TENANCY-BEFORE-END
card=NVIDIA GeForce RTX 4090 sms=128 L2=72MiB burst>=140ms reps=5
idle clocks (sm,mem,W,C): 210, 405, 24.77, 43

      geometry   M  shp  n/burst |     chain   chain64     fixup    null | fix vs best | GB/s chain
  5120x1024     6    1     8977 |     17.10     16.85     11.21   17.11 |     +33.49% |   196.0  (null -0.06%, cpu/gpu 0.89)
  5120x1024    48    8     4373 |     36.80     27.63     33.94   36.98 |     -22.87% |   102.7  (null -0.50%, cpu/gpu 0.77)
  5120x1024    64    9     3516 |     45.69     32.52     42.20   46.00 |     -29.75% |    86.3  (null -0.68%, cpu/gpu 0.71)
  5120x6144     6    2     6927 |     23.33     39.12     23.20   23.33 |      +0.56% |   677.6  (null +0.04%, cpu/gpu 0.85)
  5120x6144    48    5     2986 |     54.02     91.59     60.45   53.98 |     -11.90% |   300.7  (null +0.08%, cpu/gpu 0.66)
  5120x6144    64    5     2889 |     55.93     92.84     64.90   55.90 |     -16.03% |   293.3  (null +0.05%, cpu/gpu 0.65)
  5120x10240    6    2     4634 |     35.13     62.21     35.33   35.13 |      -0.57% |   748.8  (null -0.00%, cpu/gpu 0.78)
  5120x10240   48    5     1988 |     81.07    147.12     85.65   80.94 |      -5.66% |   329.8  (null +0.16%, cpu/gpu 0.48)
  5120x10240   64    5     1951 |     82.67    148.86     89.21   82.87 |      -7.91% |   325.4  (null -0.25%, cpu/gpu 0.47)
  5120x12288    6    2     3319 |     48.64     81.40     48.76   48.59 |      -0.26% |   810.4  (null +0.11%, cpu/gpu 0.69)
  5120x12288   48    5     1438 |    112.26    206.03    116.63  112.08 |      -3.89% |   355.0  (null +0.17%, cpu/gpu 0.28)
  5120x12288   64    5     1421 |    114.12    207.58    120.08  113.98 |      -5.22% |   350.6  (null +0.13%, cpu/gpu 0.27)
  5120x17408    6    4     2937 |     54.75     77.11     54.74   54.80 |      +0.01% |   815.9  (null -0.10%, cpu/gpu 0.65)
  5120x17408   48   10     1184 |    136.08    229.73    136.06  136.06 |      +0.01% |   331.4  (null +0.02%, cpu/gpu 0.14)
  5120x17408   64   10     1117 |    144.06    233.03    143.98  143.79 |      +0.05% |   314.2  (null +0.19%, cpu/gpu 0.09)
  6144x5120     6    2     6816 |     23.65     39.34     23.44   23.65 |      +0.91% |   669.1  (null -0.01%, cpu/gpu 0.85)
  6144x5120    48    5     2880 |     56.14     92.31     61.39   56.17 |      -9.36% |   291.1  (null -0.06%, cpu/gpu 0.64)
  6144x5120    64    5     2781 |     58.00     93.32     65.73   57.99 |     -13.32% |   285.1  (null +0.02%, cpu/gpu 0.63)
 17408x5120     6    2     2877 |     57.53    102.88     57.18   57.41 |      +0.60% |   779.0  (null +0.20%, cpu/gpu 0.64)
 17408x5120    48    5     1180 |    136.32    243.97    141.28  136.52 |      -3.64% |   339.5  (null -0.15%, cpu/gpu 0.13)
 17408x5120    64    5     1159 |    139.24    246.43    146.12  138.89 |      -4.94% |   336.4  (null +0.25%, cpu/gpu 0.11)

clocks under load, first/last: 2775, 10251, 218.76, 52 | 2745, 10251, 450.71, 72

census-weighted over the 404 body linears, against the BEST chain
   M   leg-A us  recovered us  % of leg A
   6    14751.6         206.4       1.40%
  48    35508.8       -1101.6      -3.10%
  64    37219.5       -1569.5      -4.22%

Leg-A fractions. A decode step is leg A plus attention, norms, sampling and the drafter, so the step figure is strictly smaller.
TENANCY-AFTER
rc=0
