With the 2025 OCP APAC Summit set to kick off next week in Taipei, industry attention is turning toward the next wave of AI and high-performance computing (HPC) infrastructure.
Among the key players shaping this future are AMD and the UALink Consortium. We spoke with Robert Hormuth, Corporate Vice President at AMD, and Kurtis Bowman, Chair of the UALink Consortium, to better understand how UALink aims to redefine open scale-up architectures and the role AMD is playing in advancing adoption.
Credit: Robert Hormuth
Delivering the UALink vision
Q (to Robert Hormuth): What's the key message AMD hopes the industry takes away from UALink's debut at OCP Summit?
Hormuth: AMD is one of the founding members of UALink and a driving force behind both its technical development and adoption. We want the audience to see UALink as more than a specification—it's a collaborative shift toward open, flexible infrastructure for AI and HPC. AMD's role is to help lead this change, supporting both the standardization process and enabling a vibrant market for UALink-enabled solutions.
Q (to Kurtis Bowman): How do you see UALink transforming infrastructure beyond just the technical details?
Bowman: UALink is the industry's open, scale-up fabric—designed to meet the rapid growth of AI and HPC workloads while ensuring interoperability. Our vision is to let customers combine components from multiple vendors in a unified infrastructure. This not only reduces costs but also promotes innovation and improves performance across the ecosystem.
The UALink Consortium itself doesn't comment on the individual roles of members like AMD, but we value their contributions within the broader effort.
The case for open standards
Q (to Bowman): Why does openness matter when building next-generation AI/HPC systems, and what risks do proprietary models pose?
Bowman: An open standard like UALink 1.0 fosters collaboration across cloud providers, chipmakers, and system vendors. This leads to faster innovation, fewer roadblocks, and ultimately a more adaptable ecosystem. It also helps customers avoid vendor lock-in. They're free to build infrastructure that aligns with their unique performance, cost, and supply chain needs—without being tied to a single vendor's roadmap.
Q (to Hormuth): How does AMD's open-standards philosophy fit into this picture?
Hormuth: AMD supports open innovation because it gives customers real choice. By contributing to initiatives like UALink, we help create an ecosystem where technologies can evolve quickly, interoperability is a given, and the industry isn't held back by closed systems. That ultimately benefits customers through scalability, flexibility, and cost optimization.
Credit: Kurtis-Bowman
From racks to real-world deployment
Q (to Bowman): How does UALink's collaboration with the OCP Foundation support faster, broader deployment in data centers?
Bowman: The OCP Foundation has been instrumental in standardizing designs for racks, servers, and data center infrastructure. By aligning UALink with these frameworks, we make it easier for vendors and end-users to integrate UALink-enabled components—accelerators, switches, compute nodes—into proven rack designs. This alignment will help drive global adoption, especially in hyperscale and enterprise environments.
Looking ahead in interconnect performance
Q (to Bowman): What are the next frontiers for interconnect speeds, and how is UALink preparing for them?
Bowman: We're already seeing 400Gbps and even 800Gbps links by aggregating 100Gbps and 200Gbps lanes. But as we push into 200Gbps-per-lane territory, signal integrity becomes a key technical challenge. That's why UALink is actively working with groups like IEEE, the Ethernet Alliance, and OIF to ensure we're aligned with industry-wide progress.
Lessons in building open ecosystems
Q (to Bowman): What past lessons from industry collaboration are shaping how the UALink Consortium operates?
Bowman: The biggest takeaway is that broad-based collaboration drives better standards. Our aim from the beginning was to design something that reflects the needs of the entire industry. That's why we've grown to include more than 100 member companies.
Interoperability has also been central. We built UALink 1.0 to ensure multiple vendors can create compatible devices from the start. That creates a healthy, competitive environment with more choices for users. And we've structured ourselves to move fast—we ratified version 1.0 quickly and are already planning future iterations. UALink isn't just a spec; it's an evolving platform built to support AI innovation long-term.
What comes next for adopters
Q (to Bowman): What practical steps should companies take if they want to integrate UALink into their AI/HPC setups?
Bowman: Start by engaging with UALink ecosystem vendors offering compatible accelerators and switches. Then, design AI pods to take advantage of UALink's memory-semantic communication, which can unify many accelerators into a single compute resource. And make sure your software stack is ready—UALink is designed for seamless integration with existing AI frameworks.
The balance between performance and cost
Q (to Hormuth): How does AMD strike a balance between pushing performance boundaries and enabling cost-effective adoption?
Hormuth: We focus on more than just peak performance—we also look at power efficiency, total cost of ownership, and flexibility. UALink builds on Ethernet physical layers and uses widely available components, which helps reduce cost and complexity.
On the software side, AMD's ROCm stack has always been open, allowing developers to contribute and scale AI workloads. That openness has led to a rich ecosystem—like support for over a million AI models through our work with partners such as Hugging Face and PyTorch.
And on the hardware side, our Instinct MI300X GPUs offer leadership compute and memory bandwidth for generative AI workloads. But we've also worked hard to deliver strong performance per dollar, so customers can scale without blowing past budget limits. The goal is to make high-performance solutions practical and accessible—whether you're running a hyperscale deployment or building something more tailored.
Article edited by Jack Wu
