← GPU latency results

gpu/compare_lifetime_counters

profile: nvidia-gpuno machine-readable summary

Legacy report: ingestion provenance is unrecorded; no raw artifact is inferred.

Original report

Public CUDA lifetime counters

Raw observations: data/verification/nvidia-gpu/gpu/compare_recheck/20261014_010820/lifetime_counters.jsonl.

These observations were collected separately from timing on debug GPU1. They are not performance measurements or standard operation coverage. Each case uses one entered backend session and retains outputs until native completion.

Driver free memory covers the whole device. Its change cannot isolate an output pool, prove a leak, or explain the separately observed latency growth. No public output-pool statistics or clear API is available.

Case Batches Operations/batch Maximum normalized error Driver free-byte change after final drop Workspace retained bytes Retirement in flight / leaked
compare_f32_n1024 30 128 0 0 0 0 / 0
compare_f64_n1024 30 128 0 0 0 0 / 0
compare_eq_f32_n1024 30 128 0 -8388608 0 0 / 0
compare_lt_f32_n1024 30 128 0 0 0 0 / 0
compare_le_f32_n1024 30 128 0 -8388608 0 0 / 0
compare_ge_f32_n1024 30 128 0 0 0 0 / 0
compare_eq_f64_n1024 30 128 0 -8388608 0 0 / 0
compare_lt_f64_n1024 30 128 0 0 0 0 / 0
compare_le_f64_n1024 30 128 0 -8388608 0 0 / 0
compare_ge_f64_n1024 30 128 0 0 0 0 / 0

Raw JSONL preserves the cache entries, bytes, hits, misses, evictions, clears, workspace retirement counters, and pending transfer retirements at enqueue, completion, and output drop. Queries can perturb execution; they are not synchronization primitives.