Inference
Latency and throughput for serving: time to first token, phase timing, request mix. Model, hardware and request mix must match.
Product
Compare runs, investigate, and keep the evidence.
Kernelwake is a self-hosted server and web workspace for GPU performance decisions. It receives live telemetry and imported captures, compares runs only when they are comparable, and keeps what an investigation relied on.
Three lenses
A lens sets which outcomes count and what two runs must share before they are compared.
Training
Step time and progress toward a quality target. Model and objective, hardware, rank count and effective batch must match.
GPU fleet
Delivered work within capacity: queue wait, attainment, allocation. Pool, device types, scheduler policy and eligibility history must match.
Capabilities
Comparisons
Pick a baseline and a candidate and declare the intended change. Each comparability dimension is reported as the same, intentionally changed, different or unknown. Pairs that cannot be compared, such as mixed capture formats or a lens mismatch, are refused with a stated reason instead of shown as a delta.
Campaigns
A campaign holds a baseline run, an objective metric, guardrails and a set of candidates for one workload. Analysts review each candidate, and the review is recorded. The check command gives a CI job an advisory result: improved, regressed, inconclusive or invalid, each with its own exit code.
Investigations
An investigation keeps a hypothesis and its finding apart from the observations pinned to it. Its status moves through draft, testing, inconclusive, verified and archived, and only allowed moves are accepted. Every edit saves a new revision; pinning new evidence into a verified investigation sends it back to testing.
Evidence library
Workloads, runs and their captures, scoped to your workspace. New imports keep the exact original bytes alongside the parsed data, with digests recorded in their provenance. Verifying an investigation rechecks those bytes; a missing, changed or synthetic capture leaves it inconclusive.
Capture sources
Import Kernelwake JSON captures (up to 16 MiB), Chrome or PyTorch trace JSON (up to 32 MiB) and AIPerf profile exports as zip or tar.gz (up to 64 MiB). Live telemetry arrives from the graphsignal-go client over its wire protocol: metrics, profiles, spans, logs, issues and resources.
Exports
Evidence and investigation bundles as JSON with a SHA-256 checksum, checked offline with graphsignal-server evidence verify. Admins can export the whole workspace as one JSON-lines file. Operators can export usage and an unpriced invoice preview as CSV or JSON.
Telemetry behind every comparison
The workspace also serves the raw signals the client sends: fleet and host views, runs, GPUs, traces, profiles, issues, logs, ingest and resources. Every view is bounded to a stated time window and says when a result was truncated or a value is absent.
Roles
- Viewer
- Reads the workspace.
- Analyst
- Creates comparisons, campaigns and investigations; pins and exports evidence.
- Admin
- Also manages members, ingest keys, sampling policy and workspace export or deletion.
See the workflow end to end.
The six stages, from the first question to a bundle someone else can check.