[
 {
  "node": "M2M32c",
  "chip": "Apple M2 Max",
  "gpu_cores": 30,
  "cpu_threads": "12",
  "mem_gb": 32,
  "macos": "26.5.1",
  "tests_metallib": "All tests passed (6411 assertions in 6 test cases)",
  "tests_runtime_msl": "All tests passed (6405 assertions in 5 test cases)",
  "cold_compile_s": 0.048325125,
  "cold_first_call_s": 0.0498165,
  "warm_compile_s": 0.000524916,
  "library": "runtime",
  "n_bench_records": 1242,
  "realized_fused@1000x3900_cpu1_ms": 28.88922893907875,
  "realized_fused@1000x3900_cpuMT_ms": 3.611000021919608,
  "realized_fused@1000x3900_metal_ms": 1.0015630396082997,
  "realized_fused@1000x3900_speedup_MT": 3.6053646941003636,
  "realized_fused@100x39000_cpu1_ms": 28.890353976748884,
  "realized_fused@100x39000_cpuMT_ms": 3.802749910391867,
  "realized_fused@100x39000_metal_ms": 1.1218334548175335,
  "realized_fused@100x39000_speedup_MT": 3.3897633325709524,
  "regression_stats@1000x3900_cpu1_ms": 47.923104488290846,
  "regression_stats@1000x3900_cpuMT_ms": 5.630020517855883,
  "regression_stats@1000x3900_metal_ms": 4.155708011239767,
  "regression_stats@1000x3900_speedup_MT": 1.35476806903387,
  "garch_nll_P64@1000x390_cpu1_ms": 124.73270844202489,
  "garch_nll_P64@1000x390_cpuMT_ms": 14.9338124319911,
  "garch_nll_P64@1000x390_metal_ms": 1.9055624725297093,
  "garch_nll_P64@1000x390_speedup_MT": 7.836957668549106,
  "garch_nll_P64@100x3900_cpu1_ms": 122.93493747711182,
  "garch_nll_P64@100x3900_cpuMT_ms": 14.713853946886957,
  "garch_nll_P64@100x3900_metal_ms": 2.7148124063387513,
  "garch_nll_P64@100x3900_speedup_MT": 5.419841869195797,
  "rolling_sumsq@1000x3900_cpu1_ms": 9.976478992030025,
  "rolling_sumsq@1000x3900_cpuMT_ms": 1.405874965712428,
  "rolling_sumsq@1000x3900_metal_ms": 2.717687515541911,
  "rolling_sumsq@1000x3900_speedup_MT": 0.5173055981132895
 },
 {
  "node": "M2U64",
  "chip": "Apple M2 Ultra",
  "gpu_cores": 60,
  "cpu_threads": "24",
  "mem_gb": 64,
  "macos": "26.5",
  "tests_metallib": "All tests passed (6411 assertions in 6 test cases)",
  "tests_runtime_msl": "All tests passed (6405 assertions in 5 test cases)",
  "cold_compile_s": 0.049152792,
  "cold_first_call_s": 0.051035459,
  "warm_compile_s": 0.000613,
  "library": "runtime",
  "n_bench_records": 1242,
  "realized_fused@1000x3900_cpu1_ms": 28.897770447656512,
  "realized_fused@1000x3900_cpuMT_ms": 2.279437379911542,
  "realized_fused@1000x3900_metal_ms": 1.0434375144541264,
  "realized_fused@1000x3900_speedup_MT": 2.1845461259882226,
  "realized_fused@100x39000_cpu1_ms": 28.89868791680783,
  "realized_fused@100x39000_cpuMT_ms": 2.103875041939318,
  "realized_fused@100x39000_metal_ms": 1.1007915018126369,
  "realized_fused@100x39000_speedup_MT": 1.9112384484027511,
  "regression_stats@1000x3900_cpu1_ms": 47.90822905488312,
  "regression_stats@1000x3900_cpuMT_ms": 3.968646051362157,
  "regression_stats@1000x3900_metal_ms": 3.6643960047513247,
  "regression_stats@1000x3900_speedup_MT": 1.0830287027429175,
  "garch_nll_P64@1000x390_cpu1_ms": 124.76091645658016,
  "garch_nll_P64@1000x390_cpuMT_ms": 7.946583558805287,
  "garch_nll_P64@1000x390_metal_ms": 1.6135419718921185,
  "garch_nll_P64@1000x390_speedup_MT": 4.924931422444954,
  "garch_nll_P64@100x3900_cpu1_ms": 122.93499964289367,
  "garch_nll_P64@100x3900_cpuMT_ms": 8.019687607884407,
  "garch_nll_P64@100x3900_metal_ms": 2.407687483355403,
  "garch_nll_P64@100x3900_speedup_MT": 3.330867341930941,
  "rolling_sumsq@1000x3900_cpu1_ms": 10.034603998064995,
  "rolling_sumsq@1000x3900_cpuMT_ms": 0.9129793616011739,
  "rolling_sumsq@1000x3900_metal_ms": 2.838708576746285,
  "rolling_sumsq@1000x3900_speedup_MT": 0.32161785435813445
 },
 {
  "node": "M3BA24",
  "chip": "Apple M3",
  "gpu_cores": 10,
  "cpu_threads": "8",
  "mem_gb": 24,
  "macos": "26.5.1",
  "tests_metallib": "All tests passed (6411 assertions in 6 test cases)",
  "tests_runtime_msl": "All tests passed (6405 assertions in 5 test cases)",
  "cold_compile_s": 0.043905667,
  "cold_first_call_s": 0.04613775,
  "warm_compile_s": 0.000421292,
  "library": "runtime",
  "n_bench_records": 1242,
  "realized_fused@1000x3900_cpu1_ms": 25.151999667286873,
  "realized_fused@1000x3900_cpuMT_ms": 4.603916313499212,
  "realized_fused@1000x3900_metal_ms": 1.7131036147475243,
  "realized_fused@1000x3900_speedup_MT": 2.687471016852494,
  "realized_fused@100x39000_cpu1_ms": 25.133228860795498,
  "realized_fused@100x39000_cpuMT_ms": 4.742853809148073,
  "realized_fused@100x39000_metal_ms": 2.1543754264712334,
  "realized_fused@100x39000_speedup_MT": 2.20149828617227,
  "regression_stats@1000x3900_cpu1_ms": 41.52868688106537,
  "regression_stats@1000x3900_cpuMT_ms": 10.003979317843914,
  "regression_stats@1000x3900_metal_ms": 5.957312881946564,
  "regression_stats@1000x3900_speedup_MT": 1.6792771365359433,
  "garch_nll_P64@1000x390_cpu1_ms": 107.70181193947792,
  "garch_nll_P64@1000x390_cpuMT_ms": 21.04887505993247,
  "garch_nll_P64@1000x390_metal_ms": 3.3908747136592865,
  "garch_nll_P64@1000x390_speedup_MT": 6.207505979251421,
  "garch_nll_P64@100x3900_cpu1_ms": 106.21335497125983,
  "garch_nll_P64@100x3900_cpuMT_ms": 20.496458280831575,
  "garch_nll_P64@100x3900_metal_ms": 3.9809374138712883,
  "garch_nll_P64@100x3900_speedup_MT": 5.148651222049648,
  "rolling_sumsq@1000x3900_cpu1_ms": 8.691520430147648,
  "rolling_sumsq@1000x3900_cpuMT_ms": 2.0735212601721287,
  "rolling_sumsq@1000x3900_metal_ms": 5.833750125020742,
  "rolling_sumsq@1000x3900_speedup_MT": 0.3554353915980856
 },
 {
  "node": "M3U96b",
  "chip": "Apple M3 Ultra",
  "gpu_cores": 60,
  "cpu_threads": "28",
  "mem_gb": 96,
  "macos": "26.4",
  "tests_metallib": "All tests passed (6411 assertions in 6 test cases)",
  "tests_runtime_msl": "All tests passed (6405 assertions in 5 test cases)",
  "cold_compile_s": 0.047089167,
  "cold_first_call_s": 0.053993666,
  "warm_compile_s": 0.000531084,
  "library": "runtime",
  "n_bench_records": 1242,
  "realized_fused@1000x3900_cpu1_ms": 23.458520998246968,
  "realized_fused@1000x3900_cpuMT_ms": 2.304041408933699,
  "realized_fused@1000x3900_metal_ms": 1.370021142065525,
  "realized_fused@1000x3900_speedup_MT": 1.6817560971796317,
  "realized_fused@100x39000_cpu1_ms": 23.477750015445054,
  "realized_fused@100x39000_cpuMT_ms": 2.210604608990252,
  "realized_fused@100x39000_metal_ms": 1.3697290560230613,
  "realized_fused@100x39000_speedup_MT": 1.613899186316913,
  "regression_stats@1000x3900_cpu1_ms": 40.60522955842316,
  "regression_stats@1000x3900_cpuMT_ms": 2.809292054735124,
  "regression_stats@1000x3900_metal_ms": 3.616687492467463,
  "regression_stats@1000x3900_speedup_MT": 0.7767583073146588,
  "garch_nll_P64@1000x390_cpu1_ms": 100.92091700062156,
  "garch_nll_P64@1000x390_cpuMT_ms": 5.491103976964951,
  "garch_nll_P64@1000x390_metal_ms": 1.7272288678213954,
  "garch_nll_P64@1000x390_speedup_MT": 3.179140922934574,
  "garch_nll_P64@100x3900_cpu1_ms": 98.59843761660159,
  "garch_nll_P64@100x3900_cpuMT_ms": 5.446937517262995,
  "garch_nll_P64@100x3900_metal_ms": 2.5334999663755298,
  "garch_nll_P64@100x3900_speedup_MT": 2.1499654981466136,
  "rolling_sumsq@1000x3900_cpu1_ms": 7.585395942442119,
  "rolling_sumsq@1000x3900_cpuMT_ms": 0.6446875631809235,
  "rolling_sumsq@1000x3900_metal_ms": 2.651687478646636,
  "rolling_sumsq@1000x3900_speedup_MT": 0.24312350847240793
 },
 {
  "node": "M4M64c",
  "chip": "Apple M4 Max",
  "gpu_cores": 40,
  "cpu_threads": "16",
  "mem_gb": 64,
  "macos": "26.4",
  "tests_metallib": "All tests passed (6411 assertions in 6 test cases)",
  "tests_runtime_msl": "All tests passed (6405 assertions in 5 test cases)",
  "cold_compile_s": 0.040168875,
  "cold_first_call_s": 0.044207083,
  "warm_compile_s": 0.0004675,
  "library": "runtime",
  "n_bench_records": 1242,
  "realized_fused@1000x3900_cpu1_ms": 20.198437036015093,
  "realized_fused@1000x3900_cpuMT_ms": 2.066000597551465,
  "realized_fused@1000x3900_metal_ms": 0.7389584789052606,
  "realized_fused@1000x3900_speedup_MT": 2.79582771769825,
  "realized_fused@100x39000_cpu1_ms": 20.19377041142434,
  "realized_fused@100x39000_cpuMT_ms": 2.0329373655840755,
  "realized_fused@100x39000_metal_ms": 0.6558125605806708,
  "realized_fused@100x39000_speedup_MT": 3.09987561656957,
  "regression_stats@1000x3900_cpu1_ms": 30.77616647351533,
  "regression_stats@1000x3900_cpuMT_ms": 3.5077084321528673,
  "regression_stats@1000x3900_metal_ms": 2.94706248678267,
  "regression_stats@1000x3900_speedup_MT": 1.1902389066688093,
  "garch_nll_P64@1000x390_cpu1_ms": 79.52589588239789,
  "garch_nll_P64@1000x390_cpuMT_ms": 7.422207971103489,
  "garch_nll_P64@1000x390_metal_ms": 1.5402500284835696,
  "garch_nll_P64@1000x390_speedup_MT": 4.8188331984067,
  "garch_nll_P64@100x3900_cpu1_ms": 78.29858351033181,
  "garch_nll_P64@100x3900_cpuMT_ms": 7.520229439251125,
  "garch_nll_P64@100x3900_metal_ms": 1.5180624322965741,
  "garch_nll_P64@100x3900_speedup_MT": 4.953834097504329,
  "rolling_sumsq@1000x3900_cpu1_ms": 6.132583483122289,
  "rolling_sumsq@1000x3900_cpuMT_ms": 0.787562457844615,
  "rolling_sumsq@1000x3900_metal_ms": 1.6471249982714653,
  "rolling_sumsq@1000x3900_speedup_MT": 0.4781437102048132
 },
 {
  "node": "M5M36",
  "chip": "Apple M5 Max",
  "gpu_cores": 32,
  "cpu_threads": "18",
  "mem_gb": 36,
  "macos": "27.0",
  "tests_metallib": "All tests passed (6411 assertions in 6 test cases)",
  "tests_runtime_msl": "All tests passed (6405 assertions in 5 test cases)",
  "cold_compile_s": 0.037059042,
  "cold_first_call_s": 0.040020542,
  "warm_compile_s": 0.000456708,
  "library": "runtime",
  "n_bench_records": 1242,
  "realized_fused@1000x3900_cpu1_ms": 20.459166960790753,
  "realized_fused@1000x3900_cpuMT_ms": 1.6108546406030655,
  "realized_fused@1000x3900_metal_ms": 1.4359585475176573,
  "realized_fused@1000x3900_speedup_MT": 1.1217974525711423,
  "realized_fused@100x39000_cpu1_ms": 19.205792108550668,
  "realized_fused@100x39000_cpuMT_ms": 1.6643749549984932,
  "realized_fused@100x39000_metal_ms": 1.4845620607957244,
  "realized_fused@100x39000_speedup_MT": 1.1211218439102433,
  "regression_stats@1000x3900_cpu1_ms": 28.204479021951556,
  "regression_stats@1000x3900_cpuMT_ms": 2.881812513805926,
  "regression_stats@1000x3900_metal_ms": 4.790271050296724,
  "regression_stats@1000x3900_speedup_MT": 0.6015969625826116,
  "garch_nll_P64@1000x390_cpu1_ms": 77.77429150883108,
  "garch_nll_P64@1000x390_cpuMT_ms": 6.631562486290932,
  "garch_nll_P64@1000x390_metal_ms": 3.09193751309067,
  "garch_nll_P64@1000x390_speedup_MT": 2.1447918847693943,
  "garch_nll_P64@100x3900_cpu1_ms": 79.65574995614588,
  "garch_nll_P64@100x3900_cpuMT_ms": 6.037042010575533,
  "garch_nll_P64@100x3900_metal_ms": 3.4854585537686944,
  "garch_nll_P64@100x3900_speedup_MT": 1.7320653559480452,
  "rolling_sumsq@1000x3900_cpu1_ms": 6.14762504119426,
  "rolling_sumsq@1000x3900_cpuMT_ms": 0.6659789942204952,
  "rolling_sumsq@1000x3900_metal_ms": 2.902582986280322,
  "rolling_sumsq@1000x3900_speedup_MT": 0.22944356711535452
 },
 {
  "node": "m2m8a",
  "chip": "Apple M2",
  "gpu_cores": 10,
  "cpu_threads": "8",
  "mem_gb": 8,
  "macos": "26.5.1",
  "tests_metallib": "All tests passed (6411 assertions in 6 test cases)",
  "tests_runtime_msl": "All tests passed (6405 assertions in 5 test cases)",
  "cold_compile_s": 0.109846416,
  "cold_first_call_s": 0.111059292,
  "warm_compile_s": 0.000537584,
  "library": "runtime",
  "n_bench_records": 1242,
  "realized_fused@1000x3900_cpu1_ms": 28.90385454520583,
  "realized_fused@1000x3900_cpuMT_ms": 5.524458014406264,
  "realized_fused@1000x3900_metal_ms": 1.3844374334439635,
  "realized_fused@1000x3900_speedup_MT": 3.990399191000979,
  "realized_fused@100x39000_cpu1_ms": 28.902563033625484,
  "realized_fused@100x39000_cpuMT_ms": 5.802188068628311,
  "realized_fused@100x39000_metal_ms": 1.4303125208243728,
  "realized_fused@100x39000_speedup_MT": 4.0565876227414766,
  "regression_stats@1000x3900_cpu1_ms": 48.694312339648604,
  "regression_stats@1000x3900_cpuMT_ms": 10.649729054421186,
  "regression_stats@1000x3900_metal_ms": 6.6006670240312815,
  "regression_stats@1000x3900_speedup_MT": 1.6134322509601442,
  "garch_nll_P64@1000x390_cpu1_ms": 124.81114594265819,
  "garch_nll_P64@1000x390_cpuMT_ms": 26.51604195125401,
  "garch_nll_P64@1000x390_metal_ms": 3.975166939198971,
  "garch_nll_P64@1000x390_speedup_MT": 6.670422238065105,
  "garch_nll_P64@100x3900_cpu1_ms": 123.06604196783155,
  "garch_nll_P64@100x3900_cpuMT_ms": 28.19958352483809,
  "garch_nll_P64@100x3900_metal_ms": 4.337312420830131,
  "garch_nll_P64@100x3900_speedup_MT": 6.501626073650671,
  "rolling_sumsq@1000x3900_cpu1_ms": 9.78158344514668,
  "rolling_sumsq@1000x3900_cpuMT_ms": 2.4727500276640058,
  "rolling_sumsq@1000x3900_metal_ms": 3.793541924096644,
  "rolling_sumsq@1000x3900_speedup_MT": 0.6518314749487951
 },
 {
  "node": "m4md",
  "chip": "Apple M4",
  "gpu_cores": 10,
  "cpu_threads": "10",
  "mem_gb": 16,
  "macos": "26.6.2",
  "tests_metallib": "All tests passed (6411 assertions in 6 test cases)",
  "tests_runtime_msl": "All tests passed (6405 assertions in 5 test cases)",
  "cold_compile_s": 0.042053541,
  "cold_first_call_s": 0.044052875,
  "warm_compile_s": 0.000386541,
  "library": "runtime",
  "n_bench_records": 1242,
  "realized_fused@1000x3900_cpu1_ms": 24.37687444034964,
  "realized_fused@1000x3900_cpuMT_ms": 3.558958414942026,
  "realized_fused@1000x3900_metal_ms": 1.3962920056656003,
  "realized_fused@1000x3900_speedup_MT": 2.548863991558486,
  "realized_fused@100x39000_cpu1_ms": 20.8127920050174,
  "realized_fused@100x39000_cpuMT_ms": 3.346604062244296,
  "realized_fused@100x39000_metal_ms": 1.6341664595529437,
  "realized_fused@100x39000_speedup_MT": 2.047896677037308,
  "regression_stats@1000x3900_cpu1_ms": 31.550520565360785,
  "regression_stats@1000x3900_cpuMT_ms": 7.527812034823,
  "regression_stats@1000x3900_metal_ms": 5.624479614198208,
  "regression_stats@1000x3900_speedup_MT": 1.338401514661036,
  "garch_nll_P64@1000x390_cpu1_ms": 85.11762507259846,
  "garch_nll_P64@1000x390_cpuMT_ms": 15.866832924075425,
  "garch_nll_P64@1000x390_metal_ms": 3.7722085835412145,
  "garch_nll_P64@1000x390_speedup_MT": 4.206244849053445,
  "garch_nll_P64@100x3900_cpu1_ms": 89.96483404189348,
  "garch_nll_P64@100x3900_cpuMT_ms": 15.40029130410403,
  "garch_nll_P64@100x3900_metal_ms": 4.2256665183231235,
  "garch_nll_P64@100x3900_speedup_MT": 3.644464426458183,
  "rolling_sumsq@1000x3900_cpu1_ms": 7.056104484945536,
  "rolling_sumsq@1000x3900_cpuMT_ms": 1.5508129727095366,
  "rolling_sumsq@1000x3900_metal_ms": 4.117187927477062,
  "rolling_sumsq@1000x3900_speedup_MT": 0.37666800739402895
 },
 {
  "node": "m4mm",
  "chip": "Apple M4 Pro",
  "gpu_cores": 16,
  "cpu_threads": "12",
  "mem_gb": 48,
  "macos": "26.5.1",
  "tests_metallib": "All tests passed (6411 assertions in 6 test cases)",
  "tests_runtime_msl": "All tests passed (6405 assertions in 5 test cases)",
  "cold_compile_s": 0.038549083,
  "cold_first_call_s": 0.041100125,
  "warm_compile_s": 0.000449708,
  "library": "runtime",
  "n_bench_records": 1242,
  "realized_fused@1000x3900_cpu1_ms": 20.243937498889863,
  "realized_fused@1000x3900_cpuMT_ms": 2.68670788500458,
  "realized_fused@1000x3900_metal_ms": 1.6523334197700024,
  "realized_fused@1000x3900_speedup_MT": 1.6260083182113196,
  "realized_fused@100x39000_cpu1_ms": 20.3015620354563,
  "realized_fused@100x39000_cpuMT_ms": 2.815646003000438,
  "realized_fused@100x39000_metal_ms": 1.6293750377371907,
  "realized_fused@100x39000_speedup_MT": 1.728052742793146,
  "regression_stats@1000x3900_cpu1_ms": 31.50135395117104,
  "regression_stats@1000x3900_cpuMT_ms": 3.711103927344084,
  "regression_stats@1000x3900_metal_ms": 4.070708528161049,
  "regression_stats@1000x3900_speedup_MT": 0.91166043986465,
  "garch_nll_P64@1000x390_cpu1_ms": 82.38412498030812,
  "garch_nll_P64@1000x390_cpuMT_ms": 10.105208493769169,
  "garch_nll_P64@1000x390_metal_ms": 4.723895923234522,
  "garch_nll_P64@1000x390_speedup_MT": 2.1391683174192333,
  "garch_nll_P64@100x3900_cpu1_ms": 82.38689589779824,
  "garch_nll_P64@100x3900_cpuMT_ms": 9.782104520127177,
  "garch_nll_P64@100x3900_metal_ms": 2.7961665764451027,
  "garch_nll_P64@100x3900_speedup_MT": 3.4983983438367336,
  "rolling_sumsq@1000x3900_cpu1_ms": 6.384750013239682,
  "rolling_sumsq@1000x3900_cpuMT_ms": 0.90329151134938,
  "rolling_sumsq@1000x3900_metal_ms": 3.7826255429536104,
  "rolling_sumsq@1000x3900_speedup_MT": 0.23880014056164212
 },
 {
  "node": "m5ma",
  "chip": "Apple M5 Pro",
  "gpu_cores": 16,
  "cpu_threads": "15",
  "mem_gb": 24,
  "macos": "27.0",
  "tests_metallib": "All tests passed (6411 assertions in 6 test cases)",
  "tests_runtime_msl": "All tests passed (6405 assertions in 5 test cases)",
  "cold_compile_s": 0.043711542,
  "cold_first_call_s": 0.046221459,
  "warm_compile_s": 0.000386292,
  "library": "runtime",
  "n_bench_records": 1242,
  "realized_fused@1000x3900_cpu1_ms": 21.05324948206544,
  "realized_fused@1000x3900_cpuMT_ms": 2.040499937720597,
  "realized_fused@1000x3900_metal_ms": 1.8081666203215718,
  "realized_fused@1000x3900_speedup_MT": 1.1284910996519257,
  "realized_fused@100x39000_cpu1_ms": 19.319791463203728,
  "realized_fused@100x39000_cpuMT_ms": 1.8705830443650484,
  "realized_fused@100x39000_metal_ms": 1.7046460416167974,
  "realized_fused@100x39000_speedup_MT": 1.0973439639063518,
  "regression_stats@1000x3900_cpu1_ms": 28.90241635031998,
  "regression_stats@1000x3900_cpuMT_ms": 3.107396070845425,
  "regression_stats@1000x3900_metal_ms": 5.662208539433777,
  "regression_stats@1000x3900_speedup_MT": 0.5487957656812417,
  "garch_nll_P64@1000x390_cpu1_ms": 76.39527041465044,
  "garch_nll_P64@1000x390_cpuMT_ms": 7.461854023858905,
  "garch_nll_P64@1000x390_metal_ms": 4.080208367668092,
  "garch_nll_P64@1000x390_speedup_MT": 1.8287923928070076,
  "garch_nll_P64@100x3900_cpu1_ms": 84.06835445202887,
  "garch_nll_P64@100x3900_cpuMT_ms": 7.236291537992656,
  "garch_nll_P64@100x3900_metal_ms": 2.4006880121305585,
  "garch_nll_P64@100x3900_speedup_MT": 3.0142573718150927,
  "rolling_sumsq@1000x3900_cpu1_ms": 6.243103998713195,
  "rolling_sumsq@1000x3900_cpuMT_ms": 0.767791410908103,
  "rolling_sumsq@1000x3900_metal_ms": 3.9311664877459407,
  "rolling_sumsq@1000x3900_speedup_MT": 0.19530880040350074
 },
 {
  "node": "m6ma",
  "chip": "Apple M6",
  "gpu_cores": 12,
  "cpu_threads": "12",
  "mem_gb": 24,
  "macos": "27.0",
  "tests_metallib": "All tests passed (6411 assertions in 6 test cases)",
  "tests_runtime_msl": "All tests passed (6405 assertions in 5 test cases)",
  "cold_compile_s": 0.039953334,
  "cold_first_call_s": 0.041169,
  "warm_compile_s": 0.000306458,
  "library": "runtime",
  "n_bench_records": 1242,
  "realized_fused@1000x3900_cpu1_ms": 17.909875023178756,
  "realized_fused@1000x3900_cpuMT_ms": 2.4681249633431435,
  "realized_fused@1000x3900_metal_ms": 1.5113543486222625,
  "realized_fused@1000x3900_speedup_MT": 1.6330551240965197,
  "realized_fused@100x39000_cpu1_ms": 17.88749999832362,
  "realized_fused@100x39000_cpuMT_ms": 2.598750521428883,
  "realized_fused@100x39000_metal_ms": 1.4539581025019288,
  "realized_fused@100x39000_speedup_MT": 1.7873627286488027,
  "regression_stats@1000x3900_cpu1_ms": 27.72541716694832,
  "regression_stats@1000x3900_cpuMT_ms": 4.481896059587598,
  "regression_stats@1000x3900_metal_ms": 5.708645912818611,
  "regression_stats@1000x3900_speedup_MT": 0.7851066834472289,
  "garch_nll_P64@1000x390_cpu1_ms": 73.43687547836453,
  "garch_nll_P64@1000x390_cpuMT_ms": 11.244291323237121,
  "garch_nll_P64@1000x390_metal_ms": 3.022104036062956,
  "garch_nll_P64@1000x390_speedup_MT": 3.7206830701585027,
  "garch_nll_P64@100x3900_cpu1_ms": 73.3108960557729,
  "garch_nll_P64@100x3900_cpuMT_ms": 10.704978951252997,
  "garch_nll_P64@100x3900_metal_ms": 3.263645456172526,
  "garch_nll_P64@100x3900_speedup_MT": 3.2800679776679456,
  "rolling_sumsq@1000x3900_cpu1_ms": 6.109478999860585,
  "rolling_sumsq@1000x3900_cpuMT_ms": 1.1178544955328107,
  "rolling_sumsq@1000x3900_metal_ms": 3.2437919871881604,
  "rolling_sumsq@1000x3900_speedup_MT": 0.34461349554716936
 }
]