[tenancy before] (none)

  shape  geometry                        thr  regs    smem blk/SM warps/SM   bank
  s24    128x128x128 2x4 st2 sw8         256   123   49152      2       16 64
  s27    128x128x128 1x4 st2 sw8         128   199   49152      2        8 128
  s80    128x128x128 2x4 st2 sw8 gs128   256   222   49152      1        8 64x2
  s83    128x128x128 2x8 st2 sw8 gs128   512   128   49152      1       16 32x2
  s84    128x128x128 2x8 st2 sw8         512    64   49152      2       32 32   <-- SPILLS
  s85    64x128x128 1x8 st2 sw8 gs128    256   128   32768      2       16 32x2
  s86    64x128x128 1x8 st2 sw8          256    78   32768      3       24 32
  s89    64x128x128 1x8 st2 sw8 gs256    256   126   32768      2       16 32x2
  s90    64x128x128 2x4 st2 sw8 gs128    256   128   32768      2       16 32x2
  s91    64x128x128 2x4 st2 sw8          256    80   32768      3       24 32
  s92    64x128x128 2x8 st2 sw8 gs128    512    98   32768      1       16 16x2
  s93    64x128x128 2x8 st2 sw8          512    64   32768      2       32 16
  s94    32x128x128 1x8 st2 sw8 gs128    256    99   24576      2       16 16x2
  s95    32x128x128 1x8 st2 sw8          256    64   24576      4       32 16
[l2 scrub] mode='read' -- see the module docstring for the bias it carries

  out_proj k=6144 n=5120 M=48   burst>= 50 ms, reps=7
    f24       100.54 us/call  inner=452    burst=   45.4 ms
    f27        96.02 us/call  inner=471    burst=   45.2 ms
    f80       116.43 us/call  inner=379    burst=   44.1 ms
    f83       106.46 us/call  inner=419    burst=   44.6 ms
    f84       101.88 us/call  inner=448    burst=   45.6 ms
    f85        77.51 us/call  inner=566    burst=   43.9 ms
    f86        68.93 us/call  inner=655    burst=   45.1 ms
    f89        71.17 us/call  inner=602    burst=   42.8 ms
    f90        85.24 us/call  inner=513    burst=   43.7 ms
    f91        81.78 us/call  inner=552    burst=   45.1 ms
    f92        84.07 us/call  inner=512    burst=   43.0 ms
    f93        78.35 us/call  inner=576    burst=   45.1 ms
    f94        66.60 us/call  inner=648    burst=   43.2 ms
    f95        59.55 us/call  inner=738    burst=   43.9 ms
    g24        73.10 us/call  inner=632    burst=   46.2 ms
    g27        68.15 us/call  inner=666    burst=   45.4 ms
    g80        88.87 us/call  inner=519    burst=   46.1 ms
    g83        79.11 us/call  inner=584    burst=   46.2 ms
    g84        74.55 us/call  inner=620    burst=   46.2 ms
    g85        49.65 us/call  inner=903    burst=   44.8 ms
    g86        41.71 us/call  inner=1098   burst=   45.8 ms
    g89        44.58 us/call  inner=1020   burst=   45.5 ms
    g90        58.12 us/call  inner=778    burst=   45.2 ms
    g91        54.80 us/call  inner=837    burst=   45.9 ms
    g92        57.44 us/call  inner=798    burst=   45.8 ms
    g93        51.24 us/call  inner=898    burst=   46.0 ms
    g94        39.59 us/call  inner=1123   burst=   44.5 ms
    g95        32.69 us/call  inner=1386   burst=   45.3 ms
    legB      118.12 us/call  inner=404    burst=   47.7 ms
    per-arm ratios, paired within a rep:
      s27  per-token                                         gemm/s27 1.0000x  chain/s27 1.0000x
      s24  per-token                                         gemm/s27 1.0727x  chain/s27 1.0474x
      s80  gs128      bank vs s24  1.2158x [1.2134-1.2173]  gemm/s27 1.3041x  chain/s27 1.2138x
      s84  per-token                                         gemm/s27 1.0939x  chain/s27 1.0612x
      s83  gs128      bank vs s84  1.0614x [1.0600-1.0652]  gemm/s27 1.1620x  chain/s27 1.1064x
      s86  per-token                                         gemm/s27 0.6120x  chain/s27 0.7165x
      s85  gs128      bank vs s86  1.1904x [1.1062-1.1923]  gemm/s27 0.7289x  chain/s27 0.8051x
      s91  per-token                                         gemm/s27 0.8050x  chain/s27 0.8517x
      s90  gs128      bank vs s91  1.0600x [1.0272-1.0749]  gemm/s27 0.8527x  chain/s27 0.8878x
      s93  per-token                                         gemm/s27 0.7528x  chain/s27 0.8163x
      s92  gs128      bank vs s93  1.1211x [1.1196-1.1213]  gemm/s27 0.8430x  chain/s27 0.8740x
      s95  per-token                                         gemm/s27 0.4797x  chain/s27 0.6208x
      s94  gs128      bank vs s95  1.2112x [1.1477-1.2118]  gemm/s27 0.5810x  chain/s27 0.6940x
      s89  gs256      bank vs s86  1.0687x [0.9933-1.0950]  gemm/s27 0.6548x  chain/s27 0.7431x
      BEST grouped s94 vs BEST per-token s95: gemm 1.2112x [1.1477-1.2118]  chain 1.1175x [1.1074-1.1220]

  out_proj k=6144 n=5120 M=64   burst>= 50 ms, reps=7
    f24       101.46 us/call  inner=479    burst=   48.6 ms
    f27        96.88 us/call  inner=503    burst=   48.7 ms
    f80       116.33 us/call  inner=408    burst=   47.5 ms
    f83       106.52 us/call  inner=445    burst=   47.4 ms
    f84       102.71 us/call  inner=472    burst=   48.5 ms
    f85        74.02 us/call  inner=622    burst=   46.0 ms
    f86        69.34 us/call  inner=685    burst=   47.5 ms
    f89        68.31 us/call  inner=568    burst=   38.8 ms
    f90        81.82 us/call  inner=574    burst=   47.0 ms
    f91        82.07 us/call  inner=581    burst=   47.7 ms
    f92        81.13 us/call  inner=581    burst=   47.1 ms
    f93        78.52 us/call  inner=597    burst=   46.9 ms
    f94        63.05 us/call  inner=676    burst=   42.6 ms
    f95        59.87 us/call  inner=774    burst=   46.3 ms
    g24        73.52 us/call  inner=674    burst=   49.6 ms
    g27        68.61 us/call  inner=722    burst=   49.5 ms
    g80        89.01 us/call  inner=557    burst=   49.6 ms
    g83        79.28 us/call  inner=626    burst=   49.6 ms
    g84        74.81 us/call  inner=662    burst=   49.5 ms
    g85        50.23 us/call  inner=933    burst=   46.9 ms
    g86        42.02 us/call  inner=1177   burst=   49.5 ms
    g89        44.83 us/call  inner=1077   burst=   48.3 ms
    g90        58.18 us/call  inner=835    burst=   48.6 ms
    g91        54.88 us/call  inner=894    burst=   49.1 ms
    g92        57.71 us/call  inner=842    burst=   48.6 ms
    g93        51.17 us/call  inner=963    burst=   49.3 ms
    g94        39.59 us/call  inner=1195   burst=   47.3 ms
    g95        32.69 us/call  inner=1491   burst=   48.7 ms
    legB      120.93 us/call  inner=417    burst=   50.4 ms
    per-arm ratios, paired within a rep:
      s27  per-token                                         gemm/s27 1.0000x  chain/s27 1.0000x
      s24  per-token                                         gemm/s27 1.0723x  chain/s27 1.0474x
      s80  gs128      bank vs s24  1.2099x [1.2077-1.2852]  gemm/s27 1.2970x  chain/s27 1.2003x
      s84  per-token                                         gemm/s27 1.0912x  chain/s27 1.0601x
      s83  gs128      bank vs s84  1.0596x [1.0583-1.0598]  gemm/s27 1.1563x  chain/s27 1.0996x
      s86  per-token                                         gemm/s27 0.6127x  chain/s27 0.7157x
      s85  gs128      bank vs s86  1.1956x [1.1951-1.1959]  gemm/s27 0.7325x  chain/s27 0.7636x
      s91  per-token                                         gemm/s27 0.8002x  chain/s27 0.8470x
      s90  gs128      bank vs s91  1.0597x [1.0593-1.0857]  gemm/s27 0.8481x  chain/s27 0.8444x
      s93  per-token                                         gemm/s27 0.7462x  chain/s27 0.8105x
      s92  gs128      bank vs s93  1.1271x [1.1268-1.1401]  gemm/s27 0.8412x  chain/s27 0.8375x
      s95  per-token                                         gemm/s27 0.4769x  chain/s27 0.6177x
      s94  gs128      bank vs s95  1.2101x [1.2096-1.2355]  gemm/s27 0.5771x  chain/s27 0.6503x
      s89  gs256      bank vs s86  1.0663x [1.0651-1.0836]  gemm/s27 0.6533x  chain/s27 0.7049x
      BEST grouped s94 vs BEST per-token s95: gemm 1.2101x [1.2096-1.2355]  chain 1.0532x [1.0504-1.0756]

  out_proj k=6144 n=5120 M=2048   burst>= 50 ms, reps=7
    f24       500.13 us/call  inner=103    burst=   51.5 ms
    f27       460.35 us/call  inner=114    burst=   52.5 ms
    f80       697.02 us/call  inner=77     burst=   53.7 ms
    f83       623.64 us/call  inner=83     burst=   51.8 ms
    f84       543.97 us/call  inner=97     burst=   52.8 ms
    f85       602.89 us/call  inner=86     burst=   51.8 ms
    f86       521.66 us/call  inner=99     burst=   51.6 ms
    f89       588.27 us/call  inner=89     burst=   52.4 ms
    f90       739.25 us/call  inner=71     burst=   52.5 ms
    f91       659.17 us/call  inner=79     burst=   52.1 ms
    f92       809.52 us/call  inner=62     burst=   50.2 ms
    f93       666.74 us/call  inner=77     burst=   51.3 ms
    f94       818.79 us/call  inner=63     burst=   51.6 ms
    f95       700.84 us/call  inner=74     burst=   51.9 ms
    g24       351.09 us/call  inner=148    burst=   52.0 ms
    g27       315.11 us/call  inner=170    burst=   53.6 ms
    g80       520.47 us/call  inner=111    burst=   57.8 ms
    g83       463.38 us/call  inner=115    burst=   53.3 ms
    g84       381.72 us/call  inner=138    burst=   52.7 ms
    g85       442.99 us/call  inner=121    burst=   53.6 ms
    g86       366.74 us/call  inner=140    burst=   51.3 ms
    g89       412.54 us/call  inner=125    burst=   51.6 ms
    g90       550.48 us/call  inner=93     burst=   51.2 ms
    g91       505.66 us/call  inner=104    burst=   52.6 ms
    g92       640.22 us/call  inner=78     burst=   49.9 ms
    g93       517.58 us/call  inner=100    burst=   51.8 ms
    g94       637.91 us/call  inner=81     burst=   51.7 ms
    g95       546.98 us/call  inner=95     burst=   52.0 ms
    legB      635.90 us/call  inner=81     burst=   51.5 ms
    per-arm ratios, paired within a rep:
      s27  per-token                                         gemm/s27 1.0000x  chain/s27 1.0000x
      s24  per-token                                         gemm/s27 1.1304x  chain/s27 1.0969x
      s80  gs128      bank vs s24  1.4635x [1.3292-1.5353]  gemm/s27 1.6264x  chain/s27 1.4717x
      s84  per-token                                         gemm/s27 1.2114x  chain/s27 1.1587x
      s83  gs128      bank vs s84  1.2065x [1.1276-1.2674]  gemm/s27 1.4649x  chain/s27 1.3661x
      s86  per-token                                         gemm/s27 1.1596x  chain/s27 1.1351x
      s85  gs128      bank vs s86  1.2199x [1.1392-1.2849]  gemm/s27 1.4210x  chain/s27 1.3055x
      s91  per-token                                         gemm/s27 1.6152x  chain/s27 1.4350x
      s90  gs128      bank vs s91  1.0823x [1.0315-1.1542]  gemm/s27 1.7560x  chain/s27 1.5469x
      s93  per-token                                         gemm/s27 1.6425x  chain/s27 1.4529x
      s92  gs128      bank vs s93  1.2262x [1.1746-1.3440]  gemm/s27 2.0431x  chain/s27 1.7585x
      s95  per-token                                         gemm/s27 1.7537x  chain/s27 1.5428x
      s94  gs128      bank vs s95  1.1509x [1.0860-1.2228]  gemm/s27 2.0312x  chain/s27 1.7937x
      s89  gs256      bank vs s86  1.1098x [1.0749-1.1629]  gemm/s27 1.3001x  chain/s27 1.2497x
      BEST grouped s89 vs BEST per-token s27: gemm 1.3001x [1.2771-1.3555]  chain 1.2497x [1.1377-1.2991]

  gate_proj k=5120 n=17408 M=48   burst>= 50 ms, reps=7
    f24       134.11 us/call  inner=365    burst=   48.9 ms
    f27       123.90 us/call  inner=392    burst=   48.6 ms
    f80       177.47 us/call  inner=273    burst=   48.4 ms
    f83       162.36 us/call  inner=293    burst=   47.6 ms
    f84       141.02 us/call  inner=342    burst=   48.2 ms
    f85        94.53 us/call  inner=497    burst=   47.0 ms
    f86        88.44 us/call  inner=562    burst=   49.7 ms
    f89        91.35 us/call  inner=514    burst=   47.0 ms
    f90       113.31 us/call  inner=416    burst=   47.1 ms
    f91       108.40 us/call  inner=450    burst=   48.8 ms
    f92       125.74 us/call  inner=379    burst=   47.7 ms
    f93       107.82 us/call  inner=453    burst=   48.8 ms
    f94       106.68 us/call  inner=442    burst=   47.2 ms
    f95        92.88 us/call  inner=524    burst=   48.7 ms
    g24       106.12 us/call  inner=469    burst=   49.8 ms
    g27        95.72 us/call  inner=521    burst=   49.9 ms
    g80       149.39 us/call  inner=336    burst=   50.2 ms
    g83       134.16 us/call  inner=373    burst=   50.0 ms
    g84       113.21 us/call  inner=441    burst=   49.9 ms
    g85        66.63 us/call  inner=727    burst=   48.4 ms
    g86        58.55 us/call  inner=851    burst=   49.8 ms
    g89        63.30 us/call  inner=788    burst=   49.9 ms
    g90        85.03 us/call  inner=586    burst=   49.8 ms
    g91        80.52 us/call  inner=619    burst=   49.8 ms
    g92        97.66 us/call  inner=499    burst=   48.7 ms
    g93        80.40 us/call  inner=622    burst=   50.0 ms
    g94        82.39 us/call  inner=601    burst=   49.5 ms
    g95        70.52 us/call  inner=709    burst=   50.0 ms
    legB      484.01 us/call  inner=104    burst=   50.3 ms
    per-arm ratios, paired within a rep:
      s27  per-token                                         gemm/s27 1.0000x  chain/s27 1.0000x
      s24  per-token                                         gemm/s27 1.1088x  chain/s27 1.0827x
      s80  gs128      bank vs s24  1.4076x [1.3823-1.4086]  gemm/s27 1.5607x  chain/s27 1.4328x
      s84  per-token                                         gemm/s27 1.1826x  chain/s27 1.1384x
      s83  gs128      bank vs s84  1.1851x [1.1685-1.1894]  gemm/s27 1.4016x  chain/s27 1.3100x
      s86  per-token                                         gemm/s27 0.6116x  chain/s27 0.7138x
      s85  gs128      bank vs s86  1.1394x [1.1331-1.1448]  gemm/s27 0.6961x  chain/s27 0.7632x
      s91  per-token                                         gemm/s27 0.8413x  chain/s27 0.8747x
      s90  gs128      bank vs s91  1.0561x [1.0546-1.0997]  gemm/s27 0.8884x  chain/s27 0.9145x
      s93  per-token                                         gemm/s27 0.8396x  chain/s27 0.8703x
      s92  gs128      bank vs s93  1.2161x [1.2103-1.2255]  gemm/s27 1.0203x  chain/s27 1.0152x
      s95  per-token                                         gemm/s27 0.7366x  chain/s27 0.7498x
      s94  gs128      bank vs s95  1.1680x [1.1576-1.1763]  gemm/s27 0.8600x  chain/s27 0.8608x
      s89  gs256      bank vs s86  1.0810x [1.0749-1.1265]  gemm/s27 0.6615x  chain/s27 0.7369x
      BEST grouped s89 vs BEST per-token s86: gemm 1.0810x [1.0749-1.1265]  chain 1.0371x [0.9916-1.0681]

  gate_proj k=5120 n=17408 M=64   burst>= 50 ms, reps=7
    f24       135.96 us/call  inner=361    burst=   49.1 ms
    f27       125.43 us/call  inner=391    burst=   49.0 ms
    f80       179.69 us/call  inner=270    burst=   48.5 ms
    f83       164.69 us/call  inner=295    burst=   48.6 ms
    f84       143.01 us/call  inner=343    burst=   49.1 ms
    f85        93.27 us/call  inner=509    burst=   47.5 ms
    f86        87.70 us/call  inner=556    burst=   48.8 ms
    f89        90.40 us/call  inner=528    burst=   47.7 ms
    f90       112.11 us/call  inner=426    burst=   47.8 ms
    f91       109.76 us/call  inner=445    burst=   48.8 ms
    f92       125.03 us/call  inner=386    burst=   48.3 ms
    f93       109.92 us/call  inner=446    burst=   49.0 ms
    f94       105.87 us/call  inner=452    burst=   47.9 ms
    f95        94.93 us/call  inner=514    burst=   48.8 ms
    g24       106.43 us/call  inner=466    burst=   49.6 ms
    g27        95.70 us/call  inner=521    burst=   49.9 ms
    g80       149.94 us/call  inner=332    burst=   49.8 ms
    g83       134.44 us/call  inner=367    burst=   49.3 ms
    g84       114.41 us/call  inner=435    burst=   49.8 ms
    g85        67.23 us/call  inner=710    burst=   47.7 ms
    g86        59.36 us/call  inner=834    burst=   49.5 ms
    g89        64.19 us/call  inner=777    burst=   49.9 ms
    g90        85.71 us/call  inner=578    burst=   49.5 ms
    g91        81.14 us/call  inner=612    burst=   49.7 ms
    g92        98.41 us/call  inner=494    burst=   48.6 ms
    g93        81.33 us/call  inner=609    burst=   49.5 ms
    g94        83.36 us/call  inner=592    burst=   49.3 ms
    g95        71.05 us/call  inner=704    burst=   50.0 ms
    legB      485.11 us/call  inner=104    burst=   50.5 ms
    per-arm ratios, paired within a rep:
      s27  per-token                                         gemm/s27 1.0000x  chain/s27 1.0000x
      s24  per-token                                         gemm/s27 1.1125x  chain/s27 1.0847x
      s80  gs128      bank vs s24  1.4089x [1.3144-1.4230]  gemm/s27 1.5674x  chain/s27 1.4361x
      s84  per-token                                         gemm/s27 1.1952x  chain/s27 1.1433x
      s83  gs128      bank vs s84  1.1750x [1.1745-1.1883]  gemm/s27 1.4053x  chain/s27 1.3168x
      s86  per-token                                         gemm/s27 0.6184x  chain/s27 0.7007x
      s85  gs128      bank vs s86  1.1334x [1.0895-1.1400]  gemm/s27 0.7024x  chain/s27 0.7451x
      s91  per-token                                         gemm/s27 0.8476x  chain/s27 0.8770x
      s90  gs128      bank vs s91  1.0573x [1.0538-1.1015]  gemm/s27 0.8956x  chain/s27 0.8958x
      s93  per-token                                         gemm/s27 0.8496x  chain/s27 0.8736x
      s92  gs128      bank vs s93  1.2099x [1.1330-1.2174]  gemm/s27 1.0282x  chain/s27 0.9993x
      s95  per-token                                         gemm/s27 0.7404x  chain/s27 0.7590x
      s94  gs128      bank vs s95  1.1737x [1.1667-1.1790]  gemm/s27 0.8697x  chain/s27 0.8452x
      s89  gs256      bank vs s86  1.0829x [1.0429-1.0913]  gemm/s27 0.6709x  chain/s27 0.7200x
      BEST grouped s89 vs BEST per-token s86: gemm 1.0829x [1.0429-1.0913]  chain 1.0276x [0.9591-1.1081]

  gate_proj k=5120 n=17408 M=2048   burst>= 50 ms, reps=7
    f24      1239.82 us/call  inner=42     burst=   52.1 ms
    f27      1124.48 us/call  inner=47     burst=   52.9 ms
    f80      1682.63 us/call  inner=30     burst=   50.5 ms
    f83      1656.12 us/call  inner=32     burst=   53.0 ms
    f84      1368.90 us/call  inner=39     burst=   53.4 ms
    f85      1522.03 us/call  inner=33     burst=   50.2 ms
    f86      1295.51 us/call  inner=40     burst=   51.8 ms
    f89      1463.16 us/call  inner=36     burst=   52.7 ms
    f90      1872.73 us/call  inner=28     burst=   52.4 ms
    f91      1692.40 us/call  inner=30     burst=   50.8 ms
    f92      2181.67 us/call  inner=24     burst=   52.4 ms
    f93      1777.60 us/call  inner=30     burst=   53.3 ms
    f94      2095.73 us/call  inner=24     burst=   50.3 ms
    f95      1814.75 us/call  inner=28     burst=   50.8 ms
    g24       992.80 us/call  inner=53     burst=   52.6 ms
    g27       903.83 us/call  inner=61     burst=   55.1 ms
    g80      1470.04 us/call  inner=34     burst=   50.0 ms
    g83      1358.20 us/call  inner=38     burst=   51.6 ms
    g84      1082.58 us/call  inner=48     burst=   52.0 ms
    g85      1266.09 us/call  inner=41     burst=   51.9 ms
    g86      1091.38 us/call  inner=50     burst=   54.6 ms
    g89      1216.58 us/call  inner=44     burst=   53.5 ms
    g90      1545.48 us/call  inner=31     burst=   47.9 ms
    g91      1439.63 us/call  inner=36     burst=   51.8 ms
    g92      1890.49 us/call  inner=27     burst=   51.0 ms
    g93      1511.74 us/call  inner=35     burst=   52.9 ms
    g94      1897.20 us/call  inner=27     burst=   51.2 ms
    g95      1549.60 us/call  inner=33     burst=   51.1 ms
    legB     1808.49 us/call  inner=29     burst=   52.4 ms
    per-arm ratios, paired within a rep:
      s27  per-token                                         gemm/s27 1.0000x  chain/s27 1.0000x
      s24  per-token                                         gemm/s27 1.0978x  chain/s27 1.0983x
      s80  gs128      bank vs s24  1.4807x [1.3888-1.5250]  gemm/s27 1.6439x  chain/s27 1.4752x
      s84  per-token                                         gemm/s27 1.1927x  chain/s27 1.2247x
      s83  gs128      bank vs s84  1.2701x [1.1087-1.3714]  gemm/s27 1.4947x  chain/s27 1.4610x
      s86  per-token                                         gemm/s27 1.1763x  chain/s27 1.1579x
      s85  gs128      bank vs s86  1.1601x [1.1449-1.2447]  gemm/s27 1.3996x  chain/s27 1.3395x
      s91  per-token                                         gemm/s27 1.5821x  chain/s27 1.4833x
      s90  gs128      bank vs s91  1.0765x [1.0163-1.1547]  gemm/s27 1.7076x  chain/s27 1.6701x
      s93  per-token                                         gemm/s27 1.6679x  chain/s27 1.5859x
      s92  gs128      bank vs s93  1.2386x [1.1612-1.3182]  gemm/s27 2.0660x  chain/s27 1.8805x
      s95  per-token                                         gemm/s27 1.7219x  chain/s27 1.6061x
      s94  gs128      bank vs s95  1.1947x [1.1370-1.3396]  gemm/s27 2.0388x  chain/s27 1.8307x
      s89  gs256      bank vs s86  1.1079x [1.0561-1.1686]  gemm/s27 1.3450x  chain/s27 1.2724x
      BEST grouped s89 vs BEST per-token s27: gemm 1.3450x [1.2081-1.4286]  chain 1.2724x [1.1770-1.3580]

  down_proj k=17408 n=5120 M=48   burst>= 50 ms, reps=7
    f24       228.97 us/call  inner=216    burst=   49.5 ms
    f27       215.18 us/call  inner=229    burst=   49.3 ms
    f80       280.56 us/call  inner=175    burst=   49.1 ms
    f83       252.75 us/call  inner=194    burst=   49.0 ms
    f84       232.97 us/call  inner=210    burst=   48.9 ms
    f85       167.97 us/call  inner=288    burst=   48.4 ms
    f86       138.44 us/call  inner=352    burst=   48.7 ms
    f89       153.23 us/call  inner=315    burst=   48.3 ms
    f90       192.56 us/call  inner=254    burst=   48.9 ms
    f91       176.81 us/call  inner=278    burst=   49.2 ms
    f92       190.52 us/call  inner=256    burst=   48.8 ms
    f93       166.12 us/call  inner=295    burst=   49.0 ms
    f94       139.00 us/call  inner=347    burst=   48.2 ms
    f95       113.62 us/call  inner=425    burst=   48.3 ms
    g24       202.45 us/call  inner=247    burst=   50.0 ms
    g27       188.21 us/call  inner=265    burst=   49.9 ms
    g80       248.22 us/call  inner=202    burst=   50.1 ms
    g83       220.11 us/call  inner=227    burst=   50.0 ms
    g84       206.37 us/call  inner=242    burst=   49.9 ms
    g85       137.06 us/call  inner=354    burst=   48.5 ms
    g86       113.55 us/call  inner=440    burst=   50.0 ms
    g89       123.36 us/call  inner=400    burst=   49.3 ms
    g90       161.17 us/call  inner=304    burst=   49.0 ms
    g91       151.67 us/call  inner=329    burst=   49.9 ms
    g92       159.39 us/call  inner=311    burst=   49.6 ms
    g93       141.06 us/call  inner=355    burst=   50.1 ms
    g94       108.25 us/call  inner=446    burst=   48.3 ms
    g95        88.76 us/call  inner=561    burst=   49.8 ms
    legB      489.16 us/call  inner=103    burst=   50.4 ms
    per-arm ratios, paired within a rep:
      s27  per-token                                         gemm/s27 1.0000x  chain/s27 1.0000x
      s24  per-token                                         gemm/s27 1.0757x  chain/s27 1.0642x
      s80  gs128      bank vs s24  1.2258x [1.2249-1.2310]  gemm/s27 1.3188x  chain/s27 1.3036x
      s84  per-token                                         gemm/s27 1.0965x  chain/s27 1.0827x
      s83  gs128      bank vs s84  1.0666x [1.0645-1.0695]  gemm/s27 1.1695x  chain/s27 1.1744x
      s86  per-token                                         gemm/s27 0.6034x  chain/s27 0.6432x
      s85  gs128      bank vs s86  1.2070x [1.2067-1.2071]  gemm/s27 0.7282x  chain/s27 0.7812x
      s91  per-token                                         gemm/s27 0.8059x  chain/s27 0.8223x
      s90  gs128      bank vs s91  1.0626x [1.0622-1.0628]  gemm/s27 0.8564x  chain/s27 0.8950x
      s93  per-token                                         gemm/s27 0.7495x  chain/s27 0.7722x
      s92  gs128      bank vs s93  1.1299x [1.1297-1.1305]  gemm/s27 0.8469x  chain/s27 0.8840x
      s95  per-token                                         gemm/s27 0.4716x  chain/s27 0.5279x
      s94  gs128      bank vs s95  1.2199x [1.2156-1.2212]  gemm/s27 0.5751x  chain/s27 0.6461x
      s89  gs256      bank vs s86  1.0864x [1.0861-1.0889]  gemm/s27 0.6555x  chain/s27 0.7123x
      BEST grouped s94 vs BEST per-token s95: gemm 1.2199x [1.2156-1.2212]  chain 1.2229x [1.2214-1.2452]

  down_proj k=17408 n=5120 M=64   burst>= 50 ms, reps=7
    f24       230.56 us/call  inner=215    burst=   49.6 ms
    f27       216.63 us/call  inner=229    burst=   49.6 ms
    f80       282.54 us/call  inner=174    burst=   49.2 ms
    f83       254.07 us/call  inner=193    burst=   49.0 ms
    f84       234.47 us/call  inner=211    burst=   49.5 ms
    f85       168.89 us/call  inner=289    burst=   48.8 ms
    f86       140.68 us/call  inner=349    burst=   49.1 ms
    f89       152.88 us/call  inner=317    burst=   48.5 ms
    f90       190.22 us/call  inner=258    burst=   49.1 ms
    f91       178.29 us/call  inner=277    burst=   49.4 ms
    f92       189.29 us/call  inner=260    burst=   49.2 ms
    f93       167.76 us/call  inner=293    burst=   49.2 ms
    f94       138.56 us/call  inner=354    burst=   49.1 ms
    f95       115.20 us/call  inner=423    burst=   48.7 ms
    g24       202.81 us/call  inner=246    burst=   49.9 ms
    g27       188.56 us/call  inner=265    burst=   50.0 ms
    g80       248.41 us/call  inner=201    burst=   49.9 ms
    g83       220.27 us/call  inner=227    burst=   50.0 ms
    g84       206.76 us/call  inner=242    burst=   50.0 ms
    g85       138.63 us/call  inner=343    burst=   47.5 ms
    g86       113.93 us/call  inner=436    burst=   49.7 ms
    g89       123.62 us/call  inner=400    burst=   49.4 ms
    g90       160.80 us/call  inner=304    burst=   48.9 ms
    g91       151.71 us/call  inner=329    burst=   49.9 ms
    g92       159.65 us/call  inner=309    burst=   49.3 ms
    g93       141.08 us/call  inner=354    burst=   49.9 ms
    g94       108.31 us/call  inner=445    burst=   48.2 ms
    g95        89.01 us/call  inner=560    burst=   49.8 ms
    legB      489.07 us/call  inner=103    burst=   50.4 ms
    per-arm ratios, paired within a rep:
      s27  per-token                                         gemm/s27 1.0000x  chain/s27 1.0000x
      s24  per-token                                         gemm/s27 1.0756x  chain/s27 1.0645x
      s80  gs128      bank vs s24  1.2248x [1.2247-1.2250]  gemm/s27 1.3174x  chain/s27 1.3045x
      s84  per-token                                         gemm/s27 1.0966x  chain/s27 1.0825x
      s83  gs128      bank vs s84  1.0653x [1.0646-1.0656]  gemm/s27 1.1681x  chain/s27 1.1730x
      s86  per-token                                         gemm/s27 0.6042x  chain/s27 0.6496x
      s85  gs128      bank vs s86  1.2168x [1.2155-1.2171]  gemm/s27 0.7352x  chain/s27 0.7798x
      s91  per-token                                         gemm/s27 0.8046x  chain/s27 0.8232x
      s90  gs128      bank vs s91  1.0599x [1.0598-1.0610]  gemm/s27 0.8528x  chain/s27 0.8782x
      s93  per-token                                         gemm/s27 0.7482x  chain/s27 0.7746x
      s92  gs128      bank vs s93  1.1317x [1.1314-1.1319]  gemm/s27 0.8467x  chain/s27 0.8738x
      s95  per-token                                         gemm/s27 0.4721x  chain/s27 0.5318x
      s94  gs128      bank vs s95  1.2169x [1.2160-1.2184]  gemm/s27 0.5744x  chain/s27 0.6398x
      s89  gs256      bank vs s86  1.0850x [1.0830-1.0895]  gemm/s27 0.6556x  chain/s27 0.7059x
      BEST grouped s94 vs BEST per-token s95: gemm 1.2169x [1.2160-1.2184]  chain 1.2017x [1.1952-1.2089]

  down_proj k=17408 n=5120 M=2048   burst>= 50 ms, reps=7
    f24      1493.22 us/call  inner=35     burst=   52.3 ms
    f27      1389.65 us/call  inner=38     burst=   52.8 ms
    f80      1945.87 us/call  inner=27     burst=   52.5 ms
    f83      1808.49 us/call  inner=29     burst=   52.4 ms
    f84      1580.12 us/call  inner=33     burst=   52.1 ms
    f85      1780.38 us/call  inner=29     burst=   51.6 ms
    f86      1532.77 us/call  inner=33     burst=   50.6 ms
    f89      1661.56 us/call  inner=31     burst=   51.5 ms
    f90      2103.66 us/call  inner=25     burst=   52.6 ms
    f91      1921.69 us/call  inner=26     burst=   50.0 ms
    f92      2326.95 us/call  inner=22     burst=   51.2 ms
    f93      1949.11 us/call  inner=26     burst=   50.7 ms
    f94      2359.25 us/call  inner=22     burst=   51.9 ms
    f95      2025.27 us/call  inner=25     burst=   50.6 ms
    g24       988.33 us/call  inner=52     burst=   51.4 ms
    g27       899.64 us/call  inner=61     burst=   54.9 ms
    g80      1401.16 us/call  inner=39     burst=   54.6 ms
    g83      1335.95 us/call  inner=42     burst=   56.1 ms
    g84      1072.57 us/call  inner=49     burst=   52.6 ms
    g85      1219.10 us/call  inner=42     burst=   51.2 ms
    g86      1036.76 us/call  inner=50     burst=   51.8 ms
    g89      1194.71 us/call  inner=45     burst=   53.8 ms
    g90      1583.48 us/call  inner=33     burst=   52.3 ms
    g91      1436.44 us/call  inner=36     burst=   51.7 ms
    g92      1806.78 us/call  inner=28     burst=   50.6 ms
    g93      1472.94 us/call  inner=36     burst=   53.0 ms
    g94      1816.45 us/call  inner=28     burst=   50.9 ms
    g95      1575.82 us/call  inner=34     burst=   53.6 ms
    legB     1656.87 us/call  inner=31     burst=   51.4 ms
    per-arm ratios, paired within a rep:
      s27  per-token                                         gemm/s27 1.0000x  chain/s27 1.0000x
      s24  per-token                                         gemm/s27 1.0928x  chain/s27 1.0946x
      s80  gs128      bank vs s24  1.4454x [1.3623-1.5102]  gemm/s27 1.5884x  chain/s27 1.3786x
      s84  per-token                                         gemm/s27 1.2142x  chain/s27 1.1520x
      s83  gs128      bank vs s84  1.2217x [1.1770-1.2601]  gemm/s27 1.4835x  chain/s27 1.3299x
      s86  per-token                                         gemm/s27 1.1882x  chain/s27 1.1021x
      s85  gs128      bank vs s86  1.1863x [1.1030-1.3035]  gemm/s27 1.4187x  chain/s27 1.2906x
      s91  per-token                                         gemm/s27 1.6284x  chain/s27 1.3898x
      s90  gs128      bank vs s91  1.0924x [1.0759-1.1433]  gemm/s27 1.7789x  chain/s27 1.5119x
      s93  per-token                                         gemm/s27 1.6733x  chain/s27 1.4389x
      s92  gs128      bank vs s93  1.2654x [1.1799-1.4209]  gemm/s27 2.0031x  chain/s27 1.6798x
      s95  per-token                                         gemm/s27 1.7611x  chain/s27 1.4610x
      s94  gs128      bank vs s95  1.1429x [1.1248-1.2835]  gemm/s27 2.0646x  chain/s27 1.7247x
      s89  gs256      bank vs s86  1.1128x [1.0886-1.2667]  gemm/s27 1.3336x  chain/s27 1.2185x
      BEST grouped s89 vs BEST per-token s27: gemm 1.3336x [1.2606-1.4311]  chain 1.2185x [1.1365-1.2592]

[sampler] 1868 samples over 93.5 s of process life; 91.3 s of that was inside a timed block

[under load] sm clock median 2700 MHz (min 1890, max 2715)  power median 404 W  sw_power_cap 835/1823 samples  hw_slowdown 0/1823
[tenancy after ] 2214528, 3928 MiB   (this process is pid 59)
