Quote:Ultra Accelerator Link (UALink) standard for data-centers

The largest companies in the industry are partnering to develop an open standard. 



NVIDIA NVLink is a technology that allows high-speed communication between GPUs and CPUs, surpassing existing connectivity standards which are simply not up to the task for modern high-performance accelerators. Initially showcased in 2014, NVLink  found application even in gaming GPUs before NVIDIA discontinued multi-GPU support.

For data centers housing large clusters of accelerators, efficient communication is very important. Servers require efficient networking and intercommunication methods, particularly when multiple GPUs are employed within a single server. The latest iteration, NVLink 5.0, raises bandwidth compared to 4.0, doubling it from 100 Gbit/s to 200 Gbit/s. It is to be deployed with Blackwell and Grace CPUs.

The new standard is a collaborative effort involving several prominent companies, including AMD, Broadcom, Cisco, Google, Hewlett Packard Enterprise (HPE), Intel, Meta, and Microsoft. Together, they will focus on developing this new standard tailored for scaling up AI applications in data centers.
 

Quote:Called the Ultra Accelerator Link (UALink), this initial group will define and establish an open industry standard that will enable AI accelerators to communicate more effectively. By creating an interconnect based upon open standards, UALink will enable system OEMs, IT professionals and system integrators to create a pathway for easier integration, greater flexibility and scalability of their AI-connected data centers.

As per the press release, the initial specification (1.0) aims to enable communication between up to 1024 accelerators within a single computing pod and the attached memory, including GPUs. However, there are no details about the maximum bandwidth.

The UALink Consortium is set to be established in Q3 2024, with the 1.0 specifications slated for availability in the same quarter.
...