The measurement layer for compute.

Published GPU specifications generalize across silicon and often overlook details that have to be considered in practice: the silicon lottery, heterogeneous power and networking setups, or cooling. We measure node capabilities in the context of what they're meant to do, providing custom baselines and application-specific fleet optimization. Measurements come from short reference workloads that behave like real training and inference jobs, including synchronized training steps, collective communication, sustained matmul throughput, and checkpoint I/O under load.

We have worked with neoclouds and pre-training companies to understand their compute and optimize their fleets, tailored to their clientele and specific use cases.

If you're interested in working with us, reach out. We provide anonymized sample reports from real paid engagements on request.

Our checks measure every node against a baseline tailored to that specific GPU rather than a threshold copied off a datasheet, then go deeper into fleet-specific health: catching stragglers before one slow rank sets the pace of a synchronized job, attributing slowdowns to the component responsible, and flagging a card that is merely below par as clearly as one that is broken.

• GPUDriver and device visibility, ECC health, memory bandwidth.
• ComputeSustained stress, matmul FLOPS, short training probes, and silent-data-corruption checks.
• InterconnectNVLink and PCIe link health, NCCL collectives, multi-node bandwidth and its variance.
• HostCPU and memory sanity, NUMA affinity, essential daemons, clock sync.
• NetworkNIC link health, inter-node bandwidth, RoCE checks.
• StorageFilesystem availability and checkpoint I/O performance.
• ThermalTemperature and fan behaviour, throttling under sustained load, and cooldown profiles.

We provide coverage across NVIDIA and AMD, on Slurm, Kubernetes, or plain SSH, for acceptance testing, burn-in, and continual monitoring.

The suite that produces our reports is the same one you install. Nothing is held back for the hosted version.

Every run updates a node's trust score. We combine recent test results, runtime fault events, and recovery history in a recency-weighted Bayesian update, so the score reflects both current hardware health and recent track record.