs economists, McKinsey released the report "Generative AI Economic Potential". McKinsey analyzed and surveyed 850 occupations, 63 application examples and more than 2,100 detailed work activities in 47 countries, basically covering 80% of the world. labor force. The 63 generative AI use cases analyzed by McKinsey cover 16 business scenarios, which can bring economic benefits of 2.6-4.4 trillion US dollars to the world every year. McKinsey also found that when generative AI is applied to the positions of knowledge workers, it can add 6.1-7.9 trillion US dollars of economic benefits to the world every year, which is higher than the use cases of business scenarios. McKinsey estimates that from 2023 to 2040, the automation of business processes through generative AI can provide 0.2% to 3.3% of global productivity.
In general, the business prospects of generative AI are rapidly taking shape, and can actually bring productivity and productivity improvements to everyone and all enterprises, and NVIDIA's "bet" on generative AI data centers may be for The global data center industry has brought a new storm.