{
 "stamp": "2026-09-21T18:22:15Z",
 "box": "benchbox",
 "world_size": 2,
 "torch": "2.13.0+cu130",
 "nccl": "2.29.7",
 "devices": [
  "NVIDIA GeForce RTX 3090",
  "NVIDIA GeForce RTX 3090"
 ],
 "hidden_size": 5120,
 "dtype": "bfloat16",
 "topology": "PHB, no NVLink; card 1 on a four-lane link",
 "env": {
  "NCCL_DEBUG": "INFO"
 },
 "results": [
  {
   "tokens": 1,
   "bytes": 10240,
   "iters": 50,
   "median_us": 50.175998359918594,
   "p95_us": 55.296000093221664,
   "min_us": 47.10400104522705,
   "algbw_GBs": 0.2040816393237903,
   "busbw_GBs": 0.2040816393237903,
   "what": "one decode step"
  },
  {
   "tokens": 4,
   "bytes": 40960,
   "iters": 50,
   "median_us": 82.94399827718735,
   "p95_us": 87.0399996638298,
   "min_us": 80.89599758386612,
   "algbw_GBs": 0.4938271707510091,
   "busbw_GBs": 0.4938271707510091,
   "what": "a 4-token prefill chunk"
  },
  {
   "tokens": 16,
   "bytes": 163840,
   "iters": 50,
   "median_us": 155.64799308776855,
   "p95_us": 160.76800227165222,
   "min_us": 148.47999811172485,
   "algbw_GBs": 1.052631625694088,
   "busbw_GBs": 1.052631625694088,
   "what": "a 16-token prefill chunk"
  },
  {
   "tokens": 64,
   "bytes": 655360,
   "iters": 50,
   "median_us": 423.93600940704346,
   "p95_us": 435.2000057697296,
   "min_us": 413.6959910392761,
   "algbw_GBs": 1.545893685503734,
   "busbw_GBs": 1.545893685503734,
   "what": "a 64-token prefill chunk"
  },
  {
   "tokens": 256,
   "bytes": 2621440,
   "iters": 50,
   "median_us": 1499.135971069336,
   "p95_us": 1546.239972114563,
   "min_us": 1470.4639911651611,
   "algbw_GBs": 1.7486339135269515,
   "busbw_GBs": 1.7486339135269515,
   "what": "a 256-token prefill chunk"
  },
  {
   "tokens": 1024,
   "bytes": 10485760,
   "iters": 50,
   "median_us": 5823.487997055054,
   "p95_us": 5888.000011444092,
   "min_us": 5725.183963775635,
   "algbw_GBs": 1.800597855667027,
   "busbw_GBs": 1.800597855667027,
   "what": "a 1024-token prefill chunk"
  },
  {
   "tokens": 4096,
   "bytes": 41943040,
   "iters": 50,
   "median_us": 23107.07187652588,
   "p95_us": 23227.392196655273,
   "min_us": 22773.759841918945,
   "algbw_GBs": 1.8151603207937954,
   "busbw_GBs": 1.8151603207937954,
   "what": "a 4096-token prefill chunk"
  },
  {
   "tokens": 8192,
   "bytes": 83886080,
   "iters": 50,
   "median_us": 46097.408294677734,
   "p95_us": 46388.22555541992,
   "min_us": 45421.56982421875,
   "algbw_GBs": 1.8197569690633828,
   "busbw_GBs": 1.8197569690633828,
   "what": "a 8192-token prefill chunk"
  },
  {
   "tokens": 16384,
   "bytes": 167772160,
   "iters": 50,
   "median_us": 92035.58349609375,
   "p95_us": 92360.70251464844,
   "min_us": 91054.0771484375,
   "algbw_GBs": 1.8229053766701087,
   "busbw_GBs": 1.8229053766701087,
   "what": "a 16384-token prefill chunk"
  }
 ],
 "decode_cost": {
  "one_allreduce_us": 50.175998359918594,
  "allreduces_per_output_token": 128,
  "note": "2 all-reduces per layer x 64 layers; this is the floor the link puts under a written token, before the model computes anything",
  "per_output_token_ms": 6.42252779006958
 }
}