Cluster track · Presentation
Your AI Cluster's Real Bottleneck Is How Fast It Can Change
- When
- Wednesday, February 3, 202712:35 PM to 12:50 PM PT
- Where
- Room BSan Jose McEnery Convention Center
- Track
- ClusterArchitecting 1M GPU Clusters. Programmed by CloserStill with OCP
- Format
- Presentation
About this session
As large-scale AI clusters scale to tens of thousands of accelerators, hardware failures become an inevitable, regular occurrence rather than exceptional events. Consequently, the true bottleneck to maintaining high compute efficiency ("goodput") lies in the operational speed and safety of the release process, updating the complex, co-designed stack of firmware, drivers, run-times, and diagnostics across a disaggregated fleet while training jobs continue running.
This article examines the unique challenges of release engineering in AI data centers, where silent interactions between components, tight hardware-software dependencies, and high-power density can trigger fleet-wide performance degradation or job interruptions. Key strategies to mitigate these risks include establishing baseline-relative health metrics, subjecting rollbacks to rigorous stress testing on dedicated fleets, and leveraging AI coding agents for specific tasks like failure triage and workspace cleanup. Ultimately, while hardware design sets the theoretical ceiling for AI performance, operational success is determined by how quickly and safely teams can deploy changes across the underlying infrastructure.
General rate, until Feb 2, 2027
Register for Digital Infrastructure World 2027
One pass covers this session and every other session, plus OCP Tech Day, the keynotes and the exhibit hall, on all three days.
- All Access$799
- Open Compute Project members$399
- Teams of 3 or more, each$599.25
More in the Cluster track
- Current Activities in the Open Cluster Designs for AI Strategic Initiative
- From Hardware to Useful Capacity: Provisioning AI Infrastructure at Scale
- Beyond Physical Digital Twins: Why Logical Digital Twins Are Essential for AI Factory Design, Validation and Operations
- Node-to-Rack Efficiency for Real-World AI Inference: Arm and FuriosaAI