per-shape on zen5-cpu

What was measured

Hardware

Every 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 1.47%, worst 15.77% among cells of 1 us or more (a sub-microsecond row is timer and scheduler jitter, not kernel time); 8 of 84 engine cells exceed 5% and carry !; median within-process relative SD 0.48%. A cell whose ±, ! or p is as large as the difference you are reading is not evidence of a regression. The vs plan column is the rotation-averaged ratio exp(mean over starts of log(engine median / plan median)), with the per-start ratios beside it: R > 1 means the engine was slower than that first-declared row in that start. It removes a row-position effect only under one position effect shared by the engines and stable across starts, and only where every declared engine has a measured row (* marks a cell where one has 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. The vs columns mark only differences larger than the spread this run recorded for those two rows (combined between-start SD, floored at 5%), always with the measured library as the subject: 🔵 tcontract X faster than <row> / 🔴 tcontract X slower than <row> (the compared row is named in the cell, not only in the column header), · inside the band, - that row is not measured in this arm. notes carries the cell warnings (* a declared row has no measurement, ! 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.

per-shape, f64, 1T (CPU 4)

case plan (ms) tblis (ms) vs tblis notes
ij_jk_ik_f64_n64/n1/b1 0.017392 ± 0.000240416 (W 1.9383e-05) 0.048315 ± 0.000362039 (W 0.000712029) 🔵 tcontract 2.78× faster than tblis dispatch-dominated (row under 20 us)
ij_jk_kl_il_n64/n1/b1 0.0345895 ± 0.000120915 (W 5.55944e-05) 0.096905 ± 0.000473762 (W 0.00117263) 🔵 tcontract 2.80× faster than tblis
ijk_jkl_il_8x16x8/n1/b1 0.000827 ± 7.07107e-06 (W 7.421e-06) 0.01611 ± 4.24264e-05 (W 6.48689e-05) 🔵 tcontract 19.48× faster than tblis dispatch-dominated (row under 20 us)
ikb_knb_inb_n16_b16/n1/b1 0.005961 ± 1.41421e-05 (W 9.88418e-06) 0.122082 ± 0.00116885 (W 0.00020941) 🔵 tcontract 20.48× faster than tblis dispatch-dominated (row under 20 us)
ikb_knb_inb_n16_b256/n1/b1 0.0995955 ± 0.000226981 (W 0.00189542) 1.90101 ± 0.0182921 (W 0.00151165) 🔵 tcontract 19.09× faster than tblis
ikb_knb_inb_n16_b64/n1/b1 0.0236895 ± 0.000135057 (W 0.00071655) 0.476469 ± 0.00529199 (W 0.00245011) 🔵 tcontract 20.11× faster than tblis
ikb_knb_inb_n2_b16/n1/b1 0.000706 ± 7.07107e-06 (W 2.17621e-05) 0.088945 ± 0.00128976 (W 0.000889009) 🔵 tcontract 125.98× faster than tblis dispatch-dominated (row under 20 us)
ikb_knb_inb_n2_b256/n1/b1 0.007995 ± 5.65685e-05 (W 2.59806e-05) 1.37765 ± 0.0354847 (W 0.00212261) 🔵 tcontract 172.31× faster than tblis dispatch-dominated (row under 20 us)
ikb_knb_inb_n2_b64/n1/b1 0.002194 ± 4.24264e-05 (W 0.000525887) 0.343272 ± 0.00543412 (W 0.00138063) 🔵 tcontract 156.46× faster than tblis dispatch-dominated (row under 20 us)
ikb_knb_inb_n4_b16/n1/b1 0.0009615 ± 1.48492e-05 (W 7.78054e-06) 0.090633 ± 0.00157968 (W 0.000117547) 🔵 tcontract 94.26× faster than tblis dispatch-dominated (row under 20 us)
ikb_knb_inb_n4_b256/n1/b1 0.012388 ± 0.000446891 (W 4.29094e-05) 1.41102 ± 0.0275149 (W 0.00193358) 🔵 tcontract 113.90× faster than tblis ! between-start range over 5%; dispatch-dominated (row under 20 us)
ikb_knb_inb_n4_b64/n1/b1 0.0032415 ± 9.2631e-05 (W 1.68657e-05) 0.353882 ± 0.00636891 (W 0.00150267) 🔵 tcontract 109.17× faster than tblis dispatch-dominated (row under 20 us)
ikb_knb_inb_n8_b16/n1/b1 0.0013525 ± 1.48492e-05 (W 1.13898e-05) 0.0981975 ± 0.00115471 (W 0.0017477) 🔵 tcontract 72.60× faster than tblis dispatch-dominated (row under 20 us)
ikb_knb_inb_n8_b256/n1/b1 0.0187245 ± 0.000255266 (W 4.62826e-05) 1.51725 ± 0.0276288 (W 0.00440746) 🔵 tcontract 81.03× faster than tblis dispatch-dominated (row under 20 us)
ikb_knb_inb_n8_b64/n1/b1 0.00523 ± 5.65685e-05 (W 1.94379e-05) 0.381018 ± 0.00393929 (W 0.001424) 🔵 tcontract 72.85× faster than tblis dispatch-dominated (row under 20 us)

per-shape, f64, 4T (CPU 4-7)

case plan (ms) tblis (ms) vs tblis notes
ij_jk_ik_f64_n64/n1/b1 0.0172925 ± 7.14178e-05 (W 0.000580027) 0.0345745 ± 0.00385444 (W 0.00231758) 🔵 tcontract 2.00× faster than tblis ! between-start range over 5%; p position effect over 5%; dispatch-dominated (row under 20 us)
ij_jk_kl_il_n64/n1/b1 0.034504 ± 4.24264e-05 (W 2.35634e-05) 0.0643145 ± 0.00118299 (W 0.00275097) 🔵 tcontract 1.86× faster than tblis
ijk_jkl_il_8x16x8/n1/b1 0.000827 ± 7.07107e-06 (W 4.52533e-06) 0.020593 ± 0.00033234 (W 0.000284946) 🔵 tcontract 24.90× faster than tblis dispatch-dominated (row under 20 us)
ikb_knb_inb_n16_b16/n1/b1 0.006131 ± 1.41421e-05 (W 8.83677e-06) 0.237856 ± 0.00889045 (W 0.0086884) 🔵 tcontract 38.80× faster than tblis ! between-start range over 5%; dispatch-dominated (row under 20 us)
ikb_knb_inb_n16_b256/n1/b1 0.030562 ± 3.53553e-05 (W 0.000319368) 1.23459 ± 0.0166616 (W 0.00836432) 🔵 tcontract 40.40× faster than tblis
ikb_knb_inb_n16_b64/n1/b1 0.0235535 ± 8.41457e-05 (W 3.83966e-05) 0.935297 ± 0.0180227 (W 0.0220441) 🔵 tcontract 39.71× faster than tblis
ikb_knb_inb_n2_b16/n1/b1 0.000711 ± 0 (W 1.04348e-05) 0.032796 ± 0.000432749 (W 0.00279103) 🔵 tcontract 46.13× faster than tblis dispatch-dominated (row under 20 us)
ikb_knb_inb_n2_b256/n1/b1 0.0082055 ± 0.000297692 (W 1.0423e-05) 0.397077 ± 0.00760847 (W 0.00677122) 🔵 tcontract 48.39× faster than tblis ! between-start range over 5%; dispatch-dominated (row under 20 us)
ikb_knb_inb_n2_b64/n1/b1 0.002204 ± 8.48528e-05 (W 2.93387e-05) 0.10516 ± 0.00213193 (W 0.00154837) 🔵 tcontract 47.71× faster than tblis ! between-start range over 5%; dispatch-dominated (row under 20 us)
ikb_knb_inb_n4_b16/n1/b1 0.0009765 ± 7.77817e-06 (W 7.83368e-06) 0.069344 ± 4.94975e-05 (W 0.00118229) 🔵 tcontract 71.01× faster than tblis dispatch-dominated (row under 20 us)
ikb_knb_inb_n4_b256/n1/b1 0.0125935 ± 8.55599e-05 (W 1.45335e-05) 0.408062 ± 0.00102743 (W 0.00566896) 🔵 tcontract 32.40× faster than tblis dispatch-dominated (row under 20 us)
ikb_knb_inb_n4_b64/n1/b1 0.003321 ± 7.07107e-06 (W 6.63209e-06) 0.109824 ± 0.00185686 (W 0.00266439) 🔵 tcontract 33.07× faster than tblis dispatch-dominated (row under 20 us)
ikb_knb_inb_n8_b16/n1/b1 0.0014125 ± 1.48492e-05 (W 1.11832e-05) 0.168493 ± 0.00276337 (W 0.00246216) 🔵 tcontract 119.29× faster than tblis dispatch-dominated (row under 20 us)
ikb_knb_inb_n8_b256/n1/b1 0.0193965 ± 0.000594677 (W 7.78794e-05) 0.427589 ± 0.00335098 (W 0.00686542) 🔵 tcontract 22.04× faster than tblis dispatch-dominated (row under 20 us)
ikb_knb_inb_n8_b64/n1/b1 0.004894 ± 6.36396e-05 (W 4.34835e-05) 0.283121 ± 0.00512228 (W 0.00900804) 🔵 tcontract 57.85× faster than tblis dispatch-dominated (row under 20 us)

per-shape, c64, 1T (CPU 4)

case plan (ms) tblis (ms) vs tblis notes
ij_jk_ik_c64_n32/n1/b1 0.0090165 ± 5.58614e-05 (W 1.14958e-05) 0.024796 ± 2.82843e-05 (W 6.05679e-05) 🔵 tcontract 2.75× faster than tblis dispatch-dominated (row under 20 us)
mps_chain_L32_chi16/n1/b1 0.16677 ± 7.07107e-07 (W 0.00149344) 0.902882 ± 0.00165109 (W 0.000922655) 🔵 tcontract 5.41× faster than tblis
mps_chain_L32_chi32/n1/b1 1.16337 ± 0.000226274 (W 0.00370002) 2.55319 ± 0.00716936 (W 0.0088201) 🔵 tcontract 2.19× faster than tblis
mps_chain_L32_chi4/n1/b1 0.031428 ± 0.000353553 (W 0.000315529) 0.523002 ± 0.00109107 (W 0.00140626) 🔵 tcontract 16.64× faster than tblis
mps_chain_L32_chi64/n1/b1 10.894 ± 0.00289843 (W 0.0152492) 13.7744 ± 0.0833757 (W 0.0606757) 🔵 tcontract 1.26× faster than tblis
mps_chain_L32_chi8/n1/b1 0.060502 ± 0.000581242 (W 0.000773789) 0.59756 ± 0.00423628 (W 0.00192466) 🔵 tcontract 9.88× faster than tblis

per-shape, c64, 4T (CPU 4-7)

case plan (ms) tblis (ms) vs tblis notes
ij_jk_ik_c64_n32/n1/b1 0.0089965 ± 4.31335e-05 (W 2.03568e-05) 0.023704 ± 0.000666095 (W 0.0019987) 🔵 tcontract 2.63× faster than tblis dispatch-dominated (row under 20 us)
mps_chain_L32_chi16/n1/b1 0.168502 ± 0.000523966 (W 0.00218172) 1.36423 ± 0.0557448 (W 0.0139513) 🔵 tcontract 8.10× faster than tblis ! between-start range over 5%
mps_chain_L32_chi32/n1/b1 1.1678 ± 0.00648275 (W 0.00241947) 1.91434 ± 0.04721 (W 0.0222545) 🔵 tcontract 1.64× faster than tblis
mps_chain_L32_chi4/n1/b1 0.032049 ± 0.000155563 (W 0.000311519) 0.957202 ± 0.0513112 (W 0.00253892) 🔵 tcontract 29.87× faster than tblis ! between-start range over 5%
mps_chain_L32_chi64/n1/b1 4.05505 ± 0.106823 (W 0.14407) 5.56742 ± 0.0614476 (W 0.0175714) 🔵 tcontract 1.37× faster than tblis
mps_chain_L32_chi8/n1/b1 0.0610185 ± 0.000446184 (W 0.000216776) 1.2677 ± 0.0499868 (W 0.0231547) 🔵 tcontract 20.78× faster than tblis ! between-start range over 5%

Not applicable to this population's shapes and reported as skipped: ij_jk_ik_c64_n32/1, ij_jk_ik_f64_n64/1, ij_jk_kl_il_n64/1, ijk_jkl_il_8x16x8/1, ikb_knb_inb_n16_b16/1, ikb_knb_inb_n16_b256/1, ikb_knb_inb_n16_b64/1, ikb_knb_inb_n2_b16/1, ikb_knb_inb_n2_b256/1, ikb_knb_inb_n2_b64/1, ikb_knb_inb_n4_b16/1, ikb_knb_inb_n4_b256/1, ikb_knb_inb_n4_b64/1, ikb_knb_inb_n8_b16/1, ikb_knb_inb_n8_b256/1, ikb_knb_inb_n8_b64/1, mps_chain_L32_chi16/1, mps_chain_L32_chi32/1, mps_chain_L32_chi4/1, mps_chain_L32_chi64/1, mps_chain_L32_chi8/1.

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.