LambdaGain Blog
Understanding how AI connectivity changes from inside the compute system to the data hall and across the AI campus
As AI infrastructure grows, Scale Up, Scale Out and Scale Across describe three different ways compute resources are connected. Each presents different requirements for bandwidth, latency, power, distance and connectivity.
A simple way to understand them is to follow connectivity outward from the GPU:
GPU → System → Rack → Data Hall → Building → AI Campus
Scale Up: Connecting GPUs Within the Compute Domain
Scale Up connects GPUs and accelerators into a tightly coupled compute system.
It begins inside the AI server, where GPUs communicate over extremely high-bandwidth electrical interfaces carried across PCBs, backplanes and other internal interconnects. This is typically the most performance-sensitive part of the AI infrastructure, making bandwidth, latency and power efficiency critical.
As Scale Up expands beyond a single system into trays, chassis or racks, connectivity evolves. In-system electrical links give way to external copper technologies such as DAC, ACC and AEC, with optics entering as bandwidth, reach and density requirements increase.
Scale Up does not automatically mean optics. The goal is to use the lowest-power, lowest-latency connectivity that can support the required bandwidth and reach.
Scale Out: Connecting More Compute
Scale Out connects more GPU servers and racks through a network fabric, allowing AI workloads to operate across a much larger cluster.
This is typically where Ethernet or InfiniBand leaf-spine fabrics come into play. Short server-to-switch connections can still use DACs, AECs or AOCs, while rack-to-rack and switch-to-switch connections increasingly rely on high-speed pluggable optics and fiber infrastructure.
The priorities shift from tightly coupled GPU connectivity toward bandwidth, density and scalability.
Scale Across: Connecting AI Compute Between Buildings
As AI clusters grow, power, cooling and space can make it difficult to keep everything within one data hall or building.
Scale Across extends AI connectivity across data halls, buildings and data centers.
Distances now move from hundreds of meters to kilometers. Direct fiber and conventional Ethernet optics may work for shorter links, but as distance increases or fiber becomes constrained, coherent optics, DWDM and optical amplification become increasingly relevant.
Multiple high-speed wavelengths can share the same fiber, providing a scalable way to connect distributed AI infrastructure.
Think: data hall ↔ data hall, building ↔ building, data center ↔ data center.
Scale Across Is Not Simply DCI
Scale Across can use many of the same technologies as traditional Data Center Interconnect, but the application is different.
Rather than simply connecting separate data centers, Scale Across can support the expansion of the AI compute environment itself across physical locations.
Three Domains, Different Connectivity Requirements
|
|
Scale Up |
Scale Out |
Scale Across |
|
Purpose |
Tightly couple compute |
Expand the cluster |
Connect distributed compute |
|
Typical scope |
System / tray / rack |
Racks / data hall |
Buildings / campus |
|
Priorities |
Latency, bandwidth, power |
Bandwidth, density, scale |
Reach, fiber efficiency |
|
Typical technologies |
Electrical, DAC, ACC, AEC, AOC, optics |
DAC/AEC/AOC, 100G–800G optics |
Coherent, DWDM, amplification |
The boundaries are not absolute. Architecture, speed, density and distance determine which technology makes sense.
The Bigger Trend: Optics Moves Closer to Compute
As AI bandwidth increases, optics is progressively moving closer to the GPU.
Electrical connectivity dominates at very short distances. Copper extends connectivity outside the system. High-speed optics becomes increasingly important across racks and data halls. Coherent and DWDM take over as AI infrastructure extends between buildings.
The objective isn’t to use optics everywhere. It is to use the right technology for the required bandwidth, reach, power and architecture.
LambdaGain addresses the physical connectivity and optical layers across these domains—from external copper connectivity to high-speed optics and coherent DWDM transport.
|
Domain |
Typical Technologies |
LambdaGain Solutions |
|
Scale Up |
DAC, ACC, AEC, AOC, short-reach optics |
High-speed copper/AOC assemblies, 400G/800G optics, high-density fiber |
|
Scale Out |
DAC/AEC/AOC, 100G/400G/800G optics, MPO |
High-speed transceivers, DAC/AOC, MPO, fiber shuffles and passive connectivity |
|
Scale Across |
Coherent, DWDM, amplification |
Coherent pluggables, DWDM mux/demux, pluggable EDFAs, DCMs |
The technology evolves with the network:
- Scale Up: electrical and copper first; optics where required.
- Scale Out: high-speed optics and dense fiber connectivity.
- Scale Across: coherent optics, DWDM and amplification.
There is no single technology that is right for every AI network.
The goal is to right-size connectivity—from the compute system, across the data hall and across the AI campus.
The LambdaGain™ Perspective: Right-Sizing AI Connectivity
LambdaGain addresses the physical connectivity and optical layers across these domains—from external copper connectivity to high-speed optics and coherent DWDM transport.
|
Domain |
Typical Technologies |
LambdaGain Solutions |
|
Scale Up |
DAC, ACC, AEC, AOC, short-reach optics |
High-speed copper/AOC assemblies, 400G/800G optics, high-density fiber |
|
Scale Out |
DAC/AEC/AOC, 100G/400G/800G optics, MPO |
High-speed transceivers, DAC/AOC, MPO, fiber shuffles and passive connectivity |
|
Scale Across |
Coherent, DWDM, amplification |
Coherent pluggables, DWDM mux/demux, pluggable EDFAs, DCMs |
The technology evolves with the network:
- Scale Up: electrical and copper first; optics where required.
- Scale Out: high-speed optics and dense fiber connectivity.
- Scale Across: coherent optics, DWDM and amplification.
There is no single technology that is right for every AI network.
The goal is to right-size connectivity—from the compute system, across the data hall and across the AI campus.
Date
Author
Sandro Fratarcangeli
