UALink 1.0 Specification Released: A New Era of Open AI Accelerator Connectivity

UALink 1.0: A New Standard in AI Accelerator Connectivity

The Ultra Accelerator Link Consortium has officially released the final UALink 1.0 specification, marking a significant milestone in the development of open standards for AI accelerator connectivity. This new technology aims to challenge Nvidia's proprietary NVLink, offering a competitive alternative for high-performance computing and AI applications 1

1

2

Technical Specifications and Capabilities

UALink 1.0 boasts impressive technical specifications:

• Maximum bidirectional data rate of 200 GT/s per lane
• Support for up to 1,024 accelerators in a single system
• Configurable links (x1, x2, or x4) with up to 800 GT/s in both directions
• Optimized for <4 meter cable lengths with <1 µs round-trip latency
• Four hardware-optimized layers: physical, data link, transaction, and protocol 1

  1

The new standard leverages standard Ethernet components while introducing modifications for reduced latency. It also implements compressed addressing, achieving up to 95% protocol efficiency under real workloads 1

1

Security and Management Features

UALink 1.0 incorporates advanced security and management capabilities:

• UALinkSec for hardware-level encryption and authentication
• Support for Confidential Computing through tenant-controlled Trusted Execution Environments
• Virtual Pod partitioning for isolated accelerator groups
• Full manageability through REST APIs, telemetry, and workload control 1

  1

Related Stories

UALink Consortium Incorporates to Challenge Nvidia's Dominance in AI Interconnect Technology

Broadcom's Tomahawk Ultra: Revolutionizing Ethernet for AI and HPC

Cornelis Networks raises $205M, launches Active Compute Fabric to challenge Nvidia's AI dominance

Tech giants form optical interconnect alliance to solve AI infrastructure's data bottleneck

Industry Impact and Competition with Nvidia

The release of UALink 1.0 is set to reshape the AI accelerator market:

• Enables companies like AMD, Broadcom, Google, and Intel to build rack-scale AI solutions
• Offers an open alternative to Nvidia's proprietary NVLink technology
• Potential for lower costs due to the use of industry-standard technologies 1

  1

  2

  2

Nvidia currently dominates the market with solutions like Blackwell NVL72, which can connect up to 72 GPUs in a single rack. The company plans to scale up to 576 GPUs in a single rack by 2027 with its Vera Rubin Ultra platform 1

1

Consortium Members and Future Developments

The Ultra Accelerator Link Consortium includes major players such as AMD, AWS, Broadcom, Cisco, Google, HPE, Intel, Meta, and Microsoft. The specification draws heavily on AMD's existing Infinity Fabric product while incorporating technologies from other members 2

2

Kurtis Bowman, Chair and Director of UAC, expects compliant hardware to be available in about 18 months. Major vendors like HPE, Dell, and Lenovo are anticipated to adopt the spec for their AI solutions 2

2

Work on a second specification is already underway, aiming to take advantage of 400G Ethernet variants as they become mainstream 2

2

As the AI hardware landscape continues to evolve, UALink 1.0 represents a significant step towards open standards and increased competition in the high-performance computing sector.