Intel, Google, Microsoft, Meta AMD, Hewlett Packard Enterprise, Broadcom and Cisco have established the Ultra Accelerator Link (UALink) Promoter Group to develop open source components that connect AI accelerator chips in data centers.
The group wants to standardize a standard used by the industry to connect AI accelerators in the form of CPUs or GPUs needed in data centers to train AI models. These standards are needed to prevent innovation from being dominated by one company alone.
With the involvement of major technology companies to use UALink, each member can contribute their expertise for mutual benefit. Therefore the UALink Consortium will be established in Q3 2024 which will develop future UALink specifications.
With the formation of this group, the UALink 1.0 standard was announced and it can connect 1024 AI accelerators in a computing pod in a data center. UALink 1.0 uses open standards such as AMD Infinity Fabric. The first UALink 1.0 products will be launched in a few years.
NVIDIA does not join the UALink Promoter Group as they currently use their own AI driver extension. At this point NVIDIA holds 90% of the data center GPU market which helped them post a 262% increase in revenue last quarter compared to the same quarter last year.
UALink Promoter Group is another industry effort to undercut NVIDIA's power in the AI world. Last March Intel, Google, Arm, Qualcomm, Samsung and several other technology companies established The Unified Acceleration Foundation (UXL). Through UXL, AI development will no longer be tied to the use of specific programming languages, code bases and other tools that bind them to specific architectures, such as Nvidia's CUDA platform.
