{
  "target": "Hexagon v81 (SM8850, Snapdragon 8 Elite Gen 5)",
  "method": "AI Hub per-graph profile on real v81 silicon (no Gen 5 device on QDC for end-to-end runs)",
  "qairt": "2.45.0.260326154327",
  "graphs": {
    "part2_token_ar1_cl1024":   {"latency_ms": 15.326, "peak_mem_bytes": [77946880, 88260320]},
    "part2_prompt_ar128_cl1024":{"latency_ms": 22.531, "peak_mem_bytes": [64045056, 74921952]},
    "part1_token_ar1_cl1024":   {"latency_ms": 0.040,  "peak_mem_bytes": [20480, 10073728]}
  },
  "derived": {"decode_tok_s": 65.1, "prefill_tok_s": 5681},
  "caveat": "per-graph latency, not end-to-end. On v79 the same method yields 65.8 tok/s against 72.0 measured end-to-end, so it understates by roughly 9%."
}