AI workloads are rapidly growing in scale and complexity, presenting major challenges for traditional hardware. Ultra Accelerator Link™ (UALink™) is the industry’s only open, accelerator-to-accelerator interconnect designed specifically to deliver the scale, performance, and efficiency demands of modern AI systems.
Advantages of UALink Technology
UALink delivers high-bandwidth and low-latency for AI workloads, helping to eliminate the communication bottlenecks common in legacy or proprietary interconnects. Its open, high-performance design enables real-time data exchange and more efficient workload coordination across accelerators.
Built for scalability, UALink supports up to 1,024 accelerators within an AI computing pod, empowering organizations to expand large-scale AI deployments without compromising performance or flexibility. UALink’s streamlined architecture also reduces power consumption, simplifies switch designs, and minimizes the die area of the link stack, resulting in lower capital and operational costs while improving overall efficiency.
Together, these benefits establish UALink as a future-ready foundation to enable transformative AI solutions.
What’s Next for UALink?
The UALink Consortium Technical Working Groups are continuing their work to meet the growing demands of compute and memory for AI applications.
One of the key features planned for UALink is support for in-network collectives (INC), which play an important role in improving the performance and scalability of AI systems. INC allows the interconnect fabric itself to handle collective communication operations. By offloading these data-exchange tasks into the network hardware, UALink can significantly lower software overhead, reduce latency and improve overall throughput. This approach also minimizes redundant data movement across the network, making communication between accelerators more efficient and scalable as system sizes grow into hundreds or thousands of devices
UALink management specification is also underway, which will ensure that UALink switches and accelerators provide a standard interface for diagnostics, status reporting, firmware updates, error counters, topology discovery, and security policies. By having a consistent, well-defined management plane, system software and orchestration tools can reliably monitor and operate large AI pods at scale, automate detection and recovery of link failures, enforce quality of service, and integrate data-center management frameworks.
Finally, the UALink Chiplet Specification will enable accelerators, memory, I/O and logic blocks to be packaged in smaller, interoperable building blocks that communicate via high-bandwidth, low-latency links, allowing system designers to assemble custom configurations, mix dies produced on different process nodes, and scale beyond monolithic reticle limits. A robust chiplet specification also ensures that chiplets from multiple sources can interoperate smoothly, making modular, flexible and high-performance AI systems feasible at scale.
With these advancements, UALink is not only addressing the current challenges of AI infrastructure, but also laying the groundwork for the next-generation of scalable, high-performance and flexible AI systems. By incorporating INC, management and chiplet specifications with the high-bandwidth and low-latency, UALink is empowering organizations to build large, efficient and future-ready AI deployments.
UALink Solutions in Development
UALink members are actively developing a new generation of accelerator and switch designs built on the UALink 1.0 Specifications. UALink member companies Alphawave Semi, Cadence and Synopsys have already announced IP support for UALink, demonstrating their commitment to accelerating high-performance AI and compute workloads. With more members joining and additional solutions in the pipeline, UALink is establishing itself as a foundational interconnect standard for next-generation AI and HPC systems.
As the ecosystem continues to grow and evolve, the UALink Consortium will continue to develop technical specifications that empower innovation, accelerate AI workloads and support our members to stay ahead in an increasingly demanding and fast-moving AI landscape.