JournalDAY 27 / X

FIELD NOTE / X

GPU telemetry is not an invoice ledger.

The complete written thought and the evidence behind it. The video edition will follow its public release.

Journal September 25, 2026 · X target October 24, 2026

GPU telemetry is not an invoice ledger.

Day 27 · Week 4 editorial group · X · no publication date or time assigned

Video caption

GPU telemetry is not an invoice ledger. Utilization and charged allocation are different facts. Preserve the gap instead of inventing precision. #EricFieldNotes

Full written post / accessibility read

DCGM can summarize a scheduler job, but its retained job record is not a durable accounting database. If the host engine restarts before export, the operator may lose the samples needed to explain a usage line.

Imagine a job waiting on checkpoint storage while its allocation remains held. The scheduler knows the reservation, DCGM has device observations, and the application knows progress. None alone says what the contract should charge.

Export DCGM job stats in the scheduler epilogue; persist allocation start, stop, interruptions, device UUIDs and sample gaps independently. Join application checkpoints and accepted output. Keep invoice adjustments as their own ledger entries, not a rewrite of raw usage.

If a watch was late or lost, mark the interval unknown and resolve it under the customer's agreement. Do this because a credible compute business needs replayable receipts, while a GPU counter by itself cannot decide a commercial obligation.

#EricFieldNotes

Four-beat scene transcript

1. GPU telemetry is not an invoice ledger.

DCGM can summarize a scheduler job, but its retained job record is not a durable accounting database. If the host engine restarts before export, the operator may lose the samples needed to explain a usage line.

Visual: A restarted host engine can lose the job window.

2. The customer disputes an unexplained interval.

Imagine a job waiting on checkpoint storage while its allocation remains held. The scheduler knows the reservation, DCGM has device observations, and the application knows progress. None alone says what the contract should charge.

Visual: Utilization and charged allocation are different facts.

3. Write a durable per-run receipt.

Export DCGM job stats in the scheduler epilogue; persist allocation start, stop, interruptions, device UUIDs and sample gaps independently. Join application checkpoints and accepted output. Keep invoice adjustments as their own ledger entries, not a rewrite of raw usage.

Visual: Join allocation events, telemetry, progress and contract version.

4. Make every charge explainable.

If a watch was late or lost, mark the interval unknown and resolve it under the customer's agreement. Do this because a credible compute business needs replayable receipts, while a GPU counter by itself cannot decide a commercial obligation.

Visual: Preserve the gap instead of inventing precision.

Research and claim limits

The examples identified as illustrative or simulated are design probes, not reported incidents. Vendor specifications do not establish workload performance.

More notes from the work ↗