A consortium of major IT companies led by prominent CPU manufacturers such as AMD, Intel, as well as others including Broadcom, Cisco, Google, HPE, Meta, and Microsoft have announced the formation of the Ultra Accelerator Link (UALink) group to establish standards for connecting GPUs in data centers.
UALink aims to rival NVLink, a data transfer system between GPUs from a leading market player, by introducing a new standard for data connectivity. While the specifications for UALink have yet to be released, the consortium plans to register the UALink Consortium in Q3 2024 and launch the UALink 1.0 specs in the same quarter, supporting connectivity for up to 1,024 GPUs or other processing accelerators.
HPCwire provides information on the current methods of GPU connectivity, which include using PCI Bus for general server standards supporting 4-8 GPUs, and leveraging Infiniband for inter-server GPU connections. Additionally, the Ultra Ethernet Consortium (UEC) has emerged as a competitor with Intel adopting their technology for AI processing accelerators like Gaudi 2.
The GPU to GPU connectivity market is dominated by NVLink, with AMD’s Infinity Fabric serving as a formidable rival used in their MI300 AI accelerator. However, the lack of a unified standards group has led to the formation of the UALink consortium to address this need.
Source: Businesswire, HPC Wire
TLDR: Major IT companies form UALink consortium to establish GPU connectivity standards, challenging the dominance of NVLink and other connectivity technologies in data centers.
Leave a Comment