The artificial intelligence industry is undergoing a connectivity revolution. As the number of GPUs in AI clusters has grown from dozens to hundreds and thousands, the physical and architectural limits of traditional networks are being pushed to their extremes. Today, the industry is deploying clusters with 100,000 or even more XPUs. At this scale, the question is no longer whether we can connect more computing resources, but whether the network designs we have long relied on can keep pace.
For years, copper cables have been the default choice for connecting accelerators within a rack—and for good reason. They are inexpensive, energy-efficient, and reliable. Broadcom() supports copper cable transmission distances of up to 4 meters, twice the 2 meters commonly adopted by other vendors in the industry. However, copper has physical limits, and rack designs have had to adapt accordingly. In a typical hybrid rack, compute trays are placed at the top and bottom with the switch in the middle, because that is the practical limit of cable reach.
We have pushed copper further than many expected. Today’s in-rack scale-up systems support total bandwidth of up to 1.04 Pbps. Higher-density backplanes and more advanced switches can double that to 2.08 Pbps, enabling more XPUs and higher performance per rack. But once the scale-up domain needs to span more than a single rack, alternative solutions are required.
The shift to optical connectivity changes everything. Optical links extend far beyond the 2-to-4-meter reach of copper, freeing infrastructure operators from fixed rack configurations. Disaggregated rack architectures fully separate compute racks from switch racks, allowing operators to lay out infrastructure flexibly according to their operational needs.
