einsum-opt_flops on zen5-cpu938a25fba9e76fb17dff6086db59e592f918ec22defaultcpueinsum (the provider in the measured revision; see library above)18cc16f7783f1c37ec333e754f6b6fd2ac5cff3c467573e6e3e405b8ec30b3b61d096c1d0c9893c36c524ff88eb2d548c7913b7ca6c960e3b46ac4263bc3d91f1c6cf24396c2ef5ezen5-cpudevelopment — provisional: recorded on a development host, superseded by the measurement host's dataset2026-10-27T05:46:53.998799Zscripts/v2/record_run.py zen5-cpu einsum-opt_flopsdata/v2-results/cpueinsum/zen5-cpu/einsum-opt_flops/20261027T054311Z/tenferro-einsum-benchmark problem set: pairwise matrix products, batched and elementwise products, outer products, a matrix-chain problem, MPS inner products, two MERA networks, two tensor-network permutation problems, and two likelihood contractions of 38 and 84 operands. The operands are generated by the harness (deterministic, dtype-templated) and the contraction order is taken from the corpus's own paths entry, so the rows measure one library against two declared schedules, not two libraries.
Source: The corpus is tenferro-einsum-benchmark data/instances/*.json at the revision in pins/tenferro-einsum-benchmark.rev, materialised by scripts/setup_extern_deps.sh. Each instance declares name, format_string, shapes, dtype and two opt_einsum-style positional paths; crates/cpueinsum-bench/src/corpus.rs in the measured revision translates them into cpueinsum's integer-label, SSA-order API.
Known blind spot: The corpus instance str_mps_varying_inner_product_200 is not declared: the harness's own cross-row verification refuses it (MISMATCH(opt_flops) rel_err=NaN, exit 1) at every declared thread budget. It is therefore not measured, not skipped, and not part of this suite's coverage; the evidence is in docs/verify-refusal-str_mps-varying-inner-product-200.md.
Known blind spot: The declared identity is the numeric sentinel n = 1: every instance has fixed extents, so the declaration grid has one shape and the real operand shapes are the harness's case_m (operand count), case_n (total input elements) and case_k (output elements) columns, not geometry the toolkit compares.
Known blind spot: The engine rows are two contraction orders of the same library, so a row here is a schedule comparison inside cpueinsum, not a comparison against an independent implementation; there is no reference engine and no vendor arm in the first record.
Known blind spot: Every instance has unit batch, so a cost proportional to batch elements is invisible.
Known blind spot: One size per instance: a ratio here says nothing about a neighbouring instance, and the set is not a sweep.
Known blind spot: A corpus revision is not one of the measured repository's invalidated_by paths: the declaration records the corpus pin it was generated from, and a corpus bump has to be a deliberate regeneration of this suite.opt_flops — cpueinsum executing the corpus's paths.opt_flops order: the order the corpus declares as minimizing the operation count.
Identity: the measured cpueinsum-rs revision in this run's manifest (library.commit), pinned by pins/cpueinsum-rs.rev
Caveat: The two orders coincide for most instances in this corpus; where they do, a small difference between the rows is measurement noise, not a scheduling effect.
As measured: cpueinsum (the provider in the measured revision; see library above)AMD Ryzen AI 9 HX 4702424 MiB (2 instances)0-3: 16 MiB shared; 4-11: 8 MiB sharedLinux 7.0.0-38-generic / x86_64shinaoka-EVO-X1Pro4, 4T -> 4-7Every timed cell matches a successful verification census, including operation variants.
Each cell is median of process medians ± sample SD of those medians; W is the median
within-process sample SD. All values are milliseconds. These are descriptive spreads,
not confidence intervals. The between-process spread includes the rotating-order schedule.
Each process retains 5 chronological calls; 2 complete process sets are recorded.
Spread guide (computed from this run's own rows): per-engine between-start range median 0.83%, worst 4.37% among cells of 1 us or more (a sub-microsecond row is timer and scheduler jitter, not kernel time); 0 of 26 engine cells exceed 5% and carry !; median within-process relative SD 0.30%. A cell whose ±, ! or p is as large as the difference you are reading is not evidence of a regression.
no row, so no ratio there is position-balanced). ! marks a cell where some engine's own
per-start median range exceeds 5% of its median, and p a balanced cell whose estimated
position-effect ratio exceeds 1.05. Flagged cells are retained, not discarded; the
between-start variation can exceed the difference being compared.
! between-start range over 5%, p position effect over 5%, dispatch-dominated means the
row is under 20 us so 4T/8T pay thread-region start-up). Per-start ratios and every
chronological sample are in the run directory named at the top of this page (run*.csv).
These describe the recorded revisions; they are not a regression claim.
| case | opt_flops (ms) | notes |
|---|---|
| bin_batched_matmul_b32_m64_n64_k64/n1/b1 | 0.552631 ± 0.00143118 (W 0.00161117) | |
| bin_elementwise_mul_2048x2048/n1/b1 | 4.01557 ± 0.0716992 (W 0.0155111) | |
| bin_matmul_256/n1/b1 | 1.07044 ± 0.00111228 (W 0.000668953) | |
| bin_outer_product_4096/n1/b1 | 6.1641 ± 0.00309501 (W 0.0132176) | |
| gm_queen5_5_3.wcsp/n1/b1 | 4472.66 ± 132.329 (W 12.8984) | |
| lm_batch_likelihood_brackets_4_4d/n1/b1 | 27.0846 ± 0.660961 (W 0.0820374) | |
| lm_batch_likelihood_sentence_3_12d/n1/b1 | 72.9731 ± 0.277357 (W 0.0975121) | |
| lm_batch_likelihood_sentence_4_4d/n1/b1 | 27.6675 ± 0.277119 (W 0.0558962) | |
| str_matrix_chain_multiplication_100/n1/b1 | 12.6178 ± 0.0390761 (W 0.0113993) | |
| str_nw_mera_closed_120/n1/b1 | 1765.76 ± 3.42405 (W 1.23771) | |
| str_nw_mera_open_26/n1/b1 | 1168.92 ± 5.29866 (W 0.455519) | |
| tensornetwork_permutation_focus_step409_316/n1/b1 | 242.41 ± 1.16496 (W 0.0996259) | |
| tensornetwork_permutation_light_415/n1/b1 | 237.708 ± 1.24373 (W 0.446436) | |
| case | opt_flops (ms) | notes |
|---|---|
| bin_batched_matmul_b32_m64_n64_k64/n1/b1 | 0.144515 ± 6.36396e-05 (W 0.00140673) | |
| bin_elementwise_mul_2048x2048/n1/b1 | 2.13352 ± 0.0148839 (W 0.0168885) | |
| bin_matmul_256/n1/b1 | 0.304813 ± 0.00651033 (W 0.0035421) | |
| bin_outer_product_4096/n1/b1 | 4.19119 ± 0.0319499 (W 0.00470808) | |
| gm_queen5_5_3.wcsp/n1/b1 | 1796.21 ± 15.9377 (W 6.95991) | |
| lm_batch_likelihood_brackets_4_4d/n1/b1 | 11.5318 ± 0.0597357 (W 0.176863) | |
| lm_batch_likelihood_sentence_3_12d/n1/b1 | 22.9246 ± 0.708755 (W 0.918014) | |
| lm_batch_likelihood_sentence_4_4d/n1/b1 | 12.142 ± 0.186077 (W 0.150962) | |
| str_matrix_chain_multiplication_100/n1/b1 | 6.39764 ± 0.0951702 (W 0.160765) | |
| str_nw_mera_closed_120/n1/b1 | 517.026 ± 1.98619 (W 5.24076) | |
| str_nw_mera_open_26/n1/b1 | 319.059 ± 0.120604 (W 2.2174) | |
| tensornetwork_permutation_focus_step409_316/n1/b1 | 88.8127 ± 0.574572 (W 1.09257) | |
| tensornetwork_permutation_light_415/n1/b1 | 80.3926 ± 1.18363 (W 3.30447) | |
Repeated process sets describe observed repeatability under the recorded order schedule; they are not a paired baseline/candidate experiment and do not establish equal performance between engines.