{
 "schema": "iphone-ane-vs-gpu-ab/1",
 "date": "2026-09-15",
 "device": "iPhone 17 Pro, iOS 27.0 (24A437), USB",
 "app": "AppleBenchRunnerGA (Apple llm-benchmark method: splitmix64 random prompt, greedy, 1 warmup + n timed trials, genTps excludes the first token) on the unmodified Apple main 7359dbc package; all bundles embedded in ONE app, picked per launch by AB_MODEL, order qwen3_1_7b_ane qwen3_1_7b_gpu_int4lin repeated per round, 20 s idle between launches",
 "protocol": {
  "p": 128,
  "g": 256,
  "n": 5
 },
 "note": "Same-day 3-arm interleave A-B-C-A-B-C, 20 s between launches, no-console launches with the app mirroring its log to the container. The ANE arm's round 1 came ~40 min after a 795 s cold ANE program build of a sibling bundle; its round 2 is the steadier number. Weight bits: 4 vs 4 (equal byte width).",
 "arms": {
  "qwen3_1_7b_ane": {
   "bundle": "ANE: Apple stock iOS static export, 4-bit k-means g32 (Apple iOS default; FAILS the device gate on the chat turn \u2014 speed only), AOT h18p, 31/31 ANE regions (1.1 GB)",
   "runs": [
    {
     "round": 1,
     "prompt_tps": 1146.3,
     "gen_tps": 34.36,
     "load_s": 0.278,
     "footprint_gb": 1.74,
     "trials_prompt_tps": [
      805.7,
      1253.9,
      1143.7,
      1264.0,
      1264.1
     ],
     "trials_gen_tps": [
      34.91,
      33.81,
      34.18,
      34.43,
      34.5
     ]
    },
    {
     "round": 2,
     "prompt_tps": 2211.4,
     "gen_tps": 55.47,
     "load_s": 0.235,
     "footprint_gb": 1.75,
     "trials_prompt_tps": [
      2202.4,
      2029.0,
      2350.3,
      2240.5,
      2234.8
     ],
     "trials_gen_tps": [
      53.19,
      56.12,
      56.26,
      56.07,
      55.71
     ]
    }
   ]
  },
  "qwen3_1_7b_gpu_int4lin": {
   "bundle": "GPU: the same Apple 4bit preset (int4 per-block-32 linear) exported 2026-09-15 with coreai-torch 0.4.2 as a dynamic IR (934 MB), pipelined engine, specialized on first load",
   "runs": [
    {
     "round": 1,
     "prompt_tps": 1162.3,
     "gen_tps": 63.39,
     "load_s": 3.889,
     "footprint_gb": 0.31,
     "trials_prompt_tps": [
      1318.5,
      1177.1,
      1121.9,
      1094.1,
      1100.0
     ],
     "trials_gen_tps": [
      65.78,
      65.25,
      62.94,
      62.17,
      60.8
     ]
    },
    {
     "round": 2,
     "prompt_tps": 1160.0,
     "gen_tps": 60.88,
     "load_s": 0.972,
     "footprint_gb": 0.31,
     "trials_prompt_tps": [
      1306.8,
      1155.1,
      1172.7,
      1108.5,
      1056.9
     ],
     "trials_gen_tps": [
      65.87,
      63.44,
      62.8,
      62.24,
      50.07
     ]
    }
   ]
  }
 }
}