Kernelwake
Sign in Get started

How it works

Frame, compare, inspect, hypothesize, verify, share.

Every investigation follows the same six stages. The lens changes the metrics and the comparability rules; the stages stay the same.

The six stages

  1. Frame

    Choose the lens (inference, training or GPU fleet), state the decision the investigation should inform, and declare the guardrails a change must not break.

  2. Compare

    Choose a baseline and a candidate and declare the intended change. Kernelwake checks the lens's comparability dimensions first. If the runs differ where they must match, or a dimension is unknown, you see that before any delta.

  3. Inspect

    Follow the outcome into aligned time windows, phases and operations, down to the traces, profiles and logs behind them. Windows are explicit and can be pinned, so a link shows the next reader the same data.

  4. Hypothesize

    Pin the observations that matter as evidence and write the hypothesis separately. Each edit saves a new revision, and an edit made against an older revision is refused rather than merged silently.

  5. Verify

    Attach rerun evidence under the declared guardrails. Marking the investigation verified rechecks its retained sources and derives the comparison again; a missing, changed or synthetic source leaves it inconclusive.

  6. Share

    Export the evidence or the whole investigation as a JSON bundle with a SHA-256 checksum and the recipe it was derived from. Anyone with the bundle can check offline that it is unchanged and internally consistent; that is not, by itself, a verified result.

Getting data in

There are two paths, and an investigation can use both.

  • Live telemetry. Point a graphsignal-go client at the server with an ingest key. An admin creates keys; a key's secret is shown once.
  • Imported captures. Upload a Kernelwake JSON capture, a Chrome or PyTorch trace, or an AIPerf profile export into a run in the library.
$ graphsignal-server init -db signals.db \
    -tenant acme -deployment prod \
    -email "$ADMIN_EMAIL"
$ graphsignal-server serve -db signals.db -ui

Try it on your own comparison.

Sign in to an existing workspace, or ask about a plan.

Get started Pricing