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.
Accepted or nearly accepted cluster with unproven workload behavior.
Acceptance ledger with bottlenecks ranked by operating impact.
Proceed, remediate, or defer larger model-production work.
Platform, AI infrastructure, or ML systems lead.
The surfaces we validate.
NCCL / fabric collectives
Validate that collective communication lands near the topology ceiling before long training runs are trusted.
The work inside Cluster Diagnostic.
Fabric and collectives
Measure NCCL/RCCL bandwidth, rail usage, NUMA placement, rank layout, and topology-aware performance against the cluster ceiling.
Data path and storage
Trace ingest, sharding, tokenization, prefetch, dataloader stalls, and storage throughput until GPU starvation is visible and quantified.
Scheduler and placement
Exercise gang scheduling, queue behavior, fragmented capacity, priority/preemption policy, and large-job placement across fabric domains.
Training step proof
Run a representative model step with the expected parallelism strategy and report MFU, memory headroom, step-time variance, and failure modes.
Serving smoke test
Stand up a baseline endpoint and inspect latency, throughput, KV-cache behavior, batching, and cost-per-token directionally.
Operating readout
Convert benchmark evidence into a prioritized bottleneck ledger, remediation estimate, dependency map, and next-phase SOW.
- 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