Vocabulary

CoreWeave's Slurm on Kubernetes: one bare-metal fleet serves batch and cloud-native jobs.

Source: https://www.coreweave.com/blog/sunk-slurm-on-kubernetes-implementations Checked 2026-09-18

Nebius's open-source Kubernetes operator for Slurm, with a shared root filesystem.

Source: https://github.com/nebius/soperator Checked 2026-09-18

SchedMD's Slurm and Kubernetes projects: an operator that runs Slurm in pods and a bridge that lets Slurm schedule pods. SchedMD is now part of NVIDIA.

Source: https://slinky.schedmd.com/docs/ Checked 2026-09-18

NVIDIA's AI factory management platform. Its licence bundles Base Command Manager, Run:ai, Unified Fabric Manager and NetQ.

Source: https://docs.nvidia.com/mission-control/ Checked 2026-10-06

Opened 2026-10-06: definition rewritten to match NVIDIA's Mission Control documentation, which lists Base Command Manager, Run:ai, Unified Fabric Manager and NetQ as the bundled software.

NVIDIA's provisioning tool: images, node roles, cluster administration. Free for up to 8 accelerators.

Source: https://docs.nvidia.com/pdf/base-command-manager-free-license-faq.pdf Checked 2026-09-18

Data Center GPU Manager: health, diagnostics, power and policy across a fleet, with a Prometheus exporter.

Source: https://developer.nvidia.com/dcgm Checked 2026-09-18

All-reduce benchmarks used as the acceptance test for a new cluster; they expose bad links before production.

Source: https://www.together.ai/blog/a-practitioners-guide-to-testing-and-running-large-gpu-clusters-for-training-generative-ai-models Checked 2026-09-18

Source changed from the draft's github.com/NVIDIA/nccl-tests (whose own README only claims to check performance and correctness, not cluster acceptance) to Together AI's practitioner's guide, which explicitly describes running NCCL tests across a cluster to find bad nodes, leaf switches or spine switches before production.

InfiniBand is deterministic and still leads in SuperPODs. Spectrum-X brings lossless behaviour, adaptive routing and telemetry to Ethernet and closes most of the gap.

Source: https://www.weka.io/learn/enterprise-technology/nvidia-spectrum-x/ Checked 2026-09-18

Opened this session. The page confirms InfiniBand as more deterministic and dominant in SuperPODs/TOP500 systems, and Spectrum-X's congestion-aware, adaptive routing on Ethernet; it does not use the word "lossless" itself, though it describes the same congestion-control behaviour.

UFM is NVIDIA's fabric manager. NetQ is its network monitoring and validation tool, and WJH (What Just Happened) is used to diagnose congestion and packet loss. All three are named in the NCP-AIN exam objectives.

Source: https://dam-cdn.nvd.orangelogic.com/AssetLink/32ljugfxg1hs1sd42371npw1xmcuo1yo.pdf Checked 2026-10-07

Opened 2026-10-07: NVIDIA's NCP-AIN study guide (PDF, read as text) lists NetQ, What Just Happened (WJH) and UFM among the exam's tools, for example 'Diagnose congestion or packet loss using in-band telemetry and NVIDIA What Just Happened (WJH) services' and 'Use NetQ for real-time network monitoring'. NVIDIA's NetQ 5.3 user guide has a 'Configure and Monitor What Just Happened' chapter. The UFM Enterprise manual (https://docs.nvidia.com/networking/display/UFMEnterpriseUMv6170/UFM+Overview) describes UFM as a fabric manager.

NVIDIA's Kubernetes GPU scheduler with fractional GPUs and quotas. Its scheduling core is open source as KAI Scheduler.

Source: https://www.nvidia.com/en-us/software/run-ai/ Checked 2026-09-18

Containers for Slurm jobs without a daemon: srun --container-image.

Source: https://github.com/NVIDIA/pyxis Checked 2026-09-18

Storage that feeds GPUs without passing through the CPU, and the three vendors most often named beside GPU clusters.

Source: https://docs.nvidia.com/gpudirect-storage/overview-guide/ Checked 2026-10-06

Opened 2026-10-06: NVIDIA's GPUDirect Storage overview defines GDS as a direct path between GPU memory and storage that avoids a bounce buffer through the CPU, and names DDN EXAScaler, VAST NFS and WekaFS among supported file systems. That the three are the vendors most often named is this site's reading, not NVIDIA's.

Coolant distribution unit, and the two loops it separates: technology cooling system in the rack, facility water system in the building.

Source: https://www.vertiv.com/en-us/insights/articles/educational-articles/understanding-coolant-distribution-units-cdus-for-liquid-cooling/ Checked 2026-09-18

The Vertiv page defines CDU thoroughly and confirms it isolates the facility water loop from the technology cooling loop, but only names "TCS" and "FWS" in passing without spelling out or defining either acronym directly; the two-loop separation this row describes is well supported, the specific acronym expansions are more loosely supported by this source.

The power delivered to IT equipment, as opposed to gross site power. The number leases are priced on.

Source: https://www.adventuresincre.com/glossary/it-load-critical-load/ Checked 2026-09-18