LINGBOT / FOUR-GPU ANALYSIS

What the evidence
can tell us.

Recorded, missing, and the consequences for the next analysis.

Loading source status…

Timeline tracing ≠ hardware counters

Recorded with Nsight Systems

--trace=cuda,nvtx,osrt
--sample=none --cpuctxsw=none
--cuda-graph-trace=node

CUDA activity timestamps and API events. CPU sampling was disabled.

Nsight Compute: root access verified

ERR_NVGPUCTRPERM
RmProfilingAdminOnly=unknown

Root probe recorded real SM cycles and DRAM reads. Full LingBot counter measurements remain pending. No driver policy was changed.

Permissions are feature-specific. NVIDIA limited-trace documentation · Hardware-counter permissions

Evidence → analysis impact

Show

What can proceed now?

CAN PROCEED

Timeline and repeatability analysis

Export and validate both timelines. Correlate process, rank and stage. Compare clean timing and numerical variation. Repair capture and replay identical inputs.

DEPENDENCIES REMAIN

Hardware attribution and acceptance

Counter access gates measured memory traffic and hardware attribution. Exact inputs gate valid backend comparisons. Larger samples and service measurements gate P99 and client-FPS claims.

Baseline runs CPU-ready boundary

Independent run	Steady FPS	Rollout FPS	Steady samples

80 steady samples/run is below the plan’s ≥300 screening and ≥1,000 final requirements. Numerical equivalence remains unresolved.

Capture attempts: exit success is not evidence success

Resolution / TP / SP	Workload exit	QKV files	Metadata logs	Capture evidence