Scale-up networks increase computing power within a single rack or adjacent racks by adding processors, accelerators, memory, and storage. In Al infrastructure, scale-up typically uses very high-speed, low latency connectivity so a cluster of many GPUs or other accelerators can operate as a single, tightly coordinated compute resource.

Scale-out networks increase compute capacity by connecting multiple clusters across a network, allowing workloads to be distributed among them. Scale-out is commonly used in cloud and Al data centers to expand performance and capacity by adding more compute nodes.

Scale-across networks connect multiple high-performance systems or Al clusters so they can work together across multiple data centers. Scale-across uses high-bandwidth optical links with optical amplification as needed, to move data efficiently between data centers located in the same metro area, nearby regions or across the country.