Challenge NVIDIA! Tech giants form alliance to establish AI accelerator interconnect standards

Intel, Google, Microsoft, Meta, and other tech giants announced on Thursday the establishment of a new industry organization called the "Ultra Accelerator Link (UALink) Promotion Group," aiming to set industry standards and lead the development of connection components for AI accelerator chips in data centers, challenging NVIDIA's dominant position in AI accelerators.

Jointly Establishing AI Accelerator Connection Standards to Challenge NVIDIA

In addition to the above companies, members of this organization also include AMD, Hewlett Packard Enterprise, Broadcom, Cisco, among others, with Arm not participating. The UALink Promotion Group is proposing a new industry standard to connect AI accelerator chips in the increasing number of servers.

"The industry needs an open standard that can be rapidly advanced, allowing multiple companies to add value to the entire ecosystem," said Forrest Norrod, General Manager of Data Center Solutions at AMD. "The industry needs a standard that allows for rapid innovation without being limited by any single company."

Broadly speaking, AI accelerators refer to GPUs and other custom-designed solutions used to accelerate the training, fine-tuning, and operation of AI models. The proposed standard version by UALink, UALink 1.0, will connect up to 1024 GPU AI accelerators, forming a computing "pod" in a server rack or racks. According to the UALink Promotion Group, based on "open standards" including AMD's Infinity Fabric, UALink 1.0 will allow for direct loading and storage between the memory attached to AI accelerators, improving overall speed and reducing data transfer latency compared to existing interconnect specifications.

The organization stated that an alliance called UALink will be established in the third quarter to oversee the development of UALink specifications. UALink 1.0 will be made available to companies joining the alliance during the same period, while an iterative specification UALink 1.1 with higher bandwidth is expected to be launched in the fourth quarter of 2024. Norrod mentioned that the first batch of UALink products will be launched "in the coming years."

NVIDIA: Not Interested, Not Necessary

Analysts believe that the establishment of this organization, notably without the largest AI accelerator manufacturer NVIDIA, is aimed at challenging NVIDIA's dominant position in the AI field. Currently, NVIDIA's market share in AI accelerators is estimated to be between 80% to 95%.

NVIDIA has not commented on the matter yet, but it is evident that the company is not keen on supporting UALink. Analysts suggest that NVIDIA has already provided its proprietary interconnect technology for GPU links in data center servers, so they may not be willing to support specifications based on competitor technologies Moreover, NVIDIA's position and influence in the field of AI are very strong. In the most recent fiscal quarter, NVIDIA's data center sales (including AI chip sales) grew by over 400% year-on-year. If it continues on its current trajectory, NVIDIA may surpass Apple this year to become the world's second-largest company by market capitalization.

Therefore, NVIDIA does not want to participate, nor does it need to participate in UALink.

Another major company that has not yet joined is Amazon. Analysts believe that Amazon may be in a "wait-and-see" mode, as the company is also continuously developing its internal accelerator hardware efforts. AWS may also consider that there is not much strategic significance in confronting NVIDIA, which supplies a large number of GPUs to its customers.

Microsoft, Meta, Google are the most motivated

Analysis suggests that, apart from AMD and Intel, the biggest beneficiaries of UALink seem to be Microsoft, Meta, and Google, as these three companies have collectively spent billions of dollars purchasing NVIDIA GPUs to support their cloud services and train their growing AI models. They all hope to break free from NVIDIA's dominant position as an AI hardware ecosystem supplier.

In a recent report, Gartner estimated that the value of AI accelerators used in servers this year will reach $21 billion, increasing to $33 billion by 2028. At the same time, Gartner predicts that by 2025, AI chip revenue will reach $33.4 billion.

Google has custom chips for training and running AI models - TPU and Axion. Amazon has several AI chip families. Microsoft joined the competition last year, launching Maia and Cobalt. Meta is also continuously improving its accelerator product line.

Additionally, it has been reported that Microsoft and its close partner OpenAI plan to invest $100 billion in supercomputers for training AI models equipped with future versions of Cobalt and Maia chips. These chips will need some way to connect - perhaps this is the purpose of UALink