{
 "date": "2026-10-21",
 "decision": "rejected",
 "change": "per-CTA counts: each source CTA adds its partial counts into FFN inboxes before its per-CTA arrival release; FFN ranks read them after every arrival (no SM0 grid-wide count reduction)",
 "motivation": "globaltimer trace at one token: source SM0 released its arrival 2.1-2.2 us after kernel start, other CTAs 0.45-0.51 us",
 "correctness": "BF16/FP8 small and large and the 3:1 and 1:3 small cases passed",
 "arms": {
  "afd": "per-CTA counts",
  "sm0_counts": "AFDEP_SM0_COUNTS=1 (shipped behavior)",
  "masked": "role_aware=False"
 },
 "screens": [
  {
   "hidden": 2048,
   "processes": [
    {
     "order": [
      "afd",
      "sm0_counts",
      "masked"
     ],
     "median_us": {
      "afd": {
       "1": 21.01456046104431,
       "8": 25.153119564056396,
       "32": 26.751999855041504,
       "128": 36.3155996799469,
       "512": 71.72768115997314,
       "2048": 193.2196807861328
      },
      "sm0_counts": {
       "1": 20.742640495300293,
       "8": 24.737999439239502,
       "32": 26.36080026626587,
       "128": 35.655120611190796,
       "512": 70.5406403541565,
       "2048": 192.48736381530762
      },
      "masked": {
       "1": 31.730000972747803,
       "8": 34.45616006851196,
       "32": 35.5785608291626,
       "128": 44.37359809875488,
       "512": 78.88328075408936,
       "2048": 201.15415573120117
      }
     }
    },
    {
     "order": [
      "sm0_counts",
      "masked",
      "afd"
     ],
     "median_us": {
      "sm0_counts": {
       "1": 20.517680644989014,
       "8": 24.47119951248169,
       "32": 26.74351930618286,
       "128": 35.40223956108093,
       "512": 70.65208196640015,
       "2048": 192.45247840881348
      },
      "masked": {
       "1": 31.50256037712097,
       "8": 35.49088001251221,
       "32": 35.78743934631348,
       "128": 44.48024034500122,
       "512": 79.028000831604,
       "2048": 200.81975936889648
      },
      "afd": {
       "1": 21.160240173339844,
       "8": 25.31983971595764,
       "32": 27.142720222473145,
       "128": 36.74943923950195,
       "512": 72.08832025527954,
       "2048": 193.75255584716797
      }
     }
    },
    {
     "order": [
      "masked",
      "afd",
      "sm0_counts"
     ],
     "median_us": {
      "masked": {
       "1": 31.43984079360962,
       "8": 34.383360147476196,
       "32": 35.673919916152954,
       "128": 44.064719676971436,
       "512": 78.56103897094727,
       "2048": 200.9708023071289
      },
      "afd": {
       "1": 21.008479595184326,
       "8": 25.134159326553345,
       "32": 26.87399983406067,
       "128": 36.74816012382507,
       "512": 72.09455966949463,
       "2048": 193.5209560394287
      },
      "sm0_counts": {
       "1": 20.651999711990356,
       "8": 24.989761114120483,
       "32": 26.616320610046387,
       "128": 35.37495970726013,
       "512": 70.46296119689941,
       "2048": 191.91591262817383
      }
     }
    }
   ],
   "native_source_sha256": "8ec0f5dc07490bd82a842b96140ab6ca2c2f0c7cab8084ed40dff536ab4e8edb",
   "table": [
    {
     "tokens": 1,
     "afd": {
      "median_us": 21.015,
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "sm0_counts": {
      "median_us": 20.652,
      "geomean_change_pct": -2.012,
      "repeats_faster": 3
     },
     "masked": {
      "median_us": 31.503,
      "geomean_change_pct": 49.837,
      "repeats_faster": 0
     }
    },
    {
     "tokens": 8,
     "afd": {
      "median_us": 25.153,
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "sm0_counts": {
      "median_us": 24.738,
      "geomean_change_pct": -1.866,
      "repeats_faster": 3
     },
     "masked": {
      "median_us": 34.456,
      "geomean_change_pct": 37.976,
      "repeats_faster": 0
     }
    },
    {
     "tokens": 32,
     "afd": {
      "median_us": 26.874,
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "sm0_counts": {
      "median_us": 26.616,
      "geomean_change_pct": -1.298,
      "repeats_faster": 3
     },
     "masked": {
      "median_us": 35.674,
      "geomean_change_pct": 32.529,
      "repeats_faster": 0
     }
    },
    {
     "tokens": 128,
     "afd": {
      "median_us": 36.748,
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "sm0_counts": {
      "median_us": 35.402,
      "geomean_change_pct": -3.078,
      "repeats_faster": 3
     },
     "masked": {
      "median_us": 44.374,
      "geomean_change_pct": 21.042,
      "repeats_faster": 0
     }
    },
    {
     "tokens": 512,
     "afd": {
      "median_us": 72.088,
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "sm0_counts": {
      "median_us": 70.541,
      "geomean_change_pct": -1.97,
      "repeats_faster": 3
     },
     "masked": {
      "median_us": 78.883,
      "geomean_change_pct": 9.523,
      "repeats_faster": 0
     }
    },
    {
     "tokens": 2048,
     "afd": {
      "median_us": 193.521,
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "sm0_counts": {
      "median_us": 192.452,
      "geomean_change_pct": -0.627,
      "repeats_faster": 3
     },
     "masked": {
      "median_us": 200.971,
      "geomean_change_pct": 3.868,
      "repeats_faster": 0
     }
    }
   ]
  },
  {
   "hidden": 4096,
   "processes": [
    {
     "order": [
      "afd",
      "sm0_counts",
      "masked"
     ],
     "median_us": {
      "afd": {
       "1": 28.62336039543152,
       "8": 37.73584008216858,
       "32": 44.20543909072876,
       "128": 62.82680034637451,
       "512": 127.94680118560791,
       "2048": 384.09719467163086
      },
      "sm0_counts": {
       "1": 28.390560150146484,
       "8": 39.39568042755127,
       "32": 46.411120891571045,
       "128": 61.63831949234009,
       "512": 127.03887939453125,
       "2048": 383.1898307800293
      },
      "masked": {
       "1": 38.68736028671265,
       "8": 46.0208797454834,
       "32": 54.95728015899658,
       "128": 68.7930417060852,
       "512": 136.1192798614502,
       "2048": 392.4932098388672
      }
     }
    },
    {
     "order": [
      "sm0_counts",
      "masked",
      "afd"
     ],
     "median_us": {
      "sm0_counts": {
       "1": 28.344160318374634,
       "8": 37.83423900604248,
       "32": 42.33824014663696,
       "128": 61.9380784034729,
       "512": 127.21391677856445,
       "2048": 383.54671478271484
      },
      "masked": {
       "1": 38.94032001495361,
       "8": 48.93991947174072,
       "32": 57.38247871398926,
       "128": 68.10887813568115,
       "512": 135.92567920684814,
       "2048": 392.22240447998047
      },
      "afd": {
       "1": 28.744559288024902,
       "8": 38.99503946304321,
       "32": 43.774399757385254,
       "128": 62.67911911010742,
       "512": 127.8762435913086,
       "2048": 384.2973518371582
      }
     }
    },
    {
     "order": [
      "masked",
      "afd",
      "sm0_counts"
     ],
     "median_us": {
      "masked": {
       "1": 38.69047999382019,
       "8": 48.05839776992798,
       "32": 56.05144023895264,
       "128": 68.89655828475952,
       "512": 135.89967727661133,
       "2048": 392.6358413696289
      },
      "afd": {
       "1": 28.516240119934082,
       "8": 41.88159942626953,
       "32": 47.74463891983032,
       "128": 62.4294376373291,
       "512": 128.13047885894775,
       "2048": 384.0600776672363
      },
      "sm0_counts": {
       "1": 28.31055998802185,
       "8": 39.4544792175293,
       "32": 46.73856019973755,
       "128": 61.19823932647705,
       "512": 127.13959693908691,
       "2048": 383.51470947265625
      }
     }
    }
   ],
   "native_source_sha256": "8ec0f5dc07490bd82a842b96140ab6ca2c2f0c7cab8084ed40dff536ab4e8edb",
   "table": [
    {
     "tokens": 1,
     "afd": {
      "median_us": 28.623,
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "sm0_counts": {
      "median_us": 28.344,
      "geomean_change_pct": -0.976,
      "repeats_faster": 3
     },
     "masked": {
      "median_us": 38.69,
      "geomean_change_pct": 35.436,
      "repeats_faster": 0
     }
    },
    {
     "tokens": 8,
     "afd": {
      "median_us": 38.995,
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "sm0_counts": {
      "median_us": 39.396,
      "geomean_change_pct": -1.55,
      "repeats_faster": 2
     },
     "masked": {
      "median_us": 48.058,
      "geomean_change_pct": 20.652,
      "repeats_faster": 0
     }
    },
    {
     "tokens": 32,
     "afd": {
      "median_us": 44.205,
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "sm0_counts": {
      "median_us": 46.411,
      "geomean_change_pct": -0.199,
      "repeats_faster": 2
     },
     "masked": {
      "median_us": 56.051,
      "geomean_change_pct": 24.143,
      "repeats_faster": 0
     }
    },
    {
     "tokens": 128,
     "afd": {
      "median_us": 62.679,
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "sm0_counts": {
      "median_us": 61.638,
      "geomean_change_pct": -1.683,
      "repeats_faster": 3
     },
     "masked": {
      "median_us": 68.793,
      "geomean_change_pct": 9.504,
      "repeats_faster": 0
     }
    },
    {
     "tokens": 512,
     "afd": {
      "median_us": 127.947,
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "sm0_counts": {
      "median_us": 127.14,
      "geomean_change_pct": -0.667,
      "repeats_faster": 3
     },
     "masked": {
      "median_us": 135.926,
      "geomean_change_pct": 6.248,
      "repeats_faster": 0
     }
    },
    {
     "tokens": 2048,
     "afd": {
      "median_us": 384.097,
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "sm0_counts": {
      "median_us": 383.515,
      "geomean_change_pct": -0.191,
      "repeats_faster": 3
     },
     "masked": {
      "median_us": 392.493,
      "geomean_change_pct": 2.16,
      "repeats_faster": 0
     }
    }
   ]
  }
 ],
 "conclusion": "The SM0-count control was faster in every repeat at H2048 (-2.0% at one token, 20.65 vs 21.01 us) and in most H4096 cells: moving the prefix sums behind the arrival and completing remote count atomics before each release costs more than SM0's late release. Not shipped."
}
