=== box arbi-whitebox-206 GPU1 (sole tenant, allocated) 2026-09-01T20:08:03+00:00 ===
NVIDIA GeForce RTX 4090, 1 MiB, 0 %, 210 MHz, 24.08 W

M=8192, TILE_M=128, operands bf16 (engine dtype), 9 paired rounds x 10 iters, L2 defeated (>192 MB pool + 256 MB scrub)

site                        K  G     B/e  unfus ms  fused ms  ratio   sign  saved ms  unf GB/s  fus GB/s   MHz
silu*mul -> down_proj   17408  1   9->5       1.392     0.798  1.743  9/9       0.594     922.1     893.2  2865
gate*sig -> o_proj       6144  1   9->5       0.495     0.279  1.771  9/9       0.215     915.8     901.2  2865
add+norm -> gate|up      5120  2  12->8       0.595     0.382  1.560  9/9       0.214     845.4     879.3  2865
add+norm -> qkv|z        5120  2  12->8       0.597     0.382  1.563  9/9       0.215     843.4     879.0  2850
add+norm -> q|k|v        5120  3  13->9       0.691     0.463  1.493  9/9       0.228     789.2     815.8  2850

site                   layers  ms/layer  ms total  MB/token   agreement
silu*mul -> down_proj      64     0.594      38.0     4.456   [+-1 step on 0.00001%]
gate*sig -> o_proj         16     0.215       3.4     0.393   [+-1 step on 0.00085%]
add+norm -> gate|up        64     0.214      13.7     1.311   [+-1 step on 0.00140%]
add+norm -> qkv|z          48     0.215      10.3     0.983   [+-1 step on 0.00157%]
add+norm -> q|k|v          16     0.228       3.7     0.328   [+-1 step on 0.00154%]
TOTAL                                        69.1     7.471
