CPU ≥1.2× follow-up: current public operation MWEs

2026-10-08, AMD Ryzen 9 9955HX, Linux CPU MKL devcontainer. 43 operation workloads were confirmed and reported in 9 new tenferro-rs Issues. One workload is already covered by open #2021. Two workloads remain inconclusive; seven did not remain ≥1.2× suspects in the operation-only 1T/4T scan. A scan below threshold is not a confirmed parity claim.

Selection: 53 operation groups / 139 original condition rows from the earlier inventory without open owners, selected because at least one original ratio was ≥1.2. All 53 have runnable public-API reproducers. The earlier table rows are screening evidence only. Metadata fixtures are deliberately smaller; their original large-fixture timings are not confirmed. Broadcast #2021 was identified by a renewed full-body open-Issue search before filing and excluded from new reports.

Library: b3f47296244ff7b7c55ac0a75f782cb0835418c1, pulled main then deliberately pinned for the campaign. tenferro BLAS linkage and native FFT reference: oneMKL 2026.0.1; tenferro FFT itself uses cached RustFFT 6.4.1 plans. PyTorch 2.12.0+cpu, wheel MKL 2024.2. JAX/jaxlib 0.10.1, CPU. Julia 1.13.1. strided-rs 12ff2de906f4d894a4329f7dbf195b0f874f846e. No CUDA, no runner taskset/numactl pinning; tenferro public backend-managed placement is retained. Provider/config metadata and hardware.

Setup is untimed on every arm, including calibration: fixture creation/conversion, wrapping, contexts/sessions, input views/descriptors, FFT planning, JIT and first completed initialization. Timers contain only the declared operation, intrinsic allocation, output retention assignment and native completion. Results stay alive past the clock; downloads/signatures/validation/destruction are untimed. Direct/eager paths hold one entered session. Metadata is session-free. MWE setup/commands, timing audit.

Confirmation: 3 warmups and 15 batched samples per process; four A/A pairs and four independent balanced comparison pairs AB/BA/BA/AB. Statistic: median of four round-median ratios. Declared gate: ratio ≥1.2, A/A median within 10%, every sample-set CoV ≤20%, numerical validation passed. The corrected/longer-batch phases also balance A/A and require each A/A round within 10%. No threshold was estimated from candidate results or loosened. Scan selects the strongest eligible 1T/4T condition; the table does not claim both thread counts are slow.

Batch targets are 10 ms or 50 ms for longer-batch retries, bounded by 512 MiB retained tensor payloads/owned metadata inputs. One output is still retained for the 2 GiB rotation. The final 2×2 metadata-transpose reference uses ~1.86 million operations in ~1.5 ms at its memory cap; normalized sub-ns values are specialized bulk metadata throughput, not isolated call latency. Absolute tenferro metadata costs are hundreds of ns. Output shape/count, aggregates and ~130 deterministic probes agree; activation/permutation outputs additionally pass full scalar/odometer oracles. FFT checks input preservation as well.

Compiled JAX uses dynamic arguments and completed execution. Its 1/4 active Eigen workers were verified with untimed probes. Native layouts are preserved with equivalent logical values; no one-sided layout conversion is timed. FFT compares cached native public oneMKL DFTI, not PyTorch FFT: PyTorch CPU FFT replans per call, so its historical rows are not operation-only evidence.

New reports

All selected workloads

Times are per operation in µs. Confirmed rows show independent paired results; inconclusive ratios are diagnostic and were not filed. Screen-only rows show the strongest eligible scan condition. The exact physical permutation and source strides are in the fixture definitions. The default MWE fixture catalog preserves original observation identities.

The maintainer-requested pinned Host DynRank/static-rank extension for #2040 is reported separately in CPU metadata Host comparisons; it remeasures all representations in balanced rounds and preserves the original TensorValue observations above.

Evidence and supersession

The selected prepared-trace paths are explicitly scope-ineligible, since their public API cannot borrow an execution session and includes internal admission/session work. They are not counted as confirmed, parity or covered. This confirmation campaign uses the independent crate and dedicated registered-case runner. The root CPU-provider migration landed separately on main during the campaign; the original measurement revisions are preserved. The follow-up does not refresh the complete run_all suite.

Canonical precedence is encoded in the report formatter. Original confirmations involving Julia first-calibration JIT or PyTorch FFT planning are superseded, never used as operation evidence. The old transpose-metadata 32×32 case was inconclusive and is replaced by the 2×2 long batch. Earlier PyTorch materialization closures with timed input-view construction are superseded. Earlier scatter replacement semantics were corrected to additive scatter in both arms before eligible confirmation. Preliminary CUDA-wheel, constant-captured JAX, unspecialized Julia and incorrect Fortran-symbol DFTI probes are not publication evidence. No measured timing/status values were edited.

Per-process JSON is preserved losslessly in each campaign’s processes.jsonl.gz (restore instructions); it includes exact per-process commands, source revisions, counts, durations and validation signatures. Submission identities retain MWE source commits and case membership. Workloads and reference arms are added to cpu/perf_issues at opening (manifest v4; Host extensions in v5); the exact existing activation cases are annotated instead of duplicated. Use scripts/run_cpu_followup.sh inside the Linux MKL devcontainer to run registered latest-main cases. No automated audit, registry, scheduled run or mandatory baseline campaign is introduced.