The UALink Consortium, a collaboration of major IT companies such as AMD, Intel, Broadcom, Google, Microsoft, Meta, Cisco, HPE, has been developing the UALink interconnect standard to compete with NVIDIA’s NVLink.
Nearly a year ago, the UALink group unveiled its first spec, the UALink 200G 1.0 Specification. This spec supports a maximum data transfer rate of 200 Gbps per lane, with each port supporting up to 4 lanes (equivalent to 800 Gbps per port). It can connect to GPUs or other accelerators, with a maximum of 1,024 devices.
According to The Next Platform, the UALink 1.0 utilizes the same physical layer cables and ports as Ethernet (IEEE P802.3dj), but has developed its own data link layer and transportation layer based on certain aspects of AMD xGMI and Infinity Fabric.
Following the completion of the UALink 1.0 spec, the next step is to wait for chip manufacturers like Astera Labs, Broadcom, Marvell, and Microchip to support UALink. As these devices become more widespread, we are likely to see an increased adoption of UALink in data centers. However, it’s clear that UALink products will have to compete with the long-standing market leader, NVLink.
Source: UALink, The Next Platform
TL;DR: The UALink Consortium is developing a high-speed interconnect standard to rival NVIDIA’s NVLink, with the recent release of the UALink 200G 1.0 Specification paving the way for increased data transfer rates and connectivity options in data centers.
Leave a Comment