AMD is at the center of a global collaboration to innovate next-generation artificial intelligence (AI) infrastructure.

AMD announced on the 13th that it has officially launched the 'OCI MSA (Optical Compute Interconnect Multi-Source Agreement)' consortium to jointly develop optical-based scale-up interconnect standards with Broadcom, Meta, Microsoft, NVIDIA, and OpenAI.

This council aims to define open specifications for high-performance, low-power connectivity technologies required for AI data centers and hyperscale computing environments. As a founding member, AMD will play a key role in establishing the technology roadmap and expanding the ecosystem.

With the recent rapid proliferation of large-scale language models and generative AI, GPU-to-GPU interconnect structures are facing limitations in terms of power consumption and transmission distance. To address these challenges, AMD is proposing a scale-up architecture shift centered on optical interconnects. Fiber-optic connections provide stable bandwidth over longer distances compared to conventional copper, while also significantly improving energy efficiency.

The OCI specification combines NRZ modulation with wavelength division multiplexing (WDM) technology to achieve high bandwidth density. In particular, by adopting a structure that directly integrates optical technology with computing and network semiconductors, it simultaneously secures the scalability and design flexibility of AI systems.

The AMD-led OCI MSA presented a unified roadmap supporting a variety of implementations, including pluggable modules, onboard optics, and co-packaged optics (CPO). Initially, the goal is to achieve hundreds of gigabits per fiber, with long-term expansion to terabit bandwidth. This will enable the construction of ultra-large AI clusters that combine more GPUs into a single scale-up domain.

The industry sees AMD's strategy as one aimed at leading the way beyond CPUs and GPUs to connectivity standards across the entire AI infrastructure. If the open optical interconnect ecosystem takes hold, significant changes are expected in the competitive landscape for performance and efficiency in AI data centers.