{ "device": "NVIDIA GeForce RTX 3090", "B": 1, "inc_cap": 1, "infer": { "8": [ { "p50": 9.285, "p95": 10.62 }, { "p50": 9.227, "p95": 15.144 }, { "p50": 10.015, "p95": 18.492 }, { "p50": 10.114, "p95": 15.477 } ], "16": [ { "p50": 9.281, "p95": 12.2 }, { "p50": 9.321, "p95": 15.859 }, { "p50": 11.673, "p95": 17.75 }, { "p50": 9.22, "p95": 11.764 } ] }, "plan": { "8/cb0": [ { "p50": 134.514, "p95": 146.198, "max": 151.847 }, { "p50": 147.029, "p95": 180.064, "max": 191.364 }, { "p50": 133.562, "p95": 145.323, "max": 150.765 }, { "p50": 141.62, "p95": 163.992, "max": 194.319 }, { "p50": 146.427, "p95": 186.324, "max": 189.776 }, { "p50": 144.785, "p95": 164.472, "max": 168.658 }, { "p50": 144.466, "p95": 204.952, "max": 227.804 }, { "p50": 151.671, "p95": 189.823, "max": 241.949 } ], "16/cb0": [ { "p50": 133.572, "p95": 153.984, "max": 156.276 }, { "p50": 144.001, "p95": 170.74, "max": 185.805 }, { "p50": 139.492, "p95": 162.124, "max": 165.545 }, { "p50": 139.614, "p95": 152.373, "max": 153.8 }, { "p50": 137.178, "p95": 158.119, "max": 161.382 }, { "p50": 136.955, "p95": 175.71, "max": 186.645 }, { "p50": 143.252, "p95": 166.76, "max": 182.599 }, { "p50": 149.9, "p95": 200.22, "max": 201.094 } ], "8/cb1": [ { "p50": 151.167, "p95": 164.805, "max": 187.036 }, { "p50": 148.91, "p95": 168.584, "max": 177.316 }, { "p50": 162.24, "p95": 204.403, "max": 222.327 }, { "p50": 158.656, "p95": 197.649, "max": 224.336 }, { "p50": 149.629, "p95": 176.731, "max": 187.429 }, { "p50": 149.44, "p95": 200.478, "max": 243.552 }, { "p50": 150.726, "p95": 200.097, "max": 242.956 }, { "p50": 146.359, "p95": 165.303, "max": 165.497 } ], "16/cb1": [ { "p50": 156.341, "p95": 180.399, "max": 184.267 }, { "p50": 169.067, "p95": 206.577, "max": 218.282 }, { "p50": 169.66, "p95": 190.207, "max": 202.072 }, { "p50": 163.572, "p95": 185.455, "max": 192.382 }, { "p50": 172.023, "p95": 204.316, "max": 225.098 }, { "p50": 166.825, "p95": 201.437, "max": 236.891 }, { "p50": 171.442, "p95": 207.533, "max": 226.472 }, { "p50": 159.815, "p95": 198.534, "max": 207.788 } ] }, "summary": { "infer_p50_K8": 10.015, "infer_p50_K16": 9.321, "plan_p50_8/cb0": 144.785, "plan_p50_16/cb0": 139.614, "plan_p50_8/cb1": 150.726, "plan_p50_16/cb1": 169.067, "plan_p95_8/cb0": 180.064, "plan_p95_16/cb0": 166.76, "plan_p95_8/cb1": 197.649, "plan_p95_16/cb1": 201.437 }, "note": "K16/try3 那批臂是串行认证(cert_batch=0),最坏要 K 次 gaps 调用;部署口径应是批量认证=1 次调用,所以两种口径都要量" }