device: NVIDIA GeForce RTX 4090   tenancy before: 395 MiB in use of 24109
type: IQ1_M   arms: the served spelling and its twin
  IQ1_M ggml's own 640 blocks, both tilings: EQUAL at every probed k
  IQ1_M 512x256, both tilings: EQUAL at every probed k
  IQ1_M grid vs lut at splits=1, dense 512x256, M=[1, 8]: BIT-EQUAL
L2: cold   M=1   5 races x 21 paired reps
shape      arm       MiB   bar us  live us  null us    GB/s  of bar null/bar live/null  kc/splits
gate/up    lut     45.82    54.66    64.90    57.44   740.4   0.842    1.051     1.130   320/16  
gate/up    grid    19.26    24.58    71.68    39.01   281.7   0.343    1.587     1.838   480/11  
                 grid/lut  1.094 [1.091, 1.158]  control 1.003 [0.996, 1.016]  over 5 races  SEPARATED
down       lut     45.82    55.30    66.56    64.51   721.8   0.831    1.167     1.032  1088/16  
down       grid    19.26    24.58    75.78    41.98   266.5   0.324    1.708     1.805  1088/16  
                 grid/lut  1.138 [1.095, 1.190]  control 0.977 [0.962, 1.017]  over 5 races  SEPARATED
k/v        lut      2.70     8.58    15.36    15.62   184.0   0.558    1.821     0.984   160/32  
k/v        grid     1.13     6.24    19.33    18.30    61.5   0.323    2.933     1.056   320/16  
                 grid/lut  1.012 [0.963, 1.188]  control 0.988 [0.817, 1.041]  over 5 races  INSIDE THE CONTROL — no result
tenancy after: 719 MiB in use of 24109
