The Chinese have created Ethernet for AI and HPC clusters

Last week, Chinese tech giants announced a chip to support Global Scheduling Ethernet, a networking protocol that should enable AI applications and other demanding workloads, The Register reported. According to the resource, China Mobile is the driving force behind this technology, since back in 2023 it published a description of the technical structure of GSE Ethernet.

The goal of the Chinese project coincides with the goal set by the Ultra Ethernet Consortium (UEC) to optimize Ethernet for AI and HPC applications. UEC members include Intel, AMD, HPE, Arista, Broadcom, Cisco, Meta✴ and Microsoft. The Ethernet protocol was not designed with today’s workloads in mind and makes it difficult to route traffic across very large, busy networks, resulting in high latency.

Image Source: China Mobile

China Mobile’s White Paper points out similar problems that can be solved using techniques such as fixed-size packet containers and a “dynamic global scheduling queue” that is not tied to physical ports, but takes into account the port state of the target device before establishing an optimal connection with using methods such as multi-path spraying. By the way, UEC also considers this method promising.

Last week, Chinese media reported that more than 50 cloud service providers, equipment manufacturers, chip makers and universities inside and outside China were involved in the development of the chip that makes GSE Ethernet a reality. Reports suggest that GSE Ethernet has already been deployed across a thousand-server cluster at China Mobile’s data center, where it appears to have provided significant improvements in network performance while training a large language model.

If China has decided to build and use its own version of Ethernet for some applications—and its big tech companies are willing to adopt it—that means UEC members will have a harder time targeting the Chinese market (and countries dominated by Chinese vendors), The Register writes.