Scale-up network
The fastest network tier, joining a small group of GPUs, such as a server or rack, so tightly that they work as one.
Also: scale-up
AI systems have two network tiers. The scale-up tier links a group of accelerators, eight in a classic server or 72 in an NVL72 rack, with very high bandwidth and low delay, using NVLink, AMD's Infinity Fabric or Google's inter-chip links. This tier carries the heaviest traffic, such as the pieces of a model split by tensor parallelism. Industry groups formed in 2024 and 2025 to create open alternatives, such as UALink. The larger the scale-up domain, the bigger the model that can run without leaving it.