FAQ
Questions and plain answers.
Frequently asked questions
What is Kernelwake for?
Deciding whether a change made GPU performance better or worse, and keeping the evidence for that decision. It compares a baseline and a candidate for inference, training or GPU fleet work, checks that they are comparable first, and records the investigation that follows.
What data can I bring?
Live telemetry from the graphsignal-go client, and imported captures: Kernelwake JSON captures, Chrome or PyTorch trace JSON, and AIPerf profile exports as zip or tar.gz. Nsight files are not imported directly.
Where does it run?
On a server you control. It is a single Go binary that stores data in SQLite and listens on loopback by default. It has no TLS of its own, so put it behind a TLS-terminating proxy to reach it from elsewhere.
Are the numbers on the home page real?
No. They are synthetic examples to show the workflow and are labelled that way. Kernelwake has not published benchmarks or customer results.
What happens when two runs are not comparable?
The comparison says which dimension differs or is unknown. A pair that cannot be compared at all, such as mixed capture formats or a lens mismatch, is refused with a reason rather than shown as a delta.
Can someone check an exported bundle without my server?
Yes. graphsignal-server evidence verify bundle.json rechecks the bundle's SHA-256 checksum and that its embedded data matches its digests and any verdict it states, offline. Bundles are not signed, so the check shows the bundle is unchanged and consistent, not who produced it, and a valid bundle is not by itself a verified or improved result.
How much does it cost?
Prices are not yet published. The pricing page lists the limits a plan sets.
Is it production ready?
Not yet. An independent review and qualification on a hosted deployment are still pending.
Do you have SOC 2 or other certifications?
No. See Security for what the server does today and its current limits.
Where are the terms of service and privacy policy?
They are not yet published.