2 "host": "i9-13980HX, L1d 48KB/core, L2 2MB/core, L3 36MB",
6 "iterations": 40000000,
7 "table_bytes": 2556672,
9 "hot_ns_per_direct": 2.565,
10 "hot_ns_per_table_lookup": 3.456,
11 "hot_lookup_penalty_factor": 1.347,
12 "hot_note": "4096 reused inputs -> touched table subset stays L1-hot; lookups beat arithmetic. Shows the overhead is a cache effect, not an instruction-count effect.",
13 "cold_iterations": 20000000,
15 "cold_ns_direct_work": 2.523,
16 "cold_ns_table_work": 4.782,
17 "cold_realistic_penalty_factor": 1.895,
18 "cold_note": "inputs random across the full table (gen cost subtracted). Construction B: each base-mul reads the per-component h0/h1/h2 tables (7.67 MB total, k=3) at random ciphertext offsets. Desktop L3-resident regime; embedded targets without L3 place the tables in DRAM (see working_set_sweep DRAM penalty).",
21 "working_set_sweep": {
24 {"label": "16KB", "bytes": 16384, "ns_per_access": 1.621},
25 {"label": "48KB", "bytes": 49152, "ns_per_access": 1.722},
26 {"label": "256KB", "bytes": 262144, "ns_per_access": 4.924},
27 {"label": "2MB", "bytes": 2097152, "ns_per_access": 10.253},
28 {"label": "7.67MB", "bytes": 7670000, "ns_per_access": 102.557},
29 {"label": "16MB", "bytes": 16777216, "ns_per_access": 110.140},
30 {"label": "36MB", "bytes": 37748736, "ns_per_access": 111.171},
31 {"label": "64MB", "bytes": 67108864, "ns_per_access": 114.680}
33 "baseline_16KB_ns": 1.621,
34 "table_7p67MB_ns": 102.557,
35 "dram_64MB_ns": 114.680,
36 "penalty_table_vs_L1": 63.281,
37 "penalty_dram_vs_L1": 70.761,
38 "note": "penalty_table_vs_L1 is desktop (table L3-resident). An embedded target with only L1/L2 sees the DRAM-regime penalty (penalty_dram_vs_L1) once the 7.67 MB construction-B tables exceed its cache."