Today, as the battle over AI compute power rapidly spreads from single-chip performance to cluster interconnects, the high-speed interconnect arena is undergoing rare and violent upheaval.
Two startup forces that originated from Intel’s() former networking business are simultaneously charging into the most valuable Scale-up interconnect track of the AI era—and pointing their spears directly at NVIDIA’s heartland. On the other side, an “anti-NVLink alliance” assembled by 115 industry giants is accelerating its mobilization in an attempt to break a single-vendor monopoly. NVIDIA(), meanwhile, is going with the flow and playing its hidden card, “NVLink Fusion,” trying to recreate CUDA’s path to deification at the hardware layer.
A hardware showdown over who will dominate the next generation of AI data centers has already fully begun.
Two former Intel teams charge onto the same battlefield
First, consider Cornelis. Its ties to Intel run deep. In 2020, Intel spun out its Omni-Path Architecture (OPA) high-performance interconnect business; the core team founded Cornelis Networks. Cornelis CEO Lisa Spelman herself spent about 20 years at Intel. To this day, Cornelis still explicitly describes itself as “spun out from Intel’s high-performance interconnect business,” and its technical roadmap directly inherits Omni-Path. Omni-Path was originally Intel’s important weapon for challenging Mellanox InfiniBand.
After Intel adjusted its strategy, the business was ultimately carved out. In the years that followed, Cornelis continued developing its high-speed interconnect products around HPC and AI clusters. The already-in-production CN5000 still mainly addresses Scale-out—that is, large-scale connections among servers and compute nodes. Its switches offer 48 ports at 400 Gb/s; Director-class systems can provide up to 576 ports at 400 Gb/s.
Read more
