UALink consortium set to compete with Nvidia's NVLink: AMD- and Intel-led group opens doors to contributing members.
The Consortium Ultra Accelerator Link (ULink) has been officially incorporated, meaning it is now a legal entity. 🎉 This consortium seeks to set a new standard for high-speed, low-latency communication between servers in AI data centers. It has prominent members on the board of directors, including giants such as AMD, Intel, Meta, Hewlett-Packard Enterprise, Amazon Web Services (AWS), Astera Labs, Cisco, Google and Microsoft. They are also looking for new contributing members to join this initiative. 🤝
UALink aims to become the open industry standard for expanded connections of multiple AI accelerators, competing directly with NVLink of Nvidia. NVLink is Nvidia's solution for communication between GPUs in servers, using Infiniband, other Nvidia technology for higher scaling. However, Infiniband now faces competition from the emerging Ultra Ethernet, another major consortium of tech giants looking to counter Nvidia's dominance with an open standard. 🚀
Willie Nelson, President of the Consortium UALink, is opening the doors for new companies and groups to join this exciting project. 💼 «Interested companies are encouraged to join as contributing members to support our mission: to establish an open and high-performance accelerator network performance para cargas de trabajo de IA», afirmó Nelson. La especificación 1.0 de UALink estará disponible para los miembros este año. 🔍
The standard will allow a connection of up to 200 Gbps per lane for up to 1,024 accelerators within an AI module. In a typical scenario with AI-style servers, Nvidia HGX, que contienen 8 aceleradores de IA, UALink podría conectar hasta 128 de estas máquinas en un módulo. Sin embargo, es probable que UALink se utilice con mayor frecuencia a menor escala, facilitando la comunicación entre módulos de alrededor de ocho servidores, con ampliación adicional a través de Ultra Ethernet. 🔗
Consortium members will have access to the specification this year, and a general review will be opened in the first quarter of 2025. The release of the UALink standard will coincide with the debut of version 1 of Ultra Ethernet. 🌐 AMD recently announced the first card 400 GbE ready for industry Ultra Ethernet. UALink and Ultra Ethernet are backed by industry giants looking to dethrone Nvidia and it is almost sure which will establish themselves as open standards in the field of AI data centers, thanks to their broad levels of support. 📈
The work that UALink companies are doing to create an open, high-performance accelerator fabric performance and scalable is critical to the future of AI," he said. Forrest Norrod, executive vice president and general manager of AMD's data center solutions group. 🧠 It is expected that Samsung be one of the first contributing members of the Consortium, having announced its intentions to join in June. 📅
The Ultra Consortium is likely Ethernet I also experienced many collaborations, with industry giants such as Baidu, Dell, Huawei, IBM, Nokia, Lenovo, Supermicro and Tencent, which have joined as contributors to the EU in recent months. 🌟 Nvidia is expected to remain outside the Consortium, as its technologies NVLink and Infiniband are proprietary and have already seen widespread use thanks to the company's dominance in the AI data center market. 🔒
