H200 · Live campaign
LingBot, ABot and Echo-WM Flash · Fresh runs · Two resolutions · Three repeats
Refreshes automatically every 15 seconds
Job progress
Waiting for live countsCompleted and prepared repetition entries, including CPU analyses. Blocked work is counted separately; this is not an estimate of remaining runtime.
Reserved GPU pool
Four-card campaign poolH200 identity is user-confirmed; driver may report L20X. The pool remains held between jobs. Same-model jobs can run in parallel with separate GPUs, ports and compile caches. U8 is deferred.
Running now
Waiting for a verified machine snapshot.
Baseline coverage
| Model | 480 × 832 | 704 × 1280 |
|---|
Counts require accepted results from independent fresh processes. Completion does not establish output quality or numerical equivalence.
Needs attention
Jobs by status
Filters apply to all three lists. GPU experiments and CPU analyses are labeled separately.
Running
Queued
Prepared work, ordered by scheduling priority—not a guaranteed start order. Dependencies and available cards still apply; same-model jobs may overlap. U8 remains deferred.
Blocked / not yet implemented — not in the runnable queue
Completed
Accepted results only. Expand a row for P95/P99, sample counts and measurement scope.
Recent activity
How to read these results
Each configuration requires three independent runs. P95 and P99 are empirical sample quantiles, not a guaranteed latency SLA.
Core, decoded-wall and component latencies have different measurement boundaries. They are not directly interchangeable.
Parallel jobs share host CPU, storage and power resources. Overlap is recorded locally; do not treat these as isolated-machine measurements.
Inference progress excludes any remaining decode, output writing and acceptance checks. No unverified finish-time estimate is shown.
This page observes only. It cannot cancel jobs, alter reservations or retry failures.