Why UCIe is Key to Connectivity for Next-Gen AI Chiplets

By Letizia Giuliano, VP of IP Products

February 6, 2025

Deploying AI at scale presents enormous challenges, with workloads demanding massive compute power and high-speed communication bandwidth.

Large AI clusters require significant networking infrastructure to handle the data flow between the processors, memory, and storage; without this, the performance of even the most advanced models can be bottlenecked. Data from Meta suggests that approximately 40% of the time that data resides in a data center is wasted, sitting in networking.

In short, connectivity is choking the network, and AI requires dedicated hardware with the maximum possible communication bandwidth.

Deploying AI at scale presents enormous challenges, with workloads demanding massive compute power and high-speed communication bandwidth.

Large AI clusters require significant networking infrastructure to handle the data flow between the processors, memory, and storage; without this, the performance of even the most advanced models can be bottlenecked. Data from Meta suggests that approximately 40% of the time that data resides in a data center is wasted, sitting in networking.

In short, connectivity is choking the network, and AI requires dedicated hardware with the maximum possible communication bandwidth.

The large training workloads of AI create high-bandwidth traffic on the back-end network, and this traffic generally flows in regular patterns and does not require the packet-by-packet handling needed in the front-end network. When things are working properly, they operate with very high levels of activity.

Low latency is critical, as we must have fast access to other resources, and this is enabled by a flat hierarchy. To prevent (expensive) compute being left underutilized, switching also must be non-blocking—it should be noted that the performance of AI networks can be bottlenecked by even one link that has frequent packet losses. Robustness and reliability of the networks are also critical, with the design of the back-end ML network taking this into consideration.

To read the full article, click here