How the Data Center Shapes the Workload
summary
Describes a fleet manager implemented as a state machine from power‑on and multi‑day burn‑in to Kubernetes admission that filters silent stragglers before customers see nodes, runs end‑to‑end production workloads to bake in optimal env vars and PCIe/topology fixes, and earns NVIDIA exemplar status by training half a dozen LLMs. Calls out topology facts: NVLink 72 racks today with a roadmap toward 576, prefer InfiniBand within the same scale unit, and ask about SHARP-enabled switches.
AI-written summary · about
chapters (7)
- 0:00Introduction and setup
- 1:10CPU and GPU collocation
- 3:58Networks in data centers
- 5:02Backend network and topology
- 7:11NVLink impact on AI workloads
- 8:58Switches as computers
- 10:01nscale control center
related talks
39mPanel: Cracking the Data Locality Puzzle
43mCan data drive data center hardware configuration?
16mTowards Optimal Rack-scale μs-level CPU Scheduling through In-Network Workload...
43mSystems @Scale Summer Q&A Week 2
02mStarcluster
21mHandling Traffic Spikes Through Automated Prioritized Load Shedding
21mKeynote: CERN Experiences - Ricardo Rocha & Clenimar Filemon
21mThe Evolution of Traffic Routing in a Streaming World