Frequently Asked Questions
Why was the UALink Consortium formed?
As AI models accelerate, data centers are requiring increased amounts of compute and memory to efficiently execute training and inference. The UALink Consortium was formed to develop technical specifications that facilitate direct load, store, and atomic operations between AI Accelerators. The Consortium is establishing a new industry standard that facilitates an optimized scale-up ecosystem and enables advanced models across multiple AI accelerators.
What is the primary benefit of UALink technology?
UALink offers several benefits to the industry and ecosystem. Primarily, as an open industry standard interconnect for communication between accelerators and switches in AI computing pods, it offers low-latency, high-bandwidth, and high-efficiency attributed to lower power, area, and high bandwidth utilization.
What features are included in the UALink 200G 1.0 specification?
The UALink 200G 1.0 Specification defines a low-latency, high-bandwidth interconnect for communication between accelerators and switches in AI computing pods. UALink 1.0 enables a 200G per lane scale-up connection for up to 1,024 accelerators within an AI computing pod, delivering the fastest open standard interconnect for next-generation AI cluster performance.
Are there any plans to expand beyond 1,024 accelerators?
UALink currently supports a maximum of 1,024 accelerators. However, the UALink Technical Task Force is exploring options to increase the scale-up domain size. Implementation of this addition to the specification will be based on industry demand.
Is UALink planning to address 128G?
UALink is exploring options and implementations for future specification versions, including 128G capabilities.
What are the message size capabilities with UALink?
UALink uses direct load/store access memory within the scale-up domain, and, for a number of reasons, many accelerator cache line sizes are limited to 256B, including UALink. UALink also supports smaller payloads of 64B, 128B and 192B.
When are UALink-based solutions expected to enter the market?
Products utilizing an open standard typically enter the market within 1 – 2 years following the initial specification release. For more specific information on product availability, please contact our member companies.
Can non-members of the Consortium utilize the specification for their technology developments?
Yes. As the UALink 1.0 specification is an open standard, non-members of the UALink Consortium are welcome to download the specification and develop UALink technology-based solutions. However, member companies enjoy additional benefits, including early access to the specification, influence over the technology’s direction, input in various workgroups, and more. Members also have access to testing standards, compliance, and interoperability testing to ensure their UALink device is spec-compliant and interoperable with other devices. UALink membership provides IP licensing needed to manufacture and market products based on UALink, so we highly recommend that companies planning to design and offer UALink-based products join the Consortium
How many racks can be connected with UALink?
UALink does not limit the number of racks that can be connected.
What is the maximum distance allowed between accelerators using UALink?
Hyperscalers and other end users using UALink are able to determine the appropriate configuration that works best for their individual use cases.
Is the UALink connection over optical?
The UALink 200G 1.0 Specification is based on copper and has a maximum distance of a few meters without the addition of a repeater. Future specification iterations plans include exploring the inclusion of optics.
How is UALink addressing switching operations?
UALink is focused on delivering an efficient switching fabric for direct load/store operations. When transporting messages across the SerDes link, UALink allows a single FLIT to have messages targeting multiple destinations and include requests and completions in multiple VCs within the same FLIT. This is critical for building area-efficient and low-latency designs. Switches need to unpack the FLIT and route individual UPLI packets to the target port, where it is repacked before being sent on the link. Target ports can pack messages from many sources.
How is UALink addressing collective operations?
The Consortium plans to address collective operations in future specification iterations.
When will the next UALink specification iteration be released?
The UALink Consortium intends to update the UALink Specification as the needs of the AI industry evolve. Discussions are currently underway within the Consortium’s Technical Workgroups regarding next-generation features, but a timeline for release has not been defined.
Can you share the difference between the UALink and the Ultra Ethernet (UEC) Consortia?
UALink and UEC are complementary organizations, with UALink technology addressing efficient scale-up architectures and UEC primarily addressing scale-out architectures. We are engaged to define potential collaborative efforts. Please contact UEC representatives with questions regarding UEC.
Do you have a question for the UALink Consortium? Members of the press should contact press@ualinkconsortium.org. All other inquiries can be directed to admin@ualinkconsortium.org.