device: NVIDIA GeForce RTX 4090   tenancy before: 2281 MiB in use of 24080
  IQ2_XXS ggml's own 640 blocks, both tilings: EQUAL at every probed k
  IQ2_XXS 512x256, both tilings: EQUAL at every probed k
  IQ2_XXS grid vs lut at splits=1, dense 512x256, M=[1, 8]: BIT-EQUAL
L2: cold   M=1   3 races x 11 paired reps
shape      arm       MiB   bar us  live us  null us    GB/s  of bar null/bar live/null  kc/splitsggate/up    lut     44.49    51.20    57.34    59.39   813.6   0.893    1.160     0.966   320/16  
gate/up    grid    23.24    28.67    45.06    34.82   540.9   0.636    1.214     1.294   480/11  
                 grid/lut  0.784 [0.774, 0.787]  control 1.000 [0.989, 1.011]  over 3 races  SEPARATEDddown       lut     44.49    64.51    71.68    73.73   650.9   0.900    1.143     0.972  1088/16  
down       grid    23.24    32.77    49.15    40.96   495.8   0.667    1.250     1.200  1088/16  
                 grid/lut  0.697 [0.694, 0.707]  control 1.000 [0.992, 1.000]  over 3 races  SEPARATEDkk/v        lut      2.62     7.17    11.26     9.54   243.6   0.636    1.330     1.181   160/32  
k/v        grid     1.37     5.12    13.38    11.26   107.2   0.383    2.200     1.188   320/16  
                 grid/lut  1.059 [1.059, 1.062]  control 1.000 [0.986, 1.000]  over 3 races  SEPARATED
tenancy after: 2145 MiB in use of 24080
1 MiB in use of 24080
  grid     1.37     5.12    13.31    12.29   107.7   0.385    2.400     1.083   320/16  
                 grid/lut  1.063 [1.057, 1.097]  control 1.000 [1.000, 1.025]  over 3 races  SEPARATED
tenancy after: 9061 MiB in use of 24080
