Multiple Chinese tech firms recently launched several chips that will power ‘Global Scheduling Ethernet’ or GSE, a networking protocol designed to accommodate the large volumes of data and high transfer speeds that AI and other high-performance workloads require. According to The Register, state-owned China Mobile, the largest mobile carrier in the country, has been working on the technology since 2023, after it published a white paper (PDF) on the “Technical Architecture of the Fully Schedule Ethernet.”
This technology is similar to what the Ultra Ethernet Consortium (UEC) is working on. The group is working together to build technology that allows HPC and AI clusters to communicate effectively and efficiently while using existing hardware.
Current Ethernet technology isn’t built for the massive workloads of AI training and other HPC applications; the large and busy networks you typically find in data centers make it difficult for data to move through, resulting in unacceptable latency levels. That’s why data centers that use Nvidia GPUs usually use technologies like NVLink to allow their systems to communicate directly. However, NVLink is a proprietary system that only works with Nvidia products, limiting its usefulness and making it costly to deploy.
As of October 2024, the UEC had 97 members, and AMD unveiled the first Ultra Ethernet-ready network card last month. Meanwhile, more than 50 organizations, including cloud service providers, chip makers, and educational institutions from inside and outside China, have been involved in developing the GSE.
The chips launched this month are said to make GSE a reality. In contrast, others reported that Chine Mobile had deployed the system in one of its data centers and delivered massive performance improvements for AI training.
The GSE seems to be a direct competitor to what the UEC is working on. However, even though both teams have reportedly released hardware for both technologies, we still haven’t seen a comparison of their abilities. It would be interesting to see which of these technologies works better and is more cost-effective.
