AI Architecture & Systems

Physical Deployment Hierarchy and Scale-Up/Scale-Out Domains

SourcesSurvey Introduction, Fig. 1, §4 and §6.3; NVIDIA, AMD, TPU and Neuron fabric mappings. Physical and communication-domain boundaries are distinct.

4.15.1Accelerator → Server → Rack → Pod → Datacenter#

4.15.2Organization of Compute Tray, Switch Tray, Host, and NIC#

4.15.3The Reach of a Scale-Up Fabric Across Server/Rack/Pod#

4.15.4Connecting the Scale-Out Network to the Scale-Up Domain#

4.15.5Scale-Up Interfaces: NVLink/NVSwitch, Infinity Fabric, and UALink/UALoE#

4.15.6NPU Fabrics: TPU ICI, NeuronLink/NeuronSwitch, and Huawei UnifiedBus#

4.15.7Distinguishing Physical Rack Boundaries, Network Domain Boundaries, and Failure Domains#

4.15.8Capacity, Distance, and Infrastructure Cost of Extending a Scale-Up Domain#