Solution 01

Prove what the cluster can really run.

A focused technical assessment that tests the accepted hardware against real machine-learning workloads and produces a ranked plan for closing the gaps.

01 / Cluster Diagnostic

Cluster Diagnostic

Findings + plan

Use this when the cluster is accepted, but real ML workload behavior is still unproven.

Operating outcome

What changes

A measured view of the cluster: what works, what is underperforming, what is blocked, and what should be fixed before larger model-production work starts.

Best fit

When to use it

Best when a new private GPU environment has arrived or is about to be accepted, but the internal team has not yet proven distributed training, data path, scheduling, evaluation, and serving on real workloads.

Entry state

Accepted or nearly accepted cluster with unproven workload behavior.

Primary artifact

Acceptance ledger with bottlenecks ranked by operating impact.

Decision output

Proceed, remediate, or defer larger model-production work.

Buyer owner

Platform, AI infrastructure, or ML systems lead.

Diagnostic trace

The surfaces we validate.

workload_acceptance_trace fabric validation
rack NCCL I/O MFU eval p95 runbook
input: accepted hardware output: reproducible model run + operating handoff
stage 02 / fabric

NCCL / fabric collectives

Validate that collective communication lands near the topology ceiling before long training runs are trusted.

validate
all-reduce, all-gather, rail usage, NUMA placement
failure signal
below-ceiling bandwidth, cross-socket traffic, NCCL timeouts
instrument
nccl-tests, topology dump, per-rank bandwidth profile
Workstreams

The work inside Cluster Diagnostic.

/01

Fabric and collectives

Measure NCCL/RCCL bandwidth, rail usage, NUMA placement, rank layout, and topology-aware performance against the cluster ceiling.

/02

Data path and storage

Trace ingest, sharding, tokenization, prefetch, dataloader stalls, and storage throughput until GPU starvation is visible and quantified.

/03

Scheduler and placement

Exercise gang scheduling, queue behavior, fragmented capacity, priority/preemption policy, and large-job placement across fabric domains.

/04

Training step proof

Run a representative model step with the expected parallelism strategy and report MFU, memory headroom, step-time variance, and failure modes.

/05

Serving smoke test

Stand up a baseline endpoint and inspect latency, throughput, KV-cache behavior, batching, and cost-per-token directionally.

/06

Operating readout

Convert benchmark evidence into a prioritized bottleneck ledger, remediation estimate, dependency map, and next-phase SOW.

What you keep
  • Benchmark traces and topology evidence
  • Bottleneck ledger ranked by business impact
  • Readiness score across fabric, data, scheduler, training, eval, and serving
  • Costed remediation plan
  • Executive readout and technical appendix
  • Recommended next-phase activation scope