Data Center Leaders Launch Ultra Accelerator Link Group to Enhance AI Connectivity

A coalition of data center technology leaders has established the Ultra Accelerator Link (UALink) Promoter Group to pioneer new methods for scaling AI systems within data centers.

Key Industry Players Unite

Leading companies such as Advanced Micro Devices (AMD), Broadcom, Cisco, Google, Hewlett Packard Enterprise (HPE), Intel, Meta, and Microsoft have joined forces to develop an industry standard that enhances high-speed, low-latency communication for AI systems within data centers.

Introducing UALink

The UALink initiative aims to create an open industry standard that allows AI accelerators to communicate more effectively, facilitating better integration, flexibility, and scalability for AI-connected data centers. By establishing an interconnect based on open standards, the UALink will benefit system original equipment manufacturers (OEMs), IT professionals, and system integrators alike.

Forrest Norrod, General Manager of the Data Center Solutions Group at AMD, emphasized the importance of this collaboration: "The UALink efforts to build an open, high-performance accelerator fabric are crucial for the future of AI. Our combined expertise will advance open standards and strengthen the AI ecosystem."

Notable Exclusions

Although the Promoter Group boasts extensive experience in large-scale AI and high-performance computing, it notably does not include AI chip leader Nvidia.

Commitment to Open Standards

Jas Tremblay, VP of the Data Center Solutions Group at Broadcom, stated, "As a founding member of the UALink Consortium, we aim to enhance large-scale AI technology adoption in data centers. Supporting an open ecosystem is vital for enabling scalable networks equipped with various high-speed, low-latency solutions."

Scaling AI Workloads Effectively

As AI compute demands escalate, establishing a robust, low-latency network that can easily integrate additional computing resources is imperative. The UALink group aims to standardize specifications to create a high-performance environment for AI workloads, maximizing performance outcomes.

The group plans to develop a specification that will optimize the interface for AI and Machine Learning, High-Performance Computing (HPC), and Cloud applications within future AI data centers. Their goal is to define a high-speed, low-latency interconnect for seamless communication between accelerators and switches in AI computing pods.

The forthcoming 1.0 specification aims to enable connections for up to 1,024 accelerators in a single AI computing pod, facilitating direct loads and stores between accelerator memory, such as GPUs.

Looking Ahead

The UALink Promoter Group will soon establish the UALink Consortium, anticipated to become official in Q3 of 2024, with the 1.0 specification to follow in the same timeframe for consortium members.

About UALink

The Ultra Accelerator Link (UALink) is a cutting-edge accelerator interconnect technology designed to enhance the performance of next-generation AI and Machine Learning (ML) clusters. AMD, Broadcom, Cisco, Google, HPE, Intel, Meta, and Microsoft are committed to forming an open standard body to develop technical specifications that support breakthrough performance while fostering an open ecosystem for datacenter accelerators.

Martin Lund, EVP of the Common Hardware Group at Cisco, remarked, "As AI workloads grow, ultra-high performance interconnects will be critical. Our shared goal is to create the UALink, a scalable, open solution that addresses the challenges of building AI supercomputers."