
  out_proj k=6144 n=5120 M=256   target = the EXACT int8 result in float64
    s27  FOLD16 (byte-exact) row-rel mean 3.8215e-04 CI [3.7323e-04, 3.9228e-04]
         FOLD32 (fp32 butterfly) row-rel mean 2.4015e-04 CI [2.2848e-04, 2.5331e-04]
         gain 1.5913x   fp16 results moved 645531/1310720 = 49.25%   FOLD16 == shipped chain byte for byte: True

GATE 6 fp32 BUTTERFLY: PASS
