Ryzen CPU slowdown observations and issue triage

Confirmed ≥1.2× follow-up

Current public-API MWEs and paired confirmation are in followup_12x.md. This historical observation inventory is screening evidence only; the follow-up report distinguishes confirmed new Issues, existing open owners, scope-ineligible trace paths and inconclusive results.

• Source tenferro-rs: 5cf78c7ec0ad9516dd78bab546d5e7bd42fa3102; Ryzen 9 9955HX; Linux MKL devcontainer; unpinned 1T/4T.
• Upstream main advanced after collection to b3f47296244ff7b7c55ac0a75f782cb0835418c1 (#2026). That revision removes the optional tenferro-cpu-tprims crate expected by this harness; latest-main compatibility is not established by these results. PR CI validates the measured compatible revision above.
• Issue search: all 1,173 pre-existing open/closed issues, plus focused operation searches, 2026-10-08.
• Inventory: 1585 rows with tenferro median above the fastest recorded external comparator; 377 operation/path groups. This includes flagged diagnostics, not 1585 confirmed defects.
• Full inventory and per-row source/provenance: data/results/amd-cpu/cpu/slow_analysis/20261008_review/slow-observations.csv.

Confirmed MWE and negative controls

These are current-host cross-implementation comparisons, not revision regressions. Inputs/setup/session entry/priming/cleanup are untimed; intrinsic output allocation is timed and outputs retained. Four balanced paired rounds follow same-build A/A. Immutable declarations precede the PyTorch samples. Rust/PyTorch use MKL family, with independently recorded vendor versions. Logical fixtures and requested outputs match; native layouts are preserved. The matched reshape control returns the same compact column-major [8192,4096] values.

Raw declarations, A/A and paired sample JSON live in data/results/amd-cpu/cpu/slow_analysis/20261008_{issues,extended,reshape_matched,activation_softmax,gelu_10ms,gelu_shape_matched}/. The earlier extended reshape measurement had a different logical output shape; the matched-output run supersedes it. The original eager FFT/direct session-inclusive data are not used to establish operation gaps. Initial 2 ms GELU comparisons failed the declared noise gate; the final 10 ms logical-shape-matched rerun passed the same gate and supersedes those ratios and the earlier flattened-reference control. Feature/API-request IDs alone are not focused performance owners. Large IFFT and cast did not reproduce the original large ratios; no issue is filed from those observations.

Timing and issue-match limitations

• excluded: timed session entry: source inspection found public-API generic direct wrappers or eager entry inside the measured closure. Do not interpret these as operation-performance evidence; the standalone borrowed-session MWE is the replacement where available.
• short isolated-clock diagnostic: the source samples one call per interval. Standard short-operation claims require batched intervals; these rows are candidates for reproduction, not confirmed throughput gaps.
• Prepared trace rows have graph compilation/plan preparation outside timing, but runtime dispatch is part of that execution path; they are not immediate direct calls or borrowed eager calls. Candidate observations have not individually passed paired confirmation.
• Existing closed issues remain closed in this inventory; the matching symptom does not prove the old cause/regression. Related/umbrella matches are explicitly weaker than focused ownership. Apple-specific #1900 is only a related lead for Ryzen complex GEMM, not a confirmed match.
• References using preallocated destinations are flagged and are not allocation-equivalent comparisons. Julia uses OpenBLAS while tenferro uses MKL; provider-dependent comparisons do not isolate a tenferro implementation bottleneck.
• Original cpu/perf_issues PyTorch sampling grew the retention list inside timing. The runner now preallocates it; historical ratios are screening observations, and only the separately confirmed MWEs establish the reported gaps.
• cpu/small_work currently contains no external comparator in its recorded raw rows, so it cannot establish a cross-implementation slowdown here. Unsupported/failed/missing rows are never assigned performance ratios.

Complete operation/path inventory

Each row below is the largest observed ratio in that operation/path group. CSV retains every dtype, shape, thread count and external comparator; ratios are descriptive observations until the evidence column says otherwise.