########## RUN A: out_proj 6144x5120xK4, rot 6 ##########
[seed] int8 GEMM = /opt/cache-baked/torch_extensions/arbi_serve_exl3_i8_gemm_v1/arbi_serve_exl3_i8_gemm_v1.so
[seed] a_prep op = arbi_serve_exl3_a_prep.a_prep_frag
[tenancy before] (none)
[l2 scrub] mode='read' -- see the module docstring for the bias it carries

  shape  geometry                        thr  regs    smem blk/SM warps/SM   bank
  s85    64x128x128 1x8 st2 sw8 gs128    256   128   32768      2       16 32x2
  s86    64x128x128 1x8 st2 sw8          256    78   32768      3       24 32
  s94    32x128x128 1x8 st2 sw8 gs128    256    99   24576      2       16 16x2
  s95    32x128x128 1x8 st2 sw8          256    64   24576      4       32 16
  s98    16x128x128 1x8 st2 sw8 gs128    256    63   20480      4       32 8x2
  s99    16x128x128 1x8 st2 sw8          256    63   20480      4       32 8
  s110   128x128x128 1x4 st2 sw8         128   254   49152      2        8 128   <-- SPILLS
  s111   128x128x128 1x4 st2 sw8 gs128   128   255   50176      2        8 128x2

  out_proj k=6144 n=5120 M=8   burst>= 30 ms, reps=7
    a111       84.47 us/call  inner=324    burst=   27.4 ms
    a85        53.00 us/call  inner=510    burst=   27.0 ms
    a94        41.55 us/call  inner=646    burst=   26.8 ms
    a98        34.97 us/call  inner=767    burst=   26.8 ms
    f111      104.09 us/call  inner=256    burst=   26.6 ms
    f85        73.12 us/call  inner=357    burst=   26.1 ms
    f94        61.32 us/call  inner=417    burst=   25.6 ms
    f98        61.26 us/call  inner=427    burst=   26.2 ms
    g111       79.51 us/call  inner=344    burst=   27.4 ms
    g85        49.26 us/call  inner=555    burst=   27.3 ms
    g94        37.89 us/call  inner=713    burst=   27.0 ms
    g98        32.99 us/call  inner=817    burst=   27.0 ms
    legA       23.60 us/call  inner=1202   burst=   28.4 ms
    legB      131.64 us/call  inner=214    burst=   28.2 ms
    per-arm ratios, paired within a rep:
      s111 gs128                                           
      s98  gs128                                           
      s94  gs128                                           
      s85  gs128                                           
    served chain (a_prep_frag + folded epilogue) vs leg A, paired within a rep:
      s111 served/legA 3.5782x [3.5692-3.8459]  leg A wins
      s98  served/legA 1.4819x [1.4700-1.5792]  leg A wins
      s94  served/legA 1.7604x [1.7426-1.7683]  leg A wins
      s85  served/legA 2.2465x [2.2425-2.4200]  leg A wins
      BEST served tile s98: 1.4819x leg A [1.4700-1.5792] -- LEG A is faster at this M

  out_proj k=6144 n=5120 M=16   burst>= 30 ms, reps=7
    a111       84.66 us/call  inner=349    burst=   29.5 ms
    a85        53.06 us/call  inner=543    burst=   28.8 ms
    a94        41.85 us/call  inner=696    burst=   29.1 ms
    a98        35.36 us/call  inner=814    burst=   28.8 ms
    f111      104.12 us/call  inner=271    burst=   28.2 ms
    f85        73.45 us/call  inner=380    burst=   27.9 ms
    f94        62.99 us/call  inner=424    burst=   26.7 ms
    f98        54.45 us/call  inner=482    burst=   26.2 ms
    g111       79.53 us/call  inner=375    burst=   29.8 ms
    g85        49.26 us/call  inner=597    burst=   29.4 ms
    g94        38.13 us/call  inner=764    burst=   29.1 ms
    g98        34.06 us/call  inner=850    burst=   28.9 ms
    legA       24.22 us/call  inner=1209   burst=   29.3 ms
    legB      131.86 us/call  inner=227    burst=   29.9 ms
    per-arm ratios, paired within a rep:
      s111 gs128                                           
      s98  gs128                                           
      s94  gs128                                           
      s85  gs128                                           
    served chain (a_prep_frag + folded epilogue) vs leg A, paired within a rep:
      s111 served/legA 3.4959x [3.4877-3.5015]  leg A wins
      s98  served/legA 1.4598x [1.4519-1.4630]  leg A wins
      s94  served/legA 1.7278x [1.7136-1.7300]  leg A wins
      s85  served/legA 2.1913x [2.1899-2.1986]  leg A wins
      BEST served tile s98: 1.4598x leg A [1.4519-1.4630] -- LEG A is faster at this M

  out_proj k=6144 n=5120 M=32   burst>= 30 ms, reps=7
    a111       85.14 us/call  inner=343    burst=   29.2 ms
    a85        53.78 us/call  inner=544    burst=   29.3 ms
    a94        43.43 us/call  inner=663    burst=   28.8 ms
    a98        35.53 us/call  inner=804    burst=   28.6 ms
    f111      105.02 us/call  inner=268    burst=   28.1 ms
    f85        74.14 us/call  inner=369    burst=   27.4 ms
    f94        61.90 us/call  inner=442    burst=   27.4 ms
    f98        55.79 us/call  inner=486    burst=   27.1 ms
    g111       79.78 us/call  inner=373    burst=   29.8 ms
    g85        49.62 us/call  inner=591    burst=   29.3 ms
    g94        39.53 us/call  inner=728    burst=   28.8 ms
    g98        34.35 us/call  inner=844    burst=   29.0 ms
    legA       29.34 us/call  inner=1006   burst=   29.5 ms
    legB      130.34 us/call  inner=230    burst=   30.0 ms
    per-arm ratios, paired within a rep:
      s111 gs128                                           
      s98  gs128                                           
      s94  gs128                                           
      s85  gs128                                           
    served chain (a_prep_frag + folded epilogue) vs leg A, paired within a rep:
      s111 served/legA 2.9068x [2.8602-2.9079]  leg A wins
      s98  served/legA 1.2048x [1.1982-1.2137]  leg A wins
      s94  served/legA 1.4803x [1.4600-1.4964]  leg A wins
      s85  served/legA 1.8247x [1.8083-1.8372]  leg A wins
      BEST served tile s98: 1.2048x leg A [1.1982-1.2137] -- LEG A is faster at this M

  out_proj k=6144 n=5120 M=64   burst>= 30 ms, reps=7
    a111       86.23 us/call  inner=342    burst=   29.5 ms
    a85        54.28 us/call  inner=538    burst=   29.2 ms
    a94        43.81 us/call  inner=649    burst=   28.4 ms
    a98        52.82 us/call  inner=548    burst=   28.9 ms
    f111      107.86 us/call  inner=264    burst=   28.5 ms
    f85        74.94 us/call  inner=380    burst=   28.5 ms
    f94        63.60 us/call  inner=431    burst=   27.4 ms
    f98        75.43 us/call  inner=371    burst=   28.0 ms
    g111       80.11 us/call  inner=371    burst=   29.7 ms
    g85        50.45 us/call  inner=561    burst=   28.3 ms
    g94        39.70 us/call  inner=715    burst=   28.4 ms
    g98        51.49 us/call  inner=576    burst=   29.7 ms
    legA       51.28 us/call  inner=579    burst=   29.7 ms
    legB      119.21 us/call  inner=253    burst=   30.2 ms
    per-arm ratios, paired within a rep:
      s111 gs128                                           
      s98  gs128                                           
      s94  gs128                                           
      s85  gs128                                           
    served chain (a_prep_frag + folded epilogue) vs leg A, paired within a rep:
      s111 served/legA 1.6799x [1.6734-1.7210]  leg A wins
      s98  served/legA 1.0299x [1.0251-1.0686]  leg A wins
      s94  served/legA 0.8545x [0.8476-0.8922]  int8 WINS
      s85  served/legA 1.0563x [1.0535-1.0780]  leg A wins
      BEST served tile s94: 0.8545x leg A [0.8476-0.8922] -- the int8 leg is faster at this M

[sampler] 261 samples over 13.0 s of process life; 11.3 s of that was inside a timed block

[under load] sm clock median 2715 MHz (min 2520, max 2715)  power median 272 W  sw_power_cap 0/225 samples  hw_slowdown 0/225
[tenancy after ] 4161722, 1812 MiB   (this process is pid 1)
########## RUN B: gate_proj 5120x17408 + down_proj 17408x5120, rot 2 ##########
[seed] int8 GEMM = /opt/cache-baked/torch_extensions/arbi_serve_exl3_i8_gemm_v1/arbi_serve_exl3_i8_gemm_v1.so
[seed] a_prep op = arbi_serve_exl3_a_prep.a_prep_frag
[tenancy before] (none)
[l2 scrub] mode='read' -- see the module docstring for the bias it carries

  shape  geometry                        thr  regs    smem blk/SM warps/SM   bank
  s85    64x128x128 1x8 st2 sw8 gs128    256   128   32768      2       16 32x2
  s86    64x128x128 1x8 st2 sw8          256    78   32768      3       24 32
  s94    32x128x128 1x8 st2 sw8 gs128    256    99   24576      2       16 16x2
  s95    32x128x128 1x8 st2 sw8          256    64   24576      4       32 16
  s98    16x128x128 1x8 st2 sw8 gs128    256    63   20480      4       32 8x2
  s99    16x128x128 1x8 st2 sw8          256    63   20480      4       32 8
  s110   128x128x128 1x4 st2 sw8         128   254   49152      2        8 128   <-- SPILLS
  s111   128x128x128 1x4 st2 sw8 gs128   128   255   50176      2        8 128x2

  gate_proj k=5120 n=17408 M=8   burst>= 30 ms, reps=7
    a111      106.33 us/call  inner=258    burst=   27.4 ms
    a85        68.41 us/call  inner=397    burst=   27.2 ms
    a94        55.52 us/call  inner=499    burst=   27.7 ms
    a98        53.80 us/call  inner=530    burst=   28.5 ms
    f111      125.86 us/call  inner=213    burst=   26.8 ms
    f85        90.12 us/call  inner=292    burst=   26.3 ms
    f94        75.34 us/call  inner=350    burst=   26.4 ms
    f98        75.51 us/call  inner=360    burst=   27.2 ms
    g111      101.07 us/call  inner=271    burst=   27.4 ms
    g85        65.42 us/call  inner=424    burst=   27.7 ms
    g94        51.59 us/call  inner=540    burst=   27.9 ms
    g98        50.50 us/call  inner=566    burst=   28.6 ms
    legA       52.83 us/call  inner=557    burst=   29.4 ms
    legB      489.36 us/call  inner=60     burst=   29.4 ms
    per-arm ratios, paired within a rep:
      s111 gs128                                           
      s98  gs128                                           
      s94  gs128                                           
      s85  gs128                                           
    served chain (a_prep_frag + folded epilogue) vs leg A, paired within a rep:
      s111 served/legA 2.0123x [2.0075-2.1354]  leg A wins
      s98  served/legA 1.0182x [1.0155-1.0329]  leg A wins
      s94  served/legA 1.0521x [1.0464-1.0560]  leg A wins
      s85  served/legA 1.2940x [1.2904-1.3907]  leg A wins
      BEST served tile s98: 1.0182x leg A [1.0155-1.0329] -- LEG A is faster at this M

  gate_proj k=5120 n=17408 M=16   burst>= 30 ms, reps=7
    a111      106.53 us/call  inner=278    burst=   29.6 ms
    a85        68.64 us/call  inner=428    burst=   29.4 ms
    a94        55.77 us/call  inner=513    burst=   28.6 ms
    a98        53.97 us/call  inner=537    burst=   29.0 ms
    f111      126.74 us/call  inner=226    burst=   28.6 ms
    f85        90.97 us/call  inner=310    burst=   28.2 ms
    f94        76.70 us/call  inner=359    burst=   27.5 ms
    f98        72.42 us/call  inner=386    burst=   28.0 ms
    g111      101.12 us/call  inner=295    burst=   29.8 ms
    g85        65.52 us/call  inner=449    burst=   29.4 ms
    g94        51.79 us/call  inner=563    burst=   29.2 ms
    g98        50.63 us/call  inner=569    burst=   28.8 ms
    legA       53.56 us/call  inner=555    burst=   29.7 ms
    legB      495.85 us/call  inner=61     burst=   30.2 ms
    per-arm ratios, paired within a rep:
      s111 gs128                                           
      s98  gs128                                           
      s94  gs128                                           
      s85  gs128                                           
    served chain (a_prep_frag + folded epilogue) vs leg A, paired within a rep:
      s111 served/legA 1.9895x [1.9845-2.0031]  leg A wins
      s98  served/legA 1.0077x [1.0074-1.0128]  leg A wins
      s94  served/legA 1.0412x [1.0386-1.0429]  leg A wins
      s85  served/legA 1.2816x [1.2780-1.2856]  leg A wins
      BEST served tile s98: 1.0077x leg A [1.0074-1.0128] -- LEG A is faster at this M

  gate_proj k=5120 n=17408 M=32   burst>= 30 ms, reps=7
    a111      108.29 us/call  inner=275    burst=   29.8 ms
    a85        69.98 us/call  inner=420    burst=   29.4 ms
    a94        56.48 us/call  inner=519    burst=   29.3 ms
    a98        66.48 us/call  inner=443    burst=   29.5 ms
    f111      134.04 us/call  inner=220    burst=   29.5 ms
    f85        92.42 us/call  inner=305    burst=   28.2 ms
    f94        75.14 us/call  inner=361    burst=   27.1 ms
    f98        86.39 us/call  inner=328    burst=   28.3 ms
    g111      101.94 us/call  inner=294    burst=   30.0 ms
    g85        66.11 us/call  inner=447    burst=   29.6 ms
    g94        52.25 us/call  inner=553    burst=   28.9 ms
    g98        66.63 us/call  inner=445    burst=   29.6 ms
    legA       65.90 us/call  inner=453    burst=   29.9 ms
    legB      483.99 us/call  inner=63     burst=   30.5 ms
    per-arm ratios, paired within a rep:
      s111 gs128                                           
      s98  gs128                                           
      s94  gs128                                           
      s85  gs128                                           
    served chain (a_prep_frag + folded epilogue) vs leg A, paired within a rep:
      s111 served/legA 1.6392x [1.5679-1.7164]  leg A wins
      s98  served/legA 1.0091x [0.9600-1.0149]  leg A wins
      s94  served/legA 0.8547x [0.8241-0.8668]  int8 WINS
      s85  served/legA 1.0624x [1.0075-1.1065]  leg A wins
      BEST served tile s94: 0.8547x leg A [0.8241-0.8668] -- the int8 leg is faster at this M

  gate_proj k=5120 n=17408 M=64   burst>= 30 ms, reps=7
    a111      109.21 us/call  inner=272    burst=   29.7 ms
    a85        70.35 us/call  inner=414    burst=   29.1 ms
    a94        86.29 us/call  inner=342    burst=   29.5 ms
    a98       109.67 us/call  inner=271    burst=   29.7 ms
    f111      136.20 us/call  inner=215    burst=   29.3 ms
    f85        94.04 us/call  inner=307    burst=   28.9 ms
    f94       108.57 us/call  inner=272    burst=   29.5 ms
    f98       133.68 us/call  inner=219    burst=   29.3 ms
    g111      103.44 us/call  inner=291    burst=   30.1 ms
    g85        67.49 us/call  inner=426    burst=   28.8 ms
    g94        82.01 us/call  inner=358    burst=   29.4 ms
    g98       109.98 us/call  inner=272    burst=   29.9 ms
    legA      123.29 us/call  inner=250    burst=   30.8 ms
    legB      490.06 us/call  inner=63     burst=   30.9 ms
    per-arm ratios, paired within a rep:
      s111 gs128                                           
      s98  gs128                                           
      s94  gs128                                           
      s85  gs128                                           
    served chain (a_prep_frag + folded epilogue) vs leg A, paired within a rep:
      s111 served/legA 0.9018x [0.8355-0.9219]  int8 WINS
      s98  served/legA 0.8896x [0.8269-0.9164]  int8 WINS
      s94  served/legA 0.7125x [0.6610-0.7426]  int8 WINS
      s85  served/legA 0.5791x [0.5558-0.5859]  int8 WINS
      BEST served tile s85: 0.5791x leg A [0.5558-0.5859] -- the int8 leg is faster at this M

  down_proj k=17408 n=5120 M=8   burst>= 30 ms, reps=7
    a111      229.74 us/call  inner=130    burst=   29.9 ms
    a85       142.14 us/call  inner=209    burst=   29.7 ms
    a94       109.38 us/call  inner=271    burst=   29.6 ms
    a98        91.49 us/call  inner=323    burst=   29.6 ms
    f111      248.85 us/call  inner=118    burst=   29.4 ms
    f85       160.86 us/call  inner=179    burst=   28.8 ms
    f94       127.82 us/call  inner=223    burst=   28.5 ms
    f98       114.21 us/call  inner=244    burst=   27.9 ms
    g111      221.61 us/call  inner=136    burst=   30.1 ms
    g85       135.75 us/call  inner=219    burst=   29.7 ms
    g94       103.26 us/call  inner=285    burst=   29.4 ms
    g98        89.91 us/call  inner=326    burst=   29.3 ms
    legA       55.47 us/call  inner=537    burst=   29.8 ms
    legB      486.25 us/call  inner=62     burst=   30.1 ms
    per-arm ratios, paired within a rep:
      s111 gs128                                           
      s98  gs128                                           
      s94  gs128                                           
      s85  gs128                                           
    served chain (a_prep_frag + folded epilogue) vs leg A, paired within a rep:
      s111 served/legA 4.1420x [4.1388-4.1462]  leg A wins
      s98  served/legA 1.6497x [1.6465-1.6502]  leg A wins
      s94  served/legA 1.9720x [1.9684-1.9738]  leg A wins
      s85  served/legA 2.5634x [2.5575-2.5668]  leg A wins
      BEST served tile s98: 1.6497x leg A [1.6465-1.6502] -- LEG A is faster at this M

  down_proj k=17408 n=5120 M=16   burst>= 30 ms, reps=7
    a111      230.10 us/call  inner=130    burst=   29.9 ms
    a85       142.20 us/call  inner=209    burst=   29.7 ms
    a94       109.84 us/call  inner=270    burst=   29.7 ms
    a98        91.42 us/call  inner=323    burst=   29.5 ms
    f111      249.46 us/call  inner=118    burst=   29.4 ms
    f85       162.40 us/call  inner=179    burst=   29.1 ms
    f94       129.21 us/call  inner=222    burst=   28.7 ms
    f98       114.99 us/call  inner=248    burst=   28.5 ms
    g111      221.68 us/call  inner=135    burst=   29.9 ms
    g85       135.76 us/call  inner=218    burst=   29.6 ms
    g94       103.73 us/call  inner=284    burst=   29.5 ms
    g98        92.72 us/call  inner=316    burst=   29.3 ms
    legA       59.70 us/call  inner=497    burst=   29.7 ms
    legB      489.50 us/call  inner=62     burst=   30.3 ms
    per-arm ratios, paired within a rep:
      s111 gs128                                           
      s98  gs128                                           
      s94  gs128                                           
      s85  gs128                                           
    served chain (a_prep_frag + folded epilogue) vs leg A, paired within a rep:
      s111 served/legA 3.8553x [3.8474-3.8563]  leg A wins
      s98  served/legA 1.5310x [1.5290-1.5334]  leg A wins
      s94  served/legA 1.8395x [1.8381-1.8417]  leg A wins
      s85  served/legA 2.3821x [2.3795-2.3833]  leg A wins
      BEST served tile s98: 1.5310x leg A [1.5290-1.5334] -- LEG A is faster at this M

  down_proj k=17408 n=5120 M=32   burst>= 30 ms, reps=7
    a111      230.49 us/call  inner=130    burst=   30.0 ms
    a85       142.54 us/call  inner=207    burst=   29.5 ms
    a94       114.38 us/call  inner=258    burst=   29.5 ms
    a98        91.88 us/call  inner=322    burst=   29.6 ms
    f111      252.13 us/call  inner=117    burst=   29.5 ms
    f85       166.10 us/call  inner=176    burst=   29.2 ms
    f94       133.85 us/call  inner=217    burst=   29.0 ms
    f98       119.73 us/call  inner=240    burst=   28.7 ms
    g111      221.77 us/call  inner=135    burst=   29.9 ms
    g85       136.14 us/call  inner=217    burst=   29.5 ms
    g94       107.84 us/call  inner=268    burst=   28.9 ms
    g98        93.74 us/call  inner=311    burst=   29.2 ms
    legA       67.95 us/call  inner=436    burst=   29.6 ms
    legB      480.39 us/call  inner=63     burst=   30.3 ms
    per-arm ratios, paired within a rep:
      s111 gs128                                           
      s98  gs128                                           
      s94  gs128                                           
      s85  gs128                                           
    served chain (a_prep_frag + folded epilogue) vs leg A, paired within a rep:
      s111 served/legA 3.3922x [3.3830-3.5657]  leg A wins
      s98  served/legA 1.3528x [1.3482-1.3847]  leg A wins
      s94  served/legA 1.6843x [1.6754-1.6865]  leg A wins
      s85  served/legA 2.0994x [2.0928-2.1090]  leg A wins
      BEST served tile s98: 1.3528x leg A [1.3482-1.3847] -- LEG A is faster at this M

  down_proj k=17408 n=5120 M=64   burst>= 30 ms, reps=7
    a111      231.52 us/call  inner=129    burst=   29.9 ms
    a85       143.83 us/call  inner=208    burst=   29.9 ms
    a94       114.77 us/call  inner=260    burst=   29.8 ms
    a98       142.12 us/call  inner=209    burst=   29.7 ms
    f111      256.82 us/call  inner=115    burst=   29.5 ms
    f85       168.16 us/call  inner=175    burst=   29.4 ms
    f94       137.96 us/call  inner=211    burst=   29.1 ms
    f98       172.39 us/call  inner=171    burst=   29.5 ms
    g111      223.09 us/call  inner=135    burst=   30.1 ms
    g85       139.23 us/call  inner=206    burst=   28.7 ms
    g94       108.44 us/call  inner=267    burst=   29.0 ms
    g98       142.56 us/call  inner=210    burst=   29.9 ms
    legA      124.36 us/call  inner=240    burst=   29.8 ms
    legB      486.12 us/call  inner=62     burst=   30.1 ms
    per-arm ratios, paired within a rep:
      s111 gs128                                           
      s98  gs128                                           
      s94  gs128                                           
      s85  gs128                                           
    served chain (a_prep_frag + folded epilogue) vs leg A, paired within a rep:
      s111 served/legA 1.8627x [1.8602-1.8658]  leg A wins
      s98  served/legA 1.1438x [1.1413-1.1462]  leg A wins
      s94  served/legA 0.9223x [0.9219-0.9262]  int8 WINS
      s85  served/legA 1.1550x [1.1501-1.1603]  leg A wins
      BEST served tile s94: 0.9223x leg A [0.9219-0.9262] -- the int8 leg is faster at this M

[sampler] 556 samples over 27.8 s of process life; 23.2 s of that was inside a timed block

[under load] sm clock median 2715 MHz (min 2400, max 2715)  power median 356 W  sw_power_cap 87/464 samples  hw_slowdown 0/464
[tenancy after ] 4164333, 1682 MiB   (this process is pid 1)
[seed] int8 GEMM = /opt/cache-baked/torch_extensions/arbi_serve_exl3_i8_gemm_v1/arbi_serve_exl3_i8_gemm_v1.so
[seed] a_prep op = arbi_serve_exl3_a_prep.a_prep_frag
[tenancy before] (none)
[l2 scrub] mode='read' -- see the module docstring for the bias it carries

  shape  geometry                        thr  regs    smem blk/SM warps/SM   bank
  s85    64x128x128 1x8 st2 sw8 gs128    256   128   32768      2       16 32x2
  s86    64x128x128 1x8 st2 sw8          256    78   32768      3       24 32
  s94    32x128x128 1x8 st2 sw8 gs128    256    99   24576      2       16 16x2
  s95    32x128x128 1x8 st2 sw8          256    64   24576      4       32 16
  s98    16x128x128 1x8 st2 sw8 gs128    256    63   20480      4       32 8x2
  s99    16x128x128 1x8 st2 sw8          256    63   20480      4       32 8

  out_proj k=6144 n=5120 M=40   burst>= 30 ms, reps=7
    a85        53.97 us/call  inner=503    burst=   27.1 ms
    a94        43.93 us/call  inner=616    burst=   27.1 ms
    a98        36.32 us/call  inner=744    burst=   27.0 ms
    f85        75.17 us/call  inner=344    burst=   25.9 ms
    f94        65.83 us/call  inner=391    burst=   25.7 ms
    f98        62.29 us/call  inner=424    burst=   26.4 ms
    g85        50.00 us/call  inner=541    burst=   27.0 ms
    g94        39.94 us/call  inner=682    burst=   27.2 ms
    g98        34.63 us/call  inner=757    burst=   26.2 ms
    legA       43.35 us/call  inner=603    burst=   26.1 ms
    legB      120.75 us/call  inner=234    burst=   28.3 ms
    per-arm ratios, paired within a rep:
      s98  gs128                                           
      s94  gs128                                           
      s85  gs128                                           
    served chain (a_prep_frag + folded epilogue) vs leg A, paired within a rep:
      s98  served/legA 0.8394x [0.8262-0.8859]  int8 WINS
      s94  served/legA 1.0152x [1.0085-1.0333]  leg A wins
      s85  served/legA 1.2449x [1.2384-1.3302]  leg A wins
      BEST served tile s98: 0.8394x leg A [0.8262-0.8859] -- the int8 leg is faster at this M

  out_proj k=6144 n=5120 M=48   burst>= 30 ms, reps=7
    a85        54.06 us/call  inner=529    burst=   28.6 ms
    a94        43.95 us/call  inner=661    burst=   29.0 ms
    a98        35.93 us/call  inner=790    burst=   28.4 ms
    f85        76.15 us/call  inner=361    burst=   27.5 ms
    f94        66.21 us/call  inner=413    burst=   27.3 ms
    f98        57.52 us/call  inner=478    burst=   27.5 ms
    g85        50.04 us/call  inner=566    burst=   28.3 ms
    g94        39.72 us/call  inner=724    burst=   28.8 ms
    g98        34.71 us/call  inner=829    burst=   28.8 ms
    legA       44.52 us/call  inner=648    burst=   28.9 ms
    legB      118.21 us/call  inner=253    burst=   29.9 ms
    per-arm ratios, paired within a rep:
      s98  gs128                                           
      s94  gs128                                           
      s85  gs128                                           
    served chain (a_prep_frag + folded epilogue) vs leg A, paired within a rep:
      s98  served/legA 0.8036x [0.7793-0.8308]  int8 WINS
      s94  served/legA 0.9862x [0.9573-1.0075]  int8 WINS
      s85  served/legA 1.2151x [1.1779-1.2192]  leg A wins
      BEST served tile s98: 0.8036x leg A [0.7793-0.8308] -- the int8 leg is faster at this M

  out_proj k=6144 n=5120 M=56   burst>= 30 ms, reps=7
    a85        54.38 us/call  inner=537    burst=   29.2 ms
    a94        43.82 us/call  inner=653    burst=   28.6 ms
    a98        53.02 us/call  inner=547    burst=   29.0 ms
    f85        76.60 us/call  inner=363    burst=   27.8 ms
    f94        66.60 us/call  inner=405    burst=   27.0 ms
    f98        78.91 us/call  inner=356    burst=   28.1 ms
    g85        50.22 us/call  inner=574    burst=   28.8 ms
    g94        39.86 us/call  inner=722    burst=   28.8 ms
    g98        51.64 us/call  inner=574    burst=   29.6 ms
    legA       50.60 us/call  inner=588    burst=   29.8 ms
    legB      118.28 us/call  inner=255    burst=   30.2 ms
    per-arm ratios, paired within a rep:
      s98  gs128                                           
      s94  gs128                                           
      s85  gs128                                           
    served chain (a_prep_frag + folded epilogue) vs leg A, paired within a rep:
      s98  served/legA 1.0500x [1.0357-1.0567]  leg A wins
      s94  served/legA 0.8659x [0.8592-0.8721]  int8 WINS
      s85  served/legA 1.0746x [1.0634-1.0960]  leg A wins
      BEST served tile s94: 0.8659x leg A [0.8592-0.8721] -- the int8 leg is faster at this M

[sampler] 153 samples over 7.6 s of process life; 6.6 s of that was inside a timed block

[under load] sm clock median 2715 MHz (min 2520, max 2715)  power median 363 W  sw_power_cap 38/131 samples  hw_slowdown 0/131
[tenancy after ] 4167981, 1668 MiB   (this process is pid 1)
