How to Network LLMs. Splitting Large Language Models Running on GPUs Across a Network - BRKOPS-2931#

Dave Zacks (Cisco), Patrick leMaistre (Cisco)

• Why: Not a core interest for this PM, but inverts the usual AI-fabric question: instead of "what network does training need," it asks "what the network must look like to host disaggregated inference/models across hosts" — on-prem, with private local datasets. A different lens on why DC fabric design is changing.
• What: How LLMs that don't fit in one GPU or one server are split across multiple hosts and linked by the network, focusing on what a properly designed network must provide for distributed, on-premises AI model hosting (from abstract only).
• How: Deep-dive walkthrough (by two veteran Cisco Live speakers) of model partitioning across systems and the network requirements that result.
• Where applicable: On-prem AI model hosting scenarios; likely conceptual rather than benchmarked, and inevitably Cisco-portfolio-flavored. The abstraction transfers to any fabric supporting split models.