Optical Scale Up Fabrics Are Limited by Manufacturing, Not Architecture (www.nextplatform.com)

A consortium of tech giants, including AMD and Nvidia, has established the Optical Compute Interconnect Multi-Source Agreement (OCI MSA) to standardize networking for AI infrastructure through co-packaged optics (CPO). The OCI GEN1 specification introduces a foundational architecture that supports four wavelengths at 50 Gb/sec per channel, allowing a total of 200 Gb/sec per direction per fiber, with plans to scale up to 1.6 Tb/sec. This move is significant as it resolves the architectural discussions in AI networking but leaves critical questions about enhancing bandwidth and the manufacturing technologies required to produce sufficient laser arrays for higher wavelength counts. The transition towards heterogeneous integration in photonics, akin to the advancements seen with CMOS in electronics, is pivotal for the future of AI scalability. This manufacturing evolution enables the consolidation of various optical components on a single wafer, drastically reducing costs and improving reliability. As OCI MSA aims to overcome the manufacturing bottleneck currently hindering the broader application of dense wavelength-division multiplexing (DWDM), it paves the way for future upgrades in bandwidth without extensive redesign. The ongoing development of robust manufacturing processes promises to support the growing demands of AI datacenters, ultimately facilitating the integration of thousands of GPUs into cohesive computing systems capable of handling more complex AI tasks.