AI infrastructure scaling is fundamentally a connectivity challenge, as current compute and memory advancements require massive, disaggregated systems that exceed the physical limits of copper cabling. This "copper wall" necessitates a transition toward optical interconnects, including silicon photonics and co-packaged optics, to maintain bandwidth and power efficiency. Marvell’s strategy focuses on providing a comprehensive connectivity portfolio that spans from die-to-die interfaces to long-reach data center links. By partnering with industry leaders like NVIDIA and ASE, the infrastructure is evolving toward a disaggregated model where compute and memory resources are pooled and dynamically composed. In this future, distance-independent optical connectivity will allow architects to design systems based on workload requirements rather than the physical constraints of traditional interconnects, enabling millions of processors to operate as a single, unified machine.
Sign in to continue reading, translating and more.
Open full episode in Podwise