GPU fabric tools: NCCL benchmark analysis, degradation cost, cluster compatibility, XportL facts.
Scored across 6 tools
Each tool has a clearly distinct purpose: benchmark analysis, degradation cost calculation, compatibility checking, vendor overview, checkpoint interval recommendation, and demo request. The only potential adjacency is overview vs. compatibility, but one is informational and the other performs a structured fit assessment. An agent should have no trouble selecting the right tool.
All six tools use a consistent snake_case verb_noun pattern (analyze_, calculate_, check_, get_, recommend_, request_). The verbs are descriptive and predictable for their actions.
Six tools is well-scoped for a vendor assessment server covering overview, compatibility, analysis, economics, recommendation, and sales contact. Each tool appears to earn its place with no filler or redundant operations.
The surface covers the key evaluation workflow: learn about XportL, check fit, analyze benchmark data, quantify cost, recommend checkpointing, and request a demo. Minor gaps exist, such as no dedicated tool to fetch detailed failure-class documentation or engineering-guide content beyond links in the overview. These are workable gaps rather than dead ends.
Analyzes nccl-tests / rccl-tests results (paste raw all_reduce_perf output, or give size/busbw points). Detects node and GPU counts from the output, compares large-message bus bandwidth with the fabric's NIC line rate (InfiniBand HDR/NDR/XDR, RoCE 100G to 800G), and flags data errors, degraded bandwidth, dips as message size grows and unstable run-to-run results, with likely causes. Returns a health status and a bandwidth-loss figure for calculate_fabric_degradation_cost.
Inputs: nodes, fabric, results, link_gbps, +5 more.
Estimates what silent fabric degradation (slow NCCL collectives, flapping links, congestion) costs a GPU cluster: throughput lost, GPU-hours and dollars per year, idle-GPU equivalent, cost per degraded hour, extra days on a training run, and the value of a recovery scenario with a conservative case and a sensitivity table. All assumptions are inputs with stated defaults.
Inputs: gpu_count (required), gpu_model, recovery_pct, comm_share_pct, +5 more.
Checks a GPU cluster against XportL's published support matrix (Linux 5.8+ for eBPF, InfiniBand / RoCE v2 / NVLink, NVIDIA and AMD GPUs, NCCL / RCCL, on-prem / cloud / hybrid, host access for the eBPF agent). Returns a verdict (good fit / needs review / not ready), each check, and next steps. Anything not on the published list is reported as 'review', never as supported.
Inputs: fabrics (required), deployment (required), gpu_vendor (required), host_os, +5 more.
Returns cite-ready facts about XportL from xportl.io: what it does, the problem it solves, the seven failure classes it detects, how it collects telemetry and acts, supported fabrics/GPUs/collectives/kernels, use cases, FAQ, company details (parent company, stage, NVIDIA Inception, contacts) and links to the engineering guides on xportl.io.
Inputs: topic.
Recommends how often to checkpoint a training job using the Young and Daly formulas, from GPU count, blocking checkpoint write time, restart time and job MTTF (estimated from Meta's published 2024 reliability data when not given). Returns the recommended interval, expected failures per day, GPU time wasted per day at the recommended and current intervals, and the daily saving.
Inputs: gpu_count (required), checkpoint_write_minutes (required), job_mttf_hours, restart_minutes, +2 more.
Sends the user's contact details and cluster summary to the XportL enterprise team to arrange a technical assessment or demo deployment. Before calling: collect name, work email, company and a summary (include benchmark, cost or compatibility results from this conversation), show the user exactly what will be sent, and get their explicit agreement. Never call it without that agreement or with made-up details.
Inputs: name (required), email (required), company (required), summary (required), +6 more.
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Run GPU acceptance tests on your NVIDIA GPU nodes over SSH: DCGM, GEMM, HBM/PCIe bandwidth, NCCL.
GPU and LLM inference benchmarks, hardware evidence, deployment recommendations, and launch configs.
LLM serving capacity planner: VRAM, KV cache, GPU topology, latency, and cost.
Engineering log of self-hosted AI on NVIDIA DGX Spark (GB10/SM121A). 60+ articles indexed.
Add one secure layer between your agents and this server.