May 30th local time, including Google Meta、 Microsoft AMD、 Technology giants including Intel, Broadcom, Cisco, and HP have announced the establishment of a company called Ultra Accelerator Link& quot; Super Acceleration Link; quot; The UALink is a new industry open standard developed for GPU network communication systems in artificial intelligence data centers, with a clear goal of breaking Nvidia's dominant position in the market.
As the largest participant in the artificial intelligence chip market, Nvidia has not joined the alliance. NVLink technology, unique to NVIDIA, was introduced to the market ten years ago and has been at the forefront of computing network connectivity technology, continuously driving efficient interconnection between GPUs and other components.
IBM, the early partner of NVLink, also did not join the alliance. In addition, Marvell Technologies, the main competitor of Broadcom in the network and custom chip markets, has not participated in the organization.
As of the close of the US stock market on that day, Nvidia's stock price plummeted by nearly 4%. NVIDIA's competitor AMD's stock price rose nearly 1%.
With the rapid development of artificial intelligence technology, the demand for efficient and fast data transmission and processing has become increasingly urgent. Nvidia currently holds over 80% of the global AI chip market share, and technology giants such as Google and Meta are eager to reduce their reliance on Nvidia's AI chips.
These companies stated in a statement:; quot; Industry norms are crucial for establishing standardization of next-generation AI data centers and implementing AI, machine learning The interface between HPC (high-performance computing) and cloud applications is crucial& amp;quot;
Currently, Nvidia's NVLink is not open to the industry. Therefore, The UALink expert group will set standards for managing connections between different GPUs in data centers and is expected to provide these standards to companies joining the UALink alliance in the third quarter of 2024.
Nvidia did not comment on the establishment of the UALink alliance, but told First Financial reporters:; Currently, NVLink is only used for Nvidia's chip network systems& amp;quot;
&Amp; NVLink has become a standard feature of Nvidia's artificial intelligence data center system, The emergence of UALink technology is clearly aimed at competing with NVLink& Gartner chip industry analyst Sheng Linghai told First Financial reporters.
According to UALink's plan, the first UALink 1.0 version will allow direct data transfer between AMD's Instinct GPU or Intel's dedicated processors such as Gaudi, thereby improving the performance and efficiency of AI computing.
Forrest Norrod, Executive Vice President and General Manager of AMD's Data Center Solutions Division, wrote in a press release:; The UALink consortium is working hard to create an open, high-performance, and scalable accelerator architecture, which is crucial for the future of AI& amp;quot; He also stated that the goal of the alliance is to create large-scale AI and high-performance computing solutions based on open standards, efficiency, and strong ecosystem support.
Through the interconnection of standardized AI and high-performance computing (HPC) accelerators, system OEM IT professionals and system integrators will find it easier to integrate and scale AI systems in data centers. This standard aims to promote open ecosystems and the development of large-scale AI and HPC solutions.
&Amp; UALink is an important milestone in the development of artificial intelligence computing& amp;quot; Sachin Katti, Senior Vice President and General Manager of Intel's Network and Edge Business Unit, stated that; quot; We look forward to bringing a new wave of industry innovation through the UALink standard& amp;quot;
An industry insider told a reporter from First Financial News:; The high-speed I/O communication design and protocol of UALink are very powerful, demonstrating the determination of these giant companies to challenge Nvidia& amp;quot;
Nvidia's NVLink technology has become an important support behind the release of each generation of new architecture chip systems. At the 2014 NVIDIA GTC conference, the world's first high-speed GPU interconnect technology NV Link 1.0 was released, paving the way for billions of secondary computing.
Nvidia stated that, NVLink technology is used to connect GPUs with CPUs that support NVLink technology, and also to provide high bandwidth connections directly between multiple GPUs.
At the end of 2014, IBM and Nvidia have built two new flagship supercomputer systems, namely the Summit system at Oak Ridge National Laboratory and the Sierra system at Lawrence Livermore National Laboratory, NVLink is a key technology among them, and Nvidia's GPU and IBM POWER CPU are connected through NVLink interconnection technology, providing them with at least 100 petaflops of computing performance.
IBM states:& NVLink enables fast data exchange between CPU and GPU, thereby improving the data throughput of the entire computing system and overcoming a major bottleneck in accelerating computing today. NVLink allows developers to easily modify high-performance and data analysis applications to fully utilize accelerated CPU-GPU systems& amp;quot;
NVLink technology accompanies every generation of Nvidia's GPU chip updates. In 2017, Nvidia has launched NVLink 2.0 technology, connecting two GPU V100 chips with six NVLink links; In 2020, NVLink 3.0 technology emerged, connecting two GPU A100 chips through 12 NVLink links; In 2022, NVLink technology has been upgraded to 4.0, allowing two GPU H100 chips to be interconnected through 18 NVLink links.
At the 2024 GTC conference, NVLink 5.0, along with the latest generation of Nvidia Blackwell chips, has been launched, significantly improving the scalability of large multi GPU systems. A single Nvidia Blackwell Tensor Core GPU can support up to 18 NVLink 100 gigabyte per second (GB/s) connections, with a total bandwidth of 1.8 megabytes per second (TB/s), which is twice that of the previous generation product.
NVLink has gone through a decade of development, establishing high standards for the industry and solidifying its position in the data center market. The further impact of the joint launch of the new UALink standard by the eight giants on Nvidia will also be closely monitored.
NVIDIA founder and CEO Huang Renxun will give a keynote speech on June 2nd local time before the Computex conference, during which he will also reveal more new technologies related to AI data centers.