[state] 210, 40.53, 43, Not Active   ideal divided by 900 GB/s (measured, not nominal)

  gate_proj k=5120 n=17408 M=2048   output 71.3 MB (INSIDE the 72 MB L2)   [2385, 466.11, 59, Active]
    standalone had_r_128 pass   0.1594 ms    894.7 GB/s over   142.6 MB   analytic ideal saving 0.1585 ms
    s27  128x128x128 1x4w st2 sw8   chain nofold  1.0156 fold  0.8903  paired 1.1433x [1.1148-1.1621]  saved   125.4 us =  0.79x analytic =  0.79x the measured pass
         gemm alone: nofold  0.7913  fold  0.8231  epilogue cost   +31.8 us (+19.9% of the pass it replaces)
    s12  128x128x64  1x4w st3 sw8   chain nofold  1.1220 fold  1.0108  paired 1.1079x [1.0978-1.1212]  saved   111.2 us =  0.70x analytic =  0.70x the measured pass
         gemm alone: nofold  0.9058  fold  0.9209  epilogue cost   +15.1 us ( +9.5% of the pass it replaces)
    s10  128x128x64  1x4w st3 sw1   chain nofold  1.1424 fold  0.9932  paired 1.1392x [1.1318-1.1572]  saved   149.2 us =  0.94x analytic =  0.94x the measured pass
         gemm alone: nofold  0.9194  fold  0.9119  epilogue cost    -7.5 us ( -4.7% of the pass it replaces)

  gate_proj k=5120 n=17408 M=8192   output 285.2 MB (past the 72 MB L2)   [2430, 449.10, 62, Active]
    standalone had_r_128 pass   0.6490 ms    879.0 GB/s over   570.4 MB   analytic ideal saving 0.6338 ms
    s27  128x128x128 1x4w st2 sw8   chain nofold  4.2568 fold  3.6480  paired 1.1410x [1.1289-1.1670]  saved   608.8 us =  0.96x analytic =  0.94x the measured pass
         gemm alone: nofold  3.2809  fold  3.3679  epilogue cost   +87.0 us (+13.4% of the pass it replaces)
    s12  128x128x64  1x4w st3 sw8   chain nofold  4.5949 fold  3.9980  paired 1.1646x [1.1033-1.1713]  saved   596.9 us =  0.94x analytic =  0.92x the measured pass
         gemm alone: nofold  3.5398  fold  3.6923  epilogue cost  +152.5 us (+23.5% of the pass it replaces)
    s10  128x128x64  1x4w st3 sw1   chain nofold  4.5761 fold  3.9518  paired 1.1556x [1.1129-1.1643]  saved   624.3 us =  0.98x analytic =  0.96x the measured pass
         gemm alone: nofold  3.5673  fold  3.7226  epilogue cost  +155.4 us (+23.9% of the pass it replaces)

  gate_proj k=5120 n=17408 M=16384   output 570.4 MB (past the 72 MB L2)   [2460, 450.99, 62, Active]
    standalone had_r_128 pass   1.2946 ms    881.2 GB/s over  1140.9 MB   analytic ideal saving 1.2676 ms
    s27  128x128x128 1x4w st2 sw8   chain nofold  8.3682 fold  7.5054  paired 1.1284x [1.1150-1.1447]  saved   862.8 us =  0.68x analytic =  0.67x the measured pass
         gemm alone: nofold  6.6907  fold  6.8384  epilogue cost  +147.7 us (+11.4% of the pass it replaces)
    s12  128x128x64  1x4w st3 sw8   chain nofold  9.2041 fold  8.1904  paired 1.1391x [1.0930-1.1598]  saved  1013.8 us =  0.80x analytic =  0.78x the measured pass
         gemm alone: nofold  7.2072  fold  7.4240  epilogue cost  +216.8 us (+16.7% of the pass it replaces)
    s10  128x128x64  1x4w st3 sw1   chain nofold  9.0600 fold  8.1014  paired 1.1222x [1.0994-1.1437]  saved   958.6 us =  0.76x analytic =  0.74x the measured pass
         gemm alone: nofold  7.1338  fold  7.2971  epilogue cost  +163.3 us (+12.6% of the pass it replaces)

  out_proj k=6144 n=5120 M=2048   output 21.0 MB (INSIDE the 72 MB L2)   [2850, 433.25, 62, Not Active]
    standalone had_r_128 pass   0.0399 ms   1052.4 GB/s over    41.9 MB   analytic ideal saving 0.0466 ms
    s27  128x128x128 1x4w st2 sw8   chain nofold  0.3953 fold  0.3639  paired 1.0864x [1.0838-1.0889]  saved    31.4 us =  0.67x analytic =  0.79x the measured pass
         gemm alone: nofold  0.2819  fold  0.2902  epilogue cost    +8.4 us (+21.0% of the pass it replaces)
    s12  128x128x64  1x4w st3 sw8   chain nofold  0.4690 fold  0.4405  paired 1.0639x [1.0550-1.0836]  saved    28.6 us =  0.61x analytic =  0.72x the measured pass
         gemm alone: nofold  0.3526  fold  0.3620  epilogue cost    +9.4 us (+23.5% of the pass it replaces)
    s10  128x128x64  1x4w st3 sw1   chain nofold  0.4681 fold  0.4423  paired 1.0570x [1.0343-1.0698]  saved    25.9 us =  0.55x analytic =  0.65x the measured pass
         gemm alone: nofold  0.3544  fold  0.3645  epilogue cost   +10.1 us (+25.3% of the pass it replaces)

  out_proj k=6144 n=5120 M=8192   output 83.9 MB (past the 72 MB L2)   [2550, 445.90, 61, Active]
    standalone had_r_128 pass   0.1967 ms    852.8 GB/s over   167.8 MB   analytic ideal saving 0.1864 ms
    s27  128x128x128 1x4w st2 sw8   chain nofold  1.7688 fold  1.6286  paired 1.0989x [1.0817-1.1016]  saved   140.2 us =  0.75x analytic =  0.71x the measured pass
         gemm alone: nofold  1.0950  fold  1.1305  epilogue cost   +35.5 us (+18.0% of the pass it replaces)
    s12  128x128x64  1x4w st3 sw8   chain nofold  1.8869 fold  1.7471  paired 1.0798x [1.0619-1.0988]  saved   139.8 us =  0.75x analytic =  0.71x the measured pass
         gemm alone: nofold  1.1969  fold  1.2357  epilogue cost   +38.8 us (+19.7% of the pass it replaces)
    s10  128x128x64  1x4w st3 sw1   chain nofold  1.8936 fold  1.7339  paired 1.0952x [1.0652-1.1087]  saved   159.7 us =  0.86x analytic =  0.81x the measured pass
         gemm alone: nofold  1.2346  fold  1.2422  epilogue cost    +7.6 us ( +3.9% of the pass it replaces)

  out_proj k=6144 n=5120 M=16384   output 167.8 MB (past the 72 MB L2)   [2640, 446.41, 63, Active]
    standalone had_r_128 pass   0.3854 ms    870.7 GB/s over   335.5 MB   analytic ideal saving 0.3728 ms
    s27  128x128x128 1x4w st2 sw8   chain nofold  3.5534 fold  3.2492  paired 1.0936x [1.0664-1.1020]  saved   304.1 us =  0.82x analytic =  0.79x the measured pass
         gemm alone: nofold  2.2393  fold  2.3252  epilogue cost   +85.9 us (+22.3% of the pass it replaces)
    s12  128x128x64  1x4w st3 sw8   chain nofold  3.8704 fold  3.5112  paired 1.1010x [1.0544-1.1241]  saved   359.2 us =  0.96x analytic =  0.93x the measured pass
         gemm alone: nofold  2.4008  fold  2.5017  epilogue cost  +100.9 us (+26.2% of the pass it replaces)
    s10  128x128x64  1x4w st3 sw1   chain nofold  3.7996 fold  3.4897  paired 1.1088x [1.0616-1.1260]  saved   309.9 us =  0.83x analytic =  0.80x the measured pass
         gemm alone: nofold  2.4243  fold  2.5562  epilogue cost  +131.9 us (+34.2% of the pass it replaces)

  down_proj k=17408 n=5120 M=2048   output 21.0 MB (INSIDE the 72 MB L2)   [2625, 445.77, 62, Active]
    standalone had_r_128 pass   0.0416 ms   1007.2 GB/s over    41.9 MB   analytic ideal saving 0.0466 ms
    s27  128x128x128 1x4w st2 sw8   chain nofold  1.1553 fold  1.1043  paired 1.0451x [1.0371-1.0587]  saved    51.0 us =  1.09x analytic =  1.22x the measured pass   <-- ABOVE THE PASS IT DELETES, INSTRUMENT FAULT
         gemm alone: nofold  0.7841  fold  0.7931  epilogue cost    +8.9 us (+21.5% of the pass it replaces)
    s12  128x128x64  1x4w st3 sw8   chain nofold  1.3242 fold  1.2886  paired 1.0243x [1.0140-1.0886]  saved    35.6 us =  0.76x analytic =  0.85x the measured pass
         gemm alone: nofold  0.9841  fold  0.9915  epilogue cost    +7.3 us (+17.6% of the pass it replaces)
    s10  128x128x64  1x4w st3 sw1   chain nofold  1.3298 fold  1.2929  paired 1.0287x [1.0238-1.0392]  saved    36.9 us =  0.79x analytic =  0.89x the measured pass
         gemm alone: nofold  0.9743  fold  0.9859  epilogue cost   +11.7 us (+28.1% of the pass it replaces)

  down_proj k=17408 n=5120 M=8192   output 83.9 MB (past the 72 MB L2)   [2520, 445.68, 64, Active]
    standalone had_r_128 pass   0.1919 ms    874.3 GB/s over   167.8 MB   analytic ideal saving 0.1864 ms
    s27  128x128x128 1x4w st2 sw8   chain nofold  4.6443 fold  4.4696  paired 1.0364x [1.0098-1.0391]  saved   174.7 us =  0.94x analytic =  0.91x the measured pass
         gemm alone: nofold  3.1406  fold  3.1740  epilogue cost   +33.4 us (+17.4% of the pass it replaces)
    s12  128x128x64  1x4w st3 sw8   chain nofold  5.0146 fold  4.7255  paired 1.0640x [0.9958-1.0700]  saved   289.1 us =  1.55x analytic =  1.51x the measured pass   <-- BAND CROSSES 1.0, BELOW THE NOISE FLOOR   <-- ABOVE THE PASS IT DELETES, INSTRUMENT FAULT
         gemm alone: nofold  3.3302  fold  3.4846  epilogue cost  +154.4 us (+80.4% of the pass it replaces)
    s10  128x128x64  1x4w st3 sw1   chain nofold  4.9609 fold  4.7498  paired 1.0533x [0.9958-1.0657]  saved   211.1 us =  1.13x analytic =  1.10x the measured pass   <-- BAND CROSSES 1.0, BELOW THE NOISE FLOOR   <-- ABOVE THE PASS IT DELETES, INSTRUMENT FAULT
         gemm alone: nofold  3.3755  fold  3.5511  epilogue cost  +175.5 us (+91.5% of the pass it replaces)

  down_proj k=17408 n=5120 M=16384   output 167.8 MB (past the 72 MB L2)   [2475, 452.00, 65, Active]
    standalone had_r_128 pass   0.3832 ms    875.6 GB/s over   335.5 MB   analytic ideal saving 0.3728 ms
    s27  128x128x128 1x4w st2 sw8   chain nofold  9.2500 fold  8.9032  paired 1.0372x [1.0168-1.0483]  saved   346.8 us =  0.93x analytic =  0.90x the measured pass
         gemm alone: nofold  6.5049  fold  6.5923  epilogue cost   +87.4 us (+22.8% of the pass it replaces)
    s12  128x128x64  1x4w st3 sw8   chain nofold  9.8024 fold  9.4027  paired 1.0425x [0.9939-1.0657]  saved   399.7 us =  1.07x analytic =  1.04x the measured pass   <-- BAND CROSSES 1.0, BELOW THE NOISE FLOOR
         gemm alone: nofold  6.9016  fold  7.1366  epilogue cost  +235.0 us (+61.3% of the pass it replaces)
    s10  128x128x64  1x4w st3 sw1   chain nofold  9.8594 fold  9.4685  paired 1.0453x [1.0018-1.0542]  saved   390.9 us =  1.05x analytic =  1.02x the measured pass
         gemm alone: nofold  6.8906  fold  7.0168  epilogue cost  +126.2 us (+32.9% of the pass it replaces)

  in_proj_qkv k=5120 n=10240 M=2048   output 41.9 MB (INSIDE the 72 MB L2)   [2550, 442.94, 65, Active]
    standalone had_r_128 pass   0.0840 ms    998.1 GB/s over    83.9 MB   analytic ideal saving 0.0932 ms
    s27  128x128x128 1x4w st2 sw8   chain nofold  0.6287 fold  0.5587  paired 1.1229x [1.1052-1.1322]  saved    70.1 us =  0.75x analytic =  0.83x the measured pass
         gemm alone: nofold  0.4714  fold  0.4844  epilogue cost   +13.1 us (+15.6% of the pass it replaces)
    s12  128x128x64  1x4w st3 sw8   chain nofold  0.6618 fold  0.6091  paired 1.0963x [1.0800-1.1000]  saved    52.7 us =  0.57x analytic =  0.63x the measured pass
         gemm alone: nofold  0.5193  fold  0.5357  epilogue cost   +16.5 us (+19.6% of the pass it replaces)
    s10  128x128x64  1x4w st3 sw1   chain nofold  0.6728 fold  0.6038  paired 1.1101x [1.1026-1.1195]  saved    68.9 us =  0.74x analytic =  0.82x the measured pass
         gemm alone: nofold  0.5210  fold  0.5296  epilogue cost    +8.6 us (+10.3% of the pass it replaces)

  in_proj_qkv k=5120 n=10240 M=8192   output 167.8 MB (past the 72 MB L2)   [2565, 450.11, 64, Active]
    standalone had_r_128 pass   0.3843 ms    873.0 GB/s over   335.5 MB   analytic ideal saving 0.3728 ms
    s27  128x128x128 1x4w st2 sw8   chain nofold  2.6190 fold  2.3129  paired 1.1324x [1.1123-1.1513]  saved   306.2 us =  0.82x analytic =  0.80x the measured pass
         gemm alone: nofold  1.8964  fold  1.9504  epilogue cost   +54.0 us (+14.1% of the pass it replaces)
    s12  128x128x64  1x4w st3 sw8   chain nofold  2.8448 fold  2.5295  paired 1.1487x [1.0944-1.1611]  saved   315.4 us =  0.85x analytic =  0.82x the measured pass
         gemm alone: nofold  2.0581  fold  2.1518  epilogue cost   +93.7 us (+24.4% of the pass it replaces)
    s10  128x128x64  1x4w st3 sw1   chain nofold  2.8474 fold  2.5178  paired 1.1481x [1.1016-1.1544]  saved   329.6 us =  0.88x analytic =  0.86x the measured pass
         gemm alone: nofold  2.0753  fold  2.1630  epilogue cost   +87.7 us (+22.8% of the pass it replaces)

  in_proj_qkv k=5120 n=10240 M=16384   output 335.5 MB (past the 72 MB L2)   [2625, 447.68, 64, Active]
    standalone had_r_128 pass   0.7668 ms    875.2 GB/s over   671.1 MB   analytic ideal saving 0.7457 ms
    s27  128x128x128 1x4w st2 sw8   chain nofold  5.2575 fold  4.6762  paired 1.1243x [1.1052-1.1406]  saved   581.3 us =  0.78x analytic =  0.76x the measured pass
         gemm alone: nofold  3.8263  fold  3.9752  epilogue cost  +148.8 us (+19.4% of the pass it replaces)
    s12  128x128x64  1x4w st3 sw8   chain nofold  5.7421 fold  5.0741  paired 1.1520x [1.0897-1.1615]  saved   668.1 us =  0.90x analytic =  0.87x the measured pass
         gemm alone: nofold  4.1783  fold  4.3485  epilogue cost  +170.2 us (+22.2% of the pass it replaces)
    s10  128x128x64  1x4w st3 sw1   chain nofold  5.7149 fold  5.0257  paired 1.1406x [1.0922-1.1570]  saved   689.2 us =  0.92x analytic =  0.90x the measured pass
         gemm alone: nofold  4.1711  fold  4.3668  epilogue cost  +195.7 us (+25.5% of the pass it replaces)
