GPU Networking
Measure and diagnose distributed GPU networking performance: NCCL collectives, RDMA transport, PCIe topology, link health, IOMMU and ACS. Use for current bandwidth/latency questions and follow-up bottleneck investigations or corrections.
Install
npx skills add NVIDIA/AI-Factory-Operations-Agent --skill gpu-networkinggit clone --depth 1 https://github.com/NVIDIA/AI-Factory-Operations-Agent.git /tmp/AI-Factory-Operations-Agent
mkdir -p ~/.claude/skills
cp -r /tmp/AI-Factory-Operations-Agent/helm/mosaic-stack/files/openclaw-seed/skills/gpu-networking ~/.claude/skills/gpu-networkingPersonal skills live in ~/.claude/skills. Use .claude/skills in a project folder to share a skill with your team.
From the SKILL.md
# GPU networking and performance All documentation paths below are relative to the agent workspace. These files are installed from the Helm chart. Read them locally; no web lookup is required. ## Measurement Resolve the workload's pods and nodes with `run_kubectl`. Establish the collective, payload, ranks, participating devices, units and sampling window from current metrics, bounded logs and benchmark source when needed. Report observations without rating performance against an assumed target. Use `skills/gpu-networking/references/measurement-interpretation.md` for quantitative comparisons. ## Diagnosis Complete this sequence before answering a cause question. Keep a record for **every participating host**, not just one representative machine. 1. Inspect workload configuration and initialization logs to identify the selected GPU and HCA, actual data transport and GPU-direct registration. Socket bootstrap is not the data transport. Read negotiated link speed and error counters. Do not infer them from hardware family names. 2. Use `run_kubectl_admin` with `exec` through the existing workload pods when their security context and mounted devices permit diagnostics. Pass `-i` before `--` when supplying stdin. Inspect actual visibility; ordinary sandbox `exec` is not the workload node. BCM supplies inventory and managed health, not an arbitrary shell through invented context parameters.
Details
- Skill name
gpu-networking- Repository
- NVIDIA/AI-Factory-Operations-Agent
- Path
helm/mosaic-stack/files/openclaw-seed/skills/gpu-networking/SKILL.md- License
- Apache-2.0
- Repo updated
- 17 November 2026
FAQ
What does the GPU Networking skill do?
Measure and diagnose distributed GPU networking performance: NCCL collectives, RDMA transport, PCIe topology, link health, IOMMU and ACS. Use for current bandwidth/latency questions and follow-up bottleneck investigations or corrections.
How do I install the GPU Networking skill?
skills CLI (Claude Code, Cursor, Codex and more): npx skills add NVIDIA/AI-Factory-Operations-Agent --skill gpu-networking. Skills go in ~/.claude/skills (personal) or .claude/skills (project), and Claude loads them automatically when a task matches the description.
Is GPU Networking popular?
It lives in NVIDIA/AI-Factory-Operations-Agent, which has 43 GitHub stars and 18 forks.
Related skills
AI Factory Operations Agent Headless
43NVIDIA/AI-Factory-Operations-Agent
Use AI Factory Operations Agent without opening the browser UI. Invoke it through the headless HTTP API, packaged CLI, or stdio MCP server, and deploy it first when no service is available.
Bcm
43NVIDIA/AI-Factory-Operations-Agent
Inspect Base Command Manager (BCM) inventory and health evidence, and make explicitly requested changes when Base Command Manager edit mode is enabled.
Grafana
43NVIDIA/AI-Factory-Operations-Agent
Create and open Grafana dashboards for metrics, Loki logs, and mixed panels. Use when the user asks for a dashboard, visualization, chart, panel, or graph of GPU, Base Command Manager (BCM), node, network, InfiniBand, Prometheus metrics, or Loki logs.
Serenity Analyst
36videGavin/serenity-analyst
Analyze AI-hardware / semiconductor stocks using the 'supply-chain chokepoint' methodology distilled from the X investor Serenity (@aleabitoreddit). Use when the user wants to analyze an AI-hardware/semi ticker, find upstream supply-chain bottleneck opportunities, decide which upstream supplier to buy when a popular chip stock runs, or wants a Serenity-style investment thesis. Covers substrates (InP/SOI/glass), photonics/CPO, lasers, optical modules, HBM/memory, advanced packaging & equipment, foundry, power, and NeoCloud.
Zone of Truth
113Hmbown/Wizards-of-the-Ghosts
Enter a mode where every assertion is backed by evidence, confidence is calibrated, speculation is flagged, and hallucination is actively resisted.
Slurm
43NVIDIA/AI-Factory-Operations-Agent
Inspect mounted Slurm accounting exports and job logs for failed workload root cause analysis.
Browse the most popular Claude skills, all finance & crypto skills, or write your own with the skills generator.