“What this architecture does is remove any of the bottlenecks and choke points that we would have had otherwise, and it gives us a much more scalable and modular way to keep extending the architecture going forward,” Kumar explains. “This mesh puts us on a different foundation, and it is a significant change to the architecture and the overall organization of the chip.”
This begs the question as to why Intel chose to do rings, and then interconnected rings, in the first place. With the Nehalem through Sandy Bridge chips, a ring worked just fine and a mesh was not needed, but somewhere around Ivy Bridge Intel did in fact debate going to a mesh, and we think the company decided to test it out on Knights Landing first to work out the architectural issues, and that is what HPC is precisely for. To test out ideas before they get to the wider adoption in the enterprise.