{
 "source": "experiments/results/raw/bench_primitives.jsonl",
 "rows": [
  {
   "kernel": "garch_nll_P1",
   "S": 1,
   "n": 390,
   "elements": 390,
   "backend": "cpu",
   "threads": 1,
   "median_s": 1.7909333109855652e-06,
   "ci_lo": 1.7499551177024841e-06,
   "ci_hi": 1.8328428268432617e-06,
   "device_s": 0.0,
   "elements_per_s": 217763552.44929796,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3255.451648
  },
  {
   "kernel": "garch_nll_P1",
   "S": 1,
   "n": 390,
   "elements": 390,
   "backend": "cpu",
   "threads": 18,
   "median_s": 1.7499551177024841e-06,
   "ci_lo": 1.6670674085617065e-06,
   "ci_hi": 1.8333084881305695e-06,
   "device_s": 0.0,
   "elements_per_s": 222862858.62692922,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3368.943616
  },
  {
   "kernel": "garch_nll_P1",
   "S": 1,
   "n": 390,
   "elements": 390,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0009180000051856041,
   "ci_lo": 0.00044724997133016586,
   "ci_hi": 0.0009358339011669159,
   "device_s": 0.00032679166179150343,
   "elements_per_s": 424836.59890736995,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "garch_nll_P1",
   "S": 1,
   "n": 3900,
   "elements": 3900,
   "backend": "cpu",
   "threads": 1,
   "median_s": 1.6250181943178177e-05,
   "ci_lo": 1.6207806766033173e-05,
   "ci_hi": 1.637497916817665e-05,
   "device_s": 0.0,
   "elements_per_s": 239997312.86930108,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "garch_nll_P1",
   "S": 1,
   "n": 3900,
   "elements": 3900,
   "backend": "cpu",
   "threads": 18,
   "median_s": 1.695798709988594e-05,
   "ci_lo": 1.691700890660286e-05,
   "ci_hi": 1.9416213035583496e-05,
   "device_s": 0.0,
   "elements_per_s": 229980125.41395503,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3431.858176
  },
  {
   "kernel": "garch_nll_P1",
   "S": 1,
   "n": 3900,
   "elements": 3900,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.001348957885056734,
   "ci_lo": 0.0010095001198351383,
   "ci_hi": 0.0015331669710576534,
   "device_s": 0.0011987084290012717,
   "elements_per_s": 2891120.6518771155,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "garch_nll_P1",
   "S": 10,
   "n": 390,
   "elements": 3900,
   "backend": "cpu",
   "threads": 1,
   "median_s": 2.1916814148426056e-05,
   "ci_lo": 2.1541956812143326e-05,
   "ci_hi": 7.208297029137611e-05,
   "device_s": 0.0,
   "elements_per_s": 177945570.62847915,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "garch_nll_P1",
   "S": 10,
   "n": 390,
   "elements": 3900,
   "backend": "cpu",
   "threads": 18,
   "median_s": 2.929195761680603e-05,
   "ci_lo": 1.7000362277030945e-05,
   "ci_hi": 4.200031980872154e-05,
   "device_s": 0.0,
   "elements_per_s": 133142347.50095384,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "garch_nll_P1",
   "S": 10,
   "n": 390,
   "elements": 3900,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0003432091325521469,
   "ci_lo": 0.0003193751908838749,
   "ci_hi": 0.0009956671856343746,
   "device_s": 0.00019762502051889896,
   "elements_per_s": 11363333.98531415,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 2920.169472
  },
  {
   "kernel": "garch_nll_P1",
   "S": 10,
   "n": 3900,
   "elements": 39000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.0002155420370399952,
   "ci_lo": 0.00016279099509119987,
   "ci_hi": 0.00037720799446105957,
   "device_s": 0.0,
   "elements_per_s": 180939182.60948467,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 576.782336
  },
  {
   "kernel": "garch_nll_P1",
   "S": 10,
   "n": 3900,
   "elements": 39000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 6.916606798768044e-05,
   "ci_lo": 6.245775148272514e-05,
   "ci_hi": 7.533328607678413e-05,
   "device_s": 0.0,
   "elements_per_s": 563860302.2358668,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 781.43488
  },
  {
   "kernel": "garch_nll_P1",
   "S": 10,
   "n": 3900,
   "elements": 39000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0014300826005637646,
   "ci_lo": 0.0012115412391722202,
   "ci_hi": 0.0016847080551087856,
   "device_s": 0.0010766249615699053,
   "elements_per_s": 27271152.019208886,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3360.11264
  },
  {
   "kernel": "garch_nll_P1",
   "S": 100,
   "n": 390,
   "elements": 39000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.0001916671171784401,
   "ci_lo": 0.0001916252076625824,
   "ci_hi": 0.0002199173904955387,
   "device_s": 0.0,
   "elements_per_s": 203477782.59580857,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3431.87456
  },
  {
   "kernel": "garch_nll_P1",
   "S": 100,
   "n": 390,
   "elements": 39000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.00016116676852107048,
   "ci_lo": 0.00015066703781485558,
   "ci_hi": 0.00017320923507213593,
   "device_s": 0.0,
   "elements_per_s": 241985369.30335766,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3359.932416
  },
  {
   "kernel": "garch_nll_P1",
   "S": 100,
   "n": 390,
   "elements": 39000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0006238752976059914,
   "ci_lo": 0.0006115422584116459,
   "ci_hi": 0.0006384579464793205,
   "device_s": 0.0004310833755880594,
   "elements_per_s": 62512492.720348835,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3355.901952
  },
  {
   "kernel": "garch_nll_P1",
   "S": 100,
   "n": 3900,
   "elements": 390000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.0016154590994119644,
   "ci_lo": 0.0016071251593530178,
   "ci_hi": 0.0016200002282857895,
   "device_s": 0.0,
   "elements_per_s": 241417439.87326083,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3222.339584
  },
  {
   "kernel": "garch_nll_P1",
   "S": 100,
   "n": 3900,
   "elements": 390000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.00020695803686976433,
   "ci_lo": 0.00020262505859136581,
   "ci_hi": 0.0002222089096903801,
   "device_s": 0.0,
   "elements_per_s": 1884439985.5098226,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "garch_nll_P1",
   "S": 100,
   "n": 3900,
   "elements": 390000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0016493331640958786,
   "ci_lo": 0.0015722080133855343,
   "ci_hi": 0.004429208114743233,
   "device_s": 0.0009533748961985111,
   "elements_per_s": 236459199.68740082,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "garch_nll_P1",
   "S": 1000,
   "n": 390,
   "elements": 390000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.0016390001401305199,
   "ci_lo": 0.001627500168979168,
   "ci_hi": 0.0016581248492002487,
   "device_s": 0.0,
   "elements_per_s": 237949949.14945084,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3359.932416
  },
  {
   "kernel": "garch_nll_P1",
   "S": 1000,
   "n": 390,
   "elements": 390000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.0002765827812254429,
   "ci_lo": 0.00027295900508761406,
   "ci_hi": 0.00033191684633493423,
   "device_s": 0.0,
   "elements_per_s": 1410066086.8042636,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "garch_nll_P1",
   "S": 1000,
   "n": 390,
   "elements": 390000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0011531668715178967,
   "ci_lo": 0.0010962923988699913,
   "ci_hi": 0.0011774590238928795,
   "device_s": 0.00017799995839595795,
   "elements_per_s": 338199101.65009224,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3222.339584
  },
  {
   "kernel": "garch_nll_P1",
   "S": 1000,
   "n": 3900,
   "elements": 3900000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.018310166895389557,
   "ci_lo": 0.01771716633811593,
   "ci_hi": 0.019527375232428312,
   "device_s": 0.0,
   "elements_per_s": 212996420.09172553,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 572.112896
  },
  {
   "kernel": "garch_nll_P1",
   "S": 1000,
   "n": 3900,
   "elements": 3900000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.0014978330582380295,
   "ci_lo": 0.001383374910801649,
   "ci_hi": 0.0015932917594909668,
   "device_s": 0.0,
   "elements_per_s": 2603761466.306366,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3432.742912
  },
  {
   "kernel": "garch_nll_P1",
   "S": 1000,
   "n": 3900,
   "elements": 3900000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0072068748995661736,
   "ci_lo": 0.007158916909247637,
   "ci_hi": 0.009398084133863449,
   "device_s": 0.0010924999369308352,
   "elements_per_s": 541149951.171591,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3360.11264
  },
  {
   "kernel": "garch_nll_P1",
   "S": 10000,
   "n": 390,
   "elements": 3900000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.017741417046636343,
   "ci_lo": 0.01734916679561138,
   "ci_hi": 0.017816499806940556,
   "device_s": 0.0,
   "elements_per_s": 219824605.3146817,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3355.901952
  },
  {
   "kernel": "garch_nll_P1",
   "S": 10000,
   "n": 390,
   "elements": 3900000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.0022433330304920673,
   "ci_lo": 0.002103041857481003,
   "ci_hi": 0.002292540855705738,
   "device_s": 0.0,
   "elements_per_s": 1738484632.905596,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "garch_nll_P1",
   "S": 10000,
   "n": 390,
   "elements": 3900000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.00675466563552618,
   "ci_lo": 0.006540417205542326,
   "ci_hi": 0.007071124855428934,
   "device_s": 0.00046787504106760025,
   "elements_per_s": 577378690.5880197,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "garch_nll_P1",
   "S": 10000,
   "n": 3900,
   "elements": 39000000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.16226391727104783,
   "ci_lo": 0.16133012529462576,
   "ci_hi": 0.16484179068356752,
   "device_s": 0.0,
   "elements_per_s": 240349183.3298581,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3255.451648
  },
  {
   "kernel": "garch_nll_P1",
   "S": 10000,
   "n": 3900,
   "elements": 39000000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.01383341709151864,
   "ci_lo": 0.012941500172019005,
   "ci_hi": 0.014138792175799608,
   "device_s": 0.0,
   "elements_per_s": 2819260038.3538756,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3254.272
  },
  {
   "kernel": "garch_nll_P1",
   "S": 10000,
   "n": 3900,
   "elements": 39000000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.06106641702353954,
   "ci_lo": 0.05950908316299319,
   "ci_hi": 0.0616582497023046,
   "device_s": 0.004054000019095838,
   "elements_per_s": 638648899.0334327,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3355.901952
  },
  {
   "kernel": "garch_nll_P64",
   "S": 1,
   "n": 390,
   "elements": 390,
   "backend": "cpu",
   "threads": 1,
   "median_s": 8.287513628602028e-05,
   "ci_lo": 8.27503390610218e-05,
   "ci_hi": 8.458364754915237e-05,
   "device_s": 0.0,
   "elements_per_s": 4705874.614239238,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 577.159168
  },
  {
   "kernel": "garch_nll_P64",
   "S": 1,
   "n": 390,
   "elements": 390,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.00017224997282028198,
   "ci_lo": 0.00012754183262586594,
   "ci_hi": 0.0002139592543244362,
   "device_s": 0.0,
   "elements_per_s": 2264151.3006617934,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3431.514112
  },
  {
   "kernel": "garch_nll_P64",
   "S": 1,
   "n": 390,
   "elements": 390,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0005012080073356628,
   "ci_lo": 0.00047737499698996544,
   "ci_hi": 0.0006123748607933521,
   "device_s": 0.0003255000337958336,
   "elements_per_s": 778120.0505418383,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3431.46496
  },
  {
   "kernel": "garch_nll_P64",
   "S": 1,
   "n": 3900,
   "elements": 3900,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.0008078329265117645,
   "ci_lo": 0.0008031250908970833,
   "ci_hi": 0.0008203750476241112,
   "device_s": 0.0,
   "elements_per_s": 4827730.923076213,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3431.87456
  },
  {
   "kernel": "garch_nll_P64",
   "S": 1,
   "n": 3900,
   "elements": 3900,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.00014187488704919815,
   "ci_lo": 0.000135000329464674,
   "ci_hi": 0.00021779071539640427,
   "device_s": 0.0,
   "elements_per_s": 27489008.668937948,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3222.339584
  },
  {
   "kernel": "garch_nll_P64",
   "S": 1,
   "n": 3900,
   "elements": 3900,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0014235423877835274,
   "ci_lo": 0.00134983379393816,
   "ci_hi": 0.0033927499316632748,
   "device_s": 0.00122337497305125,
   "elements_per_s": 2739644.4485733556,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3431.87456
  },
  {
   "kernel": "garch_nll_P64",
   "S": 10,
   "n": 390,
   "elements": 3900,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.0008257082663476467,
   "ci_lo": 0.000809917226433754,
   "ci_hi": 0.0008320827037096024,
   "device_s": 0.0,
   "elements_per_s": 4723217.8227436915,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3432.742912
  },
  {
   "kernel": "garch_nll_P64",
   "S": 10,
   "n": 390,
   "elements": 3900,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.0002438751980662346,
   "ci_lo": 0.00013074977323412895,
   "ci_hi": 0.0002598753198981285,
   "device_s": 0.0,
   "elements_per_s": 15991786.089460358,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3368.943616
  },
  {
   "kernel": "garch_nll_P64",
   "S": 10,
   "n": 390,
   "elements": 3900,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0010179998353123665,
   "ci_lo": 0.0004632081836462021,
   "ci_hi": 0.0010751672089099884,
   "device_s": 0.00032300001475960016,
   "elements_per_s": 3831041.877136759,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3431.530496
  },
  {
   "kernel": "garch_nll_P64",
   "S": 10,
   "n": 3900,
   "elements": 39000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.008083874825388193,
   "ci_lo": 0.008056249935179949,
   "ci_hi": 0.00812424998730421,
   "device_s": 0.0,
   "elements_per_s": 4824419.086440666,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3355.901952
  },
  {
   "kernel": "garch_nll_P64",
   "S": 10,
   "n": 3900,
   "elements": 39000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.000775500200688839,
   "ci_lo": 0.0006874999962747097,
   "ci_hi": 0.0008766250684857368,
   "device_s": 0.0,
   "elements_per_s": 50290122.38211957,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3432.742912
  },
  {
   "kernel": "garch_nll_P64",
   "S": 10,
   "n": 3900,
   "elements": 39000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0013822498731315136,
   "ci_lo": 0.0008970410563051701,
   "ci_hi": 0.0033571659587323666,
   "device_s": 0.0006641249638050795,
   "elements_per_s": 28214869.654243305,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3431.530496
  },
  {
   "kernel": "garch_nll_P64",
   "S": 100,
   "n": 390,
   "elements": 39000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.008346542250365019,
   "ci_lo": 0.008285791147500277,
   "ci_hi": 0.008543540723621845,
   "device_s": 0.0,
   "elements_per_s": 4672593.611839012,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3431.514112
  },
  {
   "kernel": "garch_nll_P64",
   "S": 100,
   "n": 390,
   "elements": 39000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.001005750149488449,
   "ci_lo": 0.0009415419772267342,
   "ci_hi": 0.0010966667905449867,
   "device_s": 0.0,
   "elements_per_s": 38777026.30204572,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 781.484032
  },
  {
   "kernel": "garch_nll_P64",
   "S": 100,
   "n": 390,
   "elements": 39000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0005332920700311661,
   "ci_lo": 0.0005230000242590904,
   "ci_hi": 0.0005429168231785297,
   "device_s": 0.00033650000113993883,
   "elements_per_s": 73130658.02332444,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "garch_nll_P64",
   "S": 100,
   "n": 3900,
   "elements": 390000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.0801338329911232,
   "ci_lo": 0.07999316696077585,
   "ci_hi": 0.08033083425834775,
   "device_s": 0.0,
   "elements_per_s": 4866858.172667244,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3359.932416
  },
  {
   "kernel": "garch_nll_P64",
   "S": 100,
   "n": 3900,
   "elements": 390000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.006173874717205763,
   "ci_lo": 0.005890249740332365,
   "ci_hi": 0.006268457975238562,
   "device_s": 0.0,
   "elements_per_s": 63169406.22606451,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3432.742912
  },
  {
   "kernel": "garch_nll_P64",
   "S": 100,
   "n": 3900,
   "elements": 390000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0021704169921576977,
   "ci_lo": 0.0018634586594998837,
   "ci_hi": 0.0022297906689345837,
   "device_s": 0.0010580417001619935,
   "elements_per_s": 179688972.86059555,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3431.858176
  },
  {
   "kernel": "garch_nll_P64",
   "S": 1000,
   "n": 390,
   "elements": 390000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.09679854195564985,
   "ci_lo": 0.08949454175308347,
   "ci_hi": 0.10223195794969797,
   "device_s": 0.0,
   "elements_per_s": 4028986.306205791,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "garch_nll_P64",
   "S": 1000,
   "n": 390,
   "elements": 390000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.00848708301782608,
   "ci_lo": 0.008417499717324972,
   "ci_hi": 0.008939292281866074,
   "device_s": 0.0,
   "elements_per_s": 45952183.94598623,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "garch_nll_P64",
   "S": 1000,
   "n": 390,
   "elements": 390000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.002336665987968445,
   "ci_lo": 0.002312499564141035,
   "ci_hi": 0.0027117920108139515,
   "device_s": 0.001452916651032865,
   "elements_per_s": 166904470.73228282,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3359.932416
  },
  {
   "kernel": "garch_nll_P64",
   "S": 1000,
   "n": 3900,
   "elements": 3900000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.8561791670508683,
   "ci_lo": 0.8505779169499874,
   "ci_hi": 0.9446879159659147,
   "device_s": 0.0,
   "elements_per_s": 4555121.34619399,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "garch_nll_P64",
   "S": 1000,
   "n": 3900,
   "elements": 3900000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.08150566695258021,
   "ci_lo": 0.0744811249896884,
   "ci_hi": 0.10778554202988744,
   "device_s": 0.0,
   "elements_per_s": 47849433.61384933,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3255.451648
  },
  {
   "kernel": "garch_nll_P64",
   "S": 1000,
   "n": 3900,
   "elements": 3900000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.010272707790136337,
   "ci_lo": 0.009604125283658504,
   "ci_hi": 0.010956583078950644,
   "device_s": 0.0029698750004172325,
   "elements_per_s": 379646737.7125929,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "realized_fused",
   "S": 1,
   "n": 390,
   "elements": 390,
   "backend": "cpu",
   "threads": 1,
   "median_s": 1.4998950064182281e-06,
   "ci_lo": 1.416075974702835e-06,
   "ci_hi": 4.041939973831177e-06,
   "device_s": 0.0,
   "elements_per_s": 260018200.16144055,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "realized_fused",
   "S": 1,
   "n": 390,
   "elements": 390,
   "backend": "cpu",
   "threads": 18,
   "median_s": 1.4998950064182281e-06,
   "ci_lo": 1.4579854905605316e-06,
   "ci_hi": 1.6670674085617065e-06,
   "device_s": 0.0,
   "elements_per_s": 260018200.16144055,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "realized_fused",
   "S": 1,
   "n": 390,
   "elements": 390,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0002344171516597271,
   "ci_lo": 0.00019499985501170158,
   "ci_hi": 0.00025337468832731247,
   "device_s": 3.199989441782236e-05,
   "elements_per_s": 1663700.7882687368,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3432.742912
  },
  {
   "kernel": "realized_fused",
   "S": 1,
   "n": 3900,
   "elements": 3900,
   "backend": "cpu",
   "threads": 1,
   "median_s": 1.1624768376350403e-05,
   "ci_lo": 1.1375173926353455e-05,
   "ci_hi": 1.2792181223630905e-05,
   "device_s": 0.0,
   "elements_per_s": 335490555.4879026,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "realized_fused",
   "S": 1,
   "n": 3900,
   "elements": 3900,
   "backend": "cpu",
   "threads": 18,
   "median_s": 1.116609200835228e-05,
   "ci_lo": 1.1124648153781891e-05,
   "ci_hi": 1.2124888598918915e-05,
   "device_s": 0.0,
   "elements_per_s": 349271705.5423496,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3432.742912
  },
  {
   "kernel": "realized_fused",
   "S": 1,
   "n": 3900,
   "elements": 3900,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0006333747878670692,
   "ci_lo": 0.0001420830376446247,
   "ci_hi": 0.000777125358581543,
   "device_s": 2.037500962615013e-05,
   "elements_per_s": 6157491.7011356,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 2913.878016
  },
  {
   "kernel": "realized_fused",
   "S": 1,
   "n": 39000,
   "elements": 39000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.00018258299678564072,
   "ci_lo": 0.0001637078821659088,
   "ci_hi": 0.00037499982863664627,
   "device_s": 0.0,
   "elements_per_s": 213601489.11230472,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "realized_fused",
   "S": 1,
   "n": 39000,
   "elements": 39000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.0001671668142080307,
   "ci_lo": 0.00015920819714665413,
   "ci_hi": 0.00017870916053652763,
   "device_s": 0.0,
   "elements_per_s": 233299893.79032168,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "realized_fused",
   "S": 1,
   "n": 39000,
   "elements": 39000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.000206624623388052,
   "ci_lo": 0.0001971670426428318,
   "ci_hi": 0.00024229194968938828,
   "device_s": 4.3874955736100674e-05,
   "elements_per_s": 188748075.42543435,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3355.901952
  },
  {
   "kernel": "realized_fused",
   "S": 1,
   "n": 390000,
   "elements": 390000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.002032375428825617,
   "ci_lo": 0.002019832842051983,
   "ci_hi": 0.002052249852567911,
   "device_s": 0.0,
   "elements_per_s": 191893679.91196227,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3222.339584
  },
  {
   "kernel": "realized_fused",
   "S": 1,
   "n": 390000,
   "elements": 390000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.0020858747884631157,
   "ci_lo": 0.002067000139504671,
   "ci_hi": 0.0027133747935295105,
   "device_s": 0.0,
   "elements_per_s": 186971913.25053322,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "realized_fused",
   "S": 1,
   "n": 390000,
   "elements": 390000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0009353337809443474,
   "ci_lo": 0.00017445813864469528,
   "ci_hi": 0.0009425836615264416,
   "device_s": 1.2583332136273384e-05,
   "elements_per_s": 416963449.78178984,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 781.1072
  },
  {
   "kernel": "realized_fused",
   "S": 10,
   "n": 390,
   "elements": 3900,
   "backend": "cpu",
   "threads": 1,
   "median_s": 1.1292286217212677e-05,
   "ci_lo": 1.1167023330926895e-05,
   "ci_hi": 1.2415926903486252e-05,
   "device_s": 0.0,
   "elements_per_s": 345368504.214433,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3432.742912
  },
  {
   "kernel": "realized_fused",
   "S": 10,
   "n": 390,
   "elements": 3900,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.00014200014993548393,
   "ci_lo": 0.0001277080737054348,
   "ci_hi": 0.0001649591140449047,
   "device_s": 0.0,
   "elements_per_s": 27464759.732802525,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 577.175552
  },
  {
   "kernel": "realized_fused",
   "S": 10,
   "n": 390,
   "elements": 3900,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0002327919937670231,
   "ci_lo": 0.00019591627642512321,
   "ci_hi": 0.0002467092126607895,
   "device_s": 3.124994691461325e-05,
   "elements_per_s": 16753153.477877328,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3255.451648
  },
  {
   "kernel": "realized_fused",
   "S": 10,
   "n": 3900,
   "elements": 39000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.00018791621550917625,
   "ci_lo": 0.00017224997282028198,
   "ci_hi": 0.00020966678857803345,
   "device_s": 0.0,
   "elements_per_s": 207539300.929012,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3431.858176
  },
  {
   "kernel": "realized_fused",
   "S": 10,
   "n": 3900,
   "elements": 39000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.000135000329464674,
   "ci_lo": 0.00012287497520446777,
   "ci_hi": 0.0001421659253537655,
   "device_s": 0.0,
   "elements_per_s": 288888183.86332357,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3360.096256
  },
  {
   "kernel": "realized_fused",
   "S": 10,
   "n": 3900,
   "elements": 39000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.00016404176130890846,
   "ci_lo": 0.00015954114496707916,
   "ci_hi": 0.0009339582175016403,
   "device_s": 1.6249949112534523e-05,
   "elements_per_s": 237744338.32467064,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3222.339584
  },
  {
   "kernel": "realized_fused",
   "S": 10,
   "n": 39000,
   "elements": 390000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.002572625409811735,
   "ci_lo": 0.0020360001362860203,
   "ci_hi": 0.0035845423117280006,
   "device_s": 0.0,
   "elements_per_s": 151596108.2062624,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "realized_fused",
   "S": 10,
   "n": 39000,
   "elements": 390000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.00036349985748529434,
   "ci_lo": 0.00035787513479590416,
   "ci_hi": 0.00039070891216397285,
   "device_s": 0.0,
   "elements_per_s": 1072902759.0217906,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3369.04192
  },
  {
   "kernel": "realized_fused",
   "S": 10,
   "n": 39000,
   "elements": 390000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0007184999994933605,
   "ci_lo": 0.000268666073679924,
   "ci_hi": 0.0007987502031028271,
   "device_s": 5.52500132471323e-05,
   "elements_per_s": 542797495.1635388,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3355.901952
  },
  {
   "kernel": "realized_fused",
   "S": 10,
   "n": 390000,
   "elements": 3900000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.02089220890775323,
   "ci_lo": 0.020823125261813402,
   "ci_hi": 0.021495667286217213,
   "device_s": 0.0,
   "elements_per_s": 186672458.48535842,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3431.497728
  },
  {
   "kernel": "realized_fused",
   "S": 10,
   "n": 390000,
   "elements": 3900000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.0025267917662858963,
   "ci_lo": 0.0025061247870326042,
   "ci_hi": 0.0027628750540316105,
   "device_s": 0.0,
   "elements_per_s": 1543459200.7289019,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3431.858176
  },
  {
   "kernel": "realized_fused",
   "S": 10,
   "n": 390000,
   "elements": 3900000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0010219174437224865,
   "ci_lo": 0.001003749668598175,
   "ci_hi": 0.0010474580340087414,
   "device_s": 8.649996016174555e-05,
   "elements_per_s": 3816355248.6134977,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3222.339584
  },
  {
   "kernel": "realized_fused",
   "S": 100,
   "n": 390,
   "elements": 39000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.0001871250569820404,
   "ci_lo": 0.00016624992713332176,
   "ci_hi": 0.00021525006741285324,
   "device_s": 0.0,
   "elements_per_s": 208416770.2016683,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "realized_fused",
   "S": 100,
   "n": 390,
   "elements": 39000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.0005520828999578953,
   "ci_lo": 0.0004998329095542431,
   "ci_hi": 0.0005905842408537865,
   "device_s": 0.0,
   "elements_per_s": 70641564.88631389,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3431.530496
  },
  {
   "kernel": "realized_fused",
   "S": 100,
   "n": 390,
   "elements": 39000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.00022633420303463936,
   "ci_lo": 0.00017816666513681412,
   "ci_hi": 0.00027566729113459587,
   "device_s": 3.15000070258975e-05,
   "elements_per_s": 172311561.73966,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3368.943616
  },
  {
   "kernel": "realized_fused",
   "S": 100,
   "n": 3900,
   "elements": 390000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.002043417189270258,
   "ci_lo": 0.0020072502084076405,
   "ci_hi": 0.0020909998565912247,
   "device_s": 0.0,
   "elements_per_s": 190856767.7945766,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3432.742912
  },
  {
   "kernel": "realized_fused",
   "S": 100,
   "n": 3900,
   "elements": 390000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.0006015831604599953,
   "ci_lo": 0.0005330839194357395,
   "ci_hi": 0.0006621670909225941,
   "device_s": 0.0,
   "elements_per_s": 648289423.0313727,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3355.8528
  },
  {
   "kernel": "realized_fused",
   "S": 100,
   "n": 3900,
   "elements": 390000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0006413329392671585,
   "ci_lo": 0.000292791984975338,
   "ci_hi": 0.0008833748288452625,
   "device_s": 5.287490785121918e-05,
   "elements_per_s": 608108481.7593294,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 490.962944
  },
  {
   "kernel": "realized_fused",
   "S": 100,
   "n": 39000,
   "elements": 3900000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.020590624772012234,
   "ci_lo": 0.020551583264023066,
   "ci_hi": 0.020679709035903215,
   "device_s": 0.0,
   "elements_per_s": 189406588.83265492,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3222.339584
  },
  {
   "kernel": "realized_fused",
   "S": 100,
   "n": 39000,
   "elements": 3900000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.0021904585883021355,
   "ci_lo": 0.001996333710849285,
   "ci_hi": 0.002361500170081854,
   "device_s": 0.0,
   "elements_per_s": 1780449089.8971806,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "realized_fused",
   "S": 100,
   "n": 39000,
   "elements": 3900000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0011692089028656483,
   "ci_lo": 0.0008370000869035721,
   "ci_hi": 0.0011960407719016075,
   "device_s": 0.000259208376519382,
   "elements_per_s": 3335588696.289753,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3360.096256
  },
  {
   "kernel": "realized_fused",
   "S": 100,
   "n": 390000,
   "elements": 39000000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.2325868341140449,
   "ci_lo": 0.2128961249254644,
   "ci_hi": 0.24819900002330542,
   "device_s": 0.0,
   "elements_per_s": 167679310.60481706,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "realized_fused",
   "S": 100,
   "n": 390000,
   "elements": 39000000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.01879258267581463,
   "ci_lo": 0.016645667143166065,
   "ci_hi": 0.02031658310443163,
   "device_s": 0.0,
   "elements_per_s": 2075286865.7159925,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "realized_fused",
   "S": 100,
   "n": 390000,
   "elements": 39000000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.008720249868929386,
   "ci_lo": 0.007397250272333622,
   "ci_hi": 0.008947291877120733,
   "device_s": 0.0021910000359639525,
   "elements_per_s": 4472348910.43187,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 486.178816
  },
  {
   "kernel": "realized_fused",
   "S": 1000,
   "n": 390,
   "elements": 390000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.002235791180282831,
   "ci_lo": 0.0021259579807519913,
   "ci_hi": 0.003065959084779024,
   "device_s": 0.0,
   "elements_per_s": 174434895.10083154,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "realized_fused",
   "S": 1000,
   "n": 390,
   "elements": 390000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.0008399169892072678,
   "ci_lo": 0.0007132496684789658,
   "ci_hi": 0.0009002499282360077,
   "device_s": 0.0,
   "elements_per_s": 464331600.6360231,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3355.8528
  },
  {
   "kernel": "realized_fused",
   "S": 1000,
   "n": 390,
   "elements": 390000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.00033845799043774605,
   "ci_lo": 0.00033012498170137405,
   "ci_hi": 0.00035137496888637543,
   "device_s": 0.00013987498823553324,
   "elements_per_s": 1152284806.4410944,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3431.87456
  },
  {
   "kernel": "realized_fused",
   "S": 1000,
   "n": 3900,
   "elements": 3900000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.021244540810585022,
   "ci_lo": 0.021230708342045546,
   "ci_hi": 0.021389917004853487,
   "device_s": 0.0,
   "elements_per_s": 183576573.14282066,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3431.530496
  },
  {
   "kernel": "realized_fused",
   "S": 1000,
   "n": 3900,
   "elements": 3900000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.0017422502860426903,
   "ci_lo": 0.0017107920721173286,
   "ci_hi": 0.0017863749526441097,
   "device_s": 0.0,
   "elements_per_s": 2238484350.522551,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3431.46496
  },
  {
   "kernel": "realized_fused",
   "S": 1000,
   "n": 3900,
   "elements": 3900000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0008750832639634609,
   "ci_lo": 0.0008294158615171909,
   "ci_hi": 0.0015468327328562737,
   "device_s": 0.0002547500189393759,
   "elements_per_s": 4456718761.065055,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3432.742912
  },
  {
   "kernel": "realized_fused",
   "S": 1000,
   "n": 39000,
   "elements": 39000000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.23877166723832488,
   "ci_lo": 0.21975037455558777,
   "ci_hi": 0.2514020833186805,
   "device_s": 0.0,
   "elements_per_s": 163335962.13939813,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "realized_fused",
   "S": 1000,
   "n": 39000,
   "elements": 39000000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.02024729223921895,
   "ci_lo": 0.019154875073581934,
   "ci_hi": 0.029473000206053257,
   "device_s": 0.0,
   "elements_per_s": 1926183488.5979028,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "realized_fused",
   "S": 1000,
   "n": 39000,
   "elements": 39000000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.005891874898225069,
   "ci_lo": 0.005617166869342327,
   "ci_hi": 0.00844891695305705,
   "device_s": 0.0009469999931752682,
   "elements_per_s": 6619285146.693929,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3369.04192
  },
  {
   "kernel": "realized_fused",
   "S": 10000,
   "n": 390,
   "elements": 3900000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.021238334011286497,
   "ci_lo": 0.020878124982118607,
   "ci_hi": 0.021370540838688612,
   "device_s": 0.0,
   "elements_per_s": 183630222.49897087,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3431.858176
  },
  {
   "kernel": "realized_fused",
   "S": 10000,
   "n": 390,
   "elements": 3900000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.002125917002558708,
   "ci_lo": 0.0019880831241607666,
   "ci_hi": 0.002400666009634733,
   "device_s": 0.0,
   "elements_per_s": 1834502473.6647966,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "realized_fused",
   "S": 10000,
   "n": 390,
   "elements": 3900000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0022132499143481255,
   "ci_lo": 0.0018930002115666866,
   "ci_hi": 0.0032122922129929066,
   "device_s": 0.001203041640110314,
   "elements_per_s": 1762114605.6380522,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3431.87456
  },
  {
   "kernel": "realized_fused",
   "S": 10000,
   "n": 3900,
   "elements": 39000000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.2336724577471614,
   "ci_lo": 0.22654250031337142,
   "ci_hi": 0.24493679078295827,
   "device_s": 0.0,
   "elements_per_s": 166900285.87878695,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "realized_fused",
   "S": 10000,
   "n": 3900,
   "elements": 39000000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.01569391693919897,
   "ci_lo": 0.015392750035971403,
   "ci_hi": 0.01590104214847088,
   "device_s": 0.0,
   "elements_per_s": 2485039276.7524476,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3432.742912
  },
  {
   "kernel": "realized_fused",
   "S": 10000,
   "n": 3900,
   "elements": 39000000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0078027499839663506,
   "ci_lo": 0.006132957991212606,
   "ci_hi": 0.008690916933119297,
   "device_s": 0.0019207500154152513,
   "elements_per_s": 4998237811.046104,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 577.1264
  },
  {
   "kernel": "regression_stats",
   "S": 1,
   "n": 390,
   "elements": 390,
   "backend": "cpu",
   "threads": 1,
   "median_s": 3.1669624149799347e-06,
   "ci_lo": 3.125052899122238e-06,
   "ci_hi": 3.2088719308376312e-06,
   "device_s": 0.0,
   "elements_per_s": 123146393.57741508,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "regression_stats",
   "S": 1,
   "n": 390,
   "elements": 390,
   "backend": "cpu",
   "threads": 18,
   "median_s": 3.2088719308376312e-06,
   "ci_lo": 3.2079406082630157e-06,
   "ci_hi": 3.3332034945487976e-06,
   "device_s": 0.0,
   "elements_per_s": 121538038.41532433,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "regression_stats",
   "S": 1,
   "n": 390,
   "elements": 390,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.000894167460501194,
   "ci_lo": 0.00017508305609226227,
   "ci_hi": 0.000988542102277279,
   "device_s": 7.087492849677801e-05,
   "elements_per_s": 436159.91100973333,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3368.910848
  },
  {
   "kernel": "regression_stats",
   "S": 1,
   "n": 3900,
   "elements": 3900,
   "backend": "cpu",
   "threads": 1,
   "median_s": 2.9958784580230713e-05,
   "ci_lo": 2.950010821223259e-05,
   "ci_hi": 3.0708033591508865e-05,
   "device_s": 0.0,
   "elements_per_s": 130178845.859239,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3368.894464
  },
  {
   "kernel": "regression_stats",
   "S": 1,
   "n": 3900,
   "elements": 3900,
   "backend": "cpu",
   "threads": 18,
   "median_s": 3.0207913368940353e-05,
   "ci_lo": 3.008265048265457e-05,
   "ci_hi": 3.0375085771083832e-05,
   "device_s": 0.0,
   "elements_per_s": 129105243.13175379,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "regression_stats",
   "S": 1,
   "n": 3900,
   "elements": 3900,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0003482499159872532,
   "ci_lo": 0.0003069578669965267,
   "ci_hi": 0.0003833337686955929,
   "device_s": 0.00019212497863918543,
   "elements_per_s": 11198854.101497471,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "regression_stats",
   "S": 1,
   "n": 39000,
   "elements": 39000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.00030233291909098625,
   "ci_lo": 0.00030054105445742607,
   "ci_hi": 0.0003049168735742569,
   "device_s": 0.0,
   "elements_per_s": 128996869.13770398,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "regression_stats",
   "S": 1,
   "n": 39000,
   "elements": 39000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.0003838753327727318,
   "ci_lo": 0.0003343336284160614,
   "ci_hi": 0.0006159590557217598,
   "device_s": 0.0,
   "elements_per_s": 101595483.4042161,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 578.486272
  },
  {
   "kernel": "regression_stats",
   "S": 1,
   "n": 39000,
   "elements": 39000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0004085828550159931,
   "ci_lo": 0.0003508748486638069,
   "ci_hi": 0.00041229184716939926,
   "device_s": 0.00018866662867367268,
   "elements_per_s": 95451875.96888623,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3359.932416
  },
  {
   "kernel": "regression_stats",
   "S": 1,
   "n": 390000,
   "elements": 390000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.0029733749106526375,
   "ci_lo": 0.002949208952486515,
   "ci_hi": 0.0029905838891863823,
   "device_s": 0.0,
   "elements_per_s": 131164085.16220291,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3368.943616
  },
  {
   "kernel": "regression_stats",
   "S": 1,
   "n": 390000,
   "elements": 390000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.0030428748577833176,
   "ci_lo": 0.0030017499811947346,
   "ci_hi": 0.003628334030508995,
   "device_s": 0.0,
   "elements_per_s": 128168267.9136231,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 572.112896
  },
  {
   "kernel": "regression_stats",
   "S": 1,
   "n": 390000,
   "elements": 390000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0007901252247393131,
   "ci_lo": 0.0007615000940859318,
   "ci_hi": 0.0008232500404119492,
   "device_s": 9.650003630667925e-05,
   "elements_per_s": 493592645.5564979,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 2920.12032
  },
  {
   "kernel": "regression_stats",
   "S": 10,
   "n": 390,
   "elements": 3900,
   "backend": "cpu",
   "threads": 1,
   "median_s": 3.1250063329935074e-05,
   "ci_lo": 3.095809370279312e-05,
   "ci_hi": 4.149973392486572e-05,
   "device_s": 0.0,
   "elements_per_s": 124799747.08608383,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2920.185856
  },
  {
   "kernel": "regression_stats",
   "S": 10,
   "n": 390,
   "elements": 3900,
   "backend": "cpu",
   "threads": 18,
   "median_s": 4.133302718400955e-05,
   "ci_lo": 2.475036308169365e-05,
   "ci_hi": 4.3540727347135544e-05,
   "device_s": 0.0,
   "elements_per_s": 94355537.58590388,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3359.932416
  },
  {
   "kernel": "regression_stats",
   "S": 10,
   "n": 390,
   "elements": 3900,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0006523327901959419,
   "ci_lo": 0.0003105420619249344,
   "ci_hi": 0.0007279999554157257,
   "device_s": 9.250000584870577e-05,
   "elements_per_s": 5978543.557236411,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "regression_stats",
   "S": 10,
   "n": 3900,
   "elements": 39000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.0002997498959302902,
   "ci_lo": 0.0002982090227305889,
   "ci_hi": 0.000300500076264143,
   "device_s": 0.0,
   "elements_per_s": 130108468.85855077,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "regression_stats",
   "S": 10,
   "n": 3900,
   "elements": 39000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 7.625017315149307e-05,
   "ci_lo": 7.187481969594955e-05,
   "ci_hi": 8.49580392241478e-05,
   "device_s": 0.0,
   "elements_per_s": 511474248.3602653,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3432.742912
  },
  {
   "kernel": "regression_stats",
   "S": 10,
   "n": 3900,
   "elements": 39000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.00036174990236759186,
   "ci_lo": 0.0003430834040045738,
   "ci_hi": 0.0005066245794296265,
   "device_s": 0.00018737500067800283,
   "elements_per_s": 107809289.63560627,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3431.514112
  },
  {
   "kernel": "regression_stats",
   "S": 10,
   "n": 39000,
   "elements": 390000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.003083667252212763,
   "ci_lo": 0.0030186669901013374,
   "ci_hi": 0.003116208128631115,
   "device_s": 0.0,
   "elements_per_s": 126472789.73441304,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3431.530496
  },
  {
   "kernel": "regression_stats",
   "S": 10,
   "n": 39000,
   "elements": 390000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.0005624173209071159,
   "ci_lo": 0.0005049579776823521,
   "ci_hi": 0.0006032092496752739,
   "device_s": 0.0,
   "elements_per_s": 693435257.9521801,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3431.563264
  },
  {
   "kernel": "regression_stats",
   "S": 10,
   "n": 39000,
   "elements": 390000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0006961673498153687,
   "ci_lo": 0.0006541251204907894,
   "ci_hi": 0.0008656666614115238,
   "device_s": 0.0002501250710338354,
   "elements_per_s": 560210127.7852694,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 576.421888
  },
  {
   "kernel": "regression_stats",
   "S": 10,
   "n": 390000,
   "elements": 3900000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.032662583980709314,
   "ci_lo": 0.030701291747391224,
   "ci_hi": 0.03406220907345414,
   "device_s": 0.0,
   "elements_per_s": 119402678.0705213,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3431.858176
  },
  {
   "kernel": "regression_stats",
   "S": 10,
   "n": 390000,
   "elements": 3900000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.006654791999608278,
   "ci_lo": 0.006012291181832552,
   "ci_hi": 0.007818708196282387,
   "device_s": 0.0,
   "elements_per_s": 586043861.3602898,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3431.46496
  },
  {
   "kernel": "regression_stats",
   "S": 10,
   "n": 390000,
   "elements": 3900000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.004117042291909456,
   "ci_lo": 0.002989500295370817,
   "ci_hi": 0.004763709381222725,
   "device_s": 0.000529208336956799,
   "elements_per_s": 947281986.3094499,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "regression_stats",
   "S": 100,
   "n": 390,
   "elements": 39000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.00036137504503130913,
   "ci_lo": 0.00033475039526820183,
   "ci_hi": 0.00036225002259016037,
   "device_s": 0.0,
   "elements_per_s": 107921121.10735561,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "regression_stats",
   "S": 100,
   "n": 390,
   "elements": 39000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.00017625000327825546,
   "ci_lo": 9.350012987852097e-05,
   "ci_hi": 0.00018370803445577621,
   "device_s": 0.0,
   "elements_per_s": 221276591.62892938,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3359.932416
  },
  {
   "kernel": "regression_stats",
   "S": 100,
   "n": 390,
   "elements": 39000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0003363746218383312,
   "ci_lo": 0.00026350002735853195,
   "ci_hi": 0.0005579171702265739,
   "device_s": 0.00019220833200961351,
   "elements_per_s": 115942159.33074829,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3255.451648
  },
  {
   "kernel": "regression_stats",
   "S": 100,
   "n": 3900,
   "elements": 390000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.0029982919804751873,
   "ci_lo": 0.0029679168947041035,
   "ci_hi": 0.003063875250518322,
   "device_s": 0.0,
   "elements_per_s": 130074056.34263493,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3432.742912
  },
  {
   "kernel": "regression_stats",
   "S": 100,
   "n": 3900,
   "elements": 390000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.0004399162717163563,
   "ci_lo": 0.00042741699144244194,
   "ci_hi": 0.00046166684478521347,
   "device_s": 0.0,
   "elements_per_s": 886532335.9792868,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 781.484032
  },
  {
   "kernel": "regression_stats",
   "S": 100,
   "n": 3900,
   "elements": 390000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0005946666933596134,
   "ci_lo": 0.0005814996547996998,
   "ci_hi": 0.0005976245738565922,
   "device_s": 0.00023925001733005047,
   "elements_per_s": 655829566.9741753,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "regression_stats",
   "S": 100,
   "n": 39000,
   "elements": 3900000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.03219050029292703,
   "ci_lo": 0.03135691722854972,
   "ci_hi": 0.03299033408984542,
   "device_s": 0.0,
   "elements_per_s": 121153755.44060485,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3355.901952
  },
  {
   "kernel": "regression_stats",
   "S": 100,
   "n": 39000,
   "elements": 3900000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.0031737093813717365,
   "ci_lo": 0.003017541952431202,
   "ci_hi": 0.0036639999598264694,
   "device_s": 0.0,
   "elements_per_s": 1228845975.2777827,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3355.901952
  },
  {
   "kernel": "regression_stats",
   "S": 100,
   "n": 39000,
   "elements": 3900000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0044694580137729645,
   "ci_lo": 0.0031338329426944256,
   "ci_hi": 0.004542499780654907,
   "device_s": 0.0016885417280718684,
   "elements_per_s": 872589022.6470106,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3431.46496
  },
  {
   "kernel": "regression_stats",
   "S": 100,
   "n": 390000,
   "elements": 39000000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.3013692502863705,
   "ci_lo": 0.2989709582179785,
   "ci_hi": 0.3020119168795645,
   "device_s": 0.0,
   "elements_per_s": 129409354.01651289,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3431.87456
  },
  {
   "kernel": "regression_stats",
   "S": 100,
   "n": 390000,
   "elements": 39000000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.033760249614715576,
   "ci_lo": 0.0324889593757689,
   "ci_hi": 0.03822179138660431,
   "device_s": 0.0,
   "elements_per_s": 1155204728.7885127,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3359.932416
  },
  {
   "kernel": "regression_stats",
   "S": 100,
   "n": 390000,
   "elements": 39000000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.03070899983868003,
   "ci_lo": 0.027952750213444233,
   "ci_hi": 0.031758958008140326,
   "device_s": 0.005229625036008656,
   "elements_per_s": 1269986004.2617507,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 2913.845248
  },
  {
   "kernel": "regression_stats",
   "S": 1000,
   "n": 390,
   "elements": 390000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.0031211250461637974,
   "ci_lo": 0.0030487077310681343,
   "ci_hi": 0.0040359580889344215,
   "device_s": 0.0,
   "elements_per_s": 124954942.28254408,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3432.742912
  },
  {
   "kernel": "regression_stats",
   "S": 1000,
   "n": 390,
   "elements": 390000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.0006971252150833607,
   "ci_lo": 0.0005761669017374516,
   "ci_hi": 0.0007857088930904865,
   "device_s": 0.0,
   "elements_per_s": 559440386.8369108,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "regression_stats",
   "S": 1000,
   "n": 390,
   "elements": 390000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0005814996547996998,
   "ci_lo": 0.0005787923000752926,
   "ci_hi": 0.0009665000252425671,
   "device_s": 0.00020362494979053736,
   "elements_per_s": 670679675.870723,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3431.858176
  },
  {
   "kernel": "regression_stats",
   "S": 1000,
   "n": 3900,
   "elements": 3900000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.03439616598188877,
   "ci_lo": 0.033030624967068434,
   "ci_hi": 0.03468058304861188,
   "device_s": 0.0,
   "elements_per_s": 113384730.20666131,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 781.484032
  },
  {
   "kernel": "regression_stats",
   "S": 1000,
   "n": 3900,
   "elements": 3900000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.003421790897846222,
   "ci_lo": 0.0031987917609512806,
   "ci_hi": 0.003632874693721533,
   "device_s": 0.0,
   "elements_per_s": 1139754040.0422413,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3431.514112
  },
  {
   "kernel": "regression_stats",
   "S": 1000,
   "n": 3900,
   "elements": 3900000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.003796167206019163,
   "ci_lo": 0.0033167502842843533,
   "ci_hi": 0.004350459203124046,
   "device_s": 0.000579125015065074,
   "elements_per_s": 1027352007.5238521,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 781.1072
  },
  {
   "kernel": "regression_stats",
   "S": 1000,
   "n": 39000,
   "elements": 39000000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.3240363751538098,
   "ci_lo": 0.3117389581166208,
   "ci_hi": 0.326583041343838,
   "device_s": 0.0,
   "elements_per_s": 120356858.02709013,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3431.46496
  },
  {
   "kernel": "regression_stats",
   "S": 1000,
   "n": 39000,
   "elements": 39000000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.031240916810929775,
   "ci_lo": 0.029546833131462336,
   "ci_hi": 0.03264595801010728,
   "device_s": 0.0,
   "elements_per_s": 1248362851.7059293,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3431.87456
  },
  {
   "kernel": "regression_stats",
   "S": 1000,
   "n": 39000,
   "elements": 39000000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.026743208058178425,
   "ci_lo": 0.025766875129193068,
   "ci_hi": 0.02711204206570983,
   "device_s": 0.004475083318538964,
   "elements_per_s": 1458314197.5771036,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3360.11264
  },
  {
   "kernel": "regression_stats",
   "S": 10000,
   "n": 390,
   "elements": 3900000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.030644499696791172,
   "ci_lo": 0.030419666320085526,
   "ci_hi": 0.030872833915054798,
   "device_s": 0.0,
   "elements_per_s": 127265905.41820379,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3431.87456
  },
  {
   "kernel": "regression_stats",
   "S": 10000,
   "n": 390,
   "elements": 3900000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.003959249705076218,
   "ci_lo": 0.0036838343366980553,
   "ci_hi": 0.00489212479442358,
   "device_s": 0.0,
   "elements_per_s": 985035117.8910861,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "regression_stats",
   "S": 10000,
   "n": 390,
   "elements": 3900000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.004190749954432249,
   "ci_lo": 0.00392800010740757,
   "ci_hi": 0.005979666020721197,
   "device_s": 0.0011161250295117497,
   "elements_per_s": 930621020.6779948,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3432.742912
  },
  {
   "kernel": "regression_stats",
   "S": 10000,
   "n": 3900,
   "elements": 39000000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.3569141672924161,
   "ci_lo": 0.3479222496971488,
   "ci_hi": 0.35834929207339883,
   "device_s": 0.0,
   "elements_per_s": 109269969.01203898,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "regression_stats",
   "S": 10000,
   "n": 3900,
   "elements": 39000000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.03749012481421232,
   "ci_lo": 0.03214179212227464,
   "ci_hi": 0.03947866568341851,
   "device_s": 0.0,
   "elements_per_s": 1040273943.9591113,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "regression_stats",
   "S": 10000,
   "n": 3900,
   "elements": 39000000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.026574499905109406,
   "ci_lo": 0.026411292143166065,
   "ci_hi": 0.026990290731191635,
   "device_s": 0.004499250091612339,
   "elements_per_s": 1467572301.9909616,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3221.83168
  },
  {
   "kernel": "rolling_sumsq",
   "S": 1,
   "n": 390,
   "elements": 390,
   "backend": "cpu",
   "threads": 1,
   "median_s": 9.578652679920197e-07,
   "ci_lo": 8.330680429935455e-07,
   "ci_hi": 1.0831281542778015e-06,
   "device_s": 0.0,
   "elements_per_s": 407155382.94603795,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3431.497728
  },
  {
   "kernel": "rolling_sumsq",
   "S": 1,
   "n": 390,
   "elements": 390,
   "backend": "cpu",
   "threads": 18,
   "median_s": 7.920898497104645e-07,
   "ci_lo": 7.920898497104645e-07,
   "ci_hi": 8.749775588512421e-07,
   "device_s": 0.0,
   "elements_per_s": 492368384.9029982,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3431.87456
  },
  {
   "kernel": "rolling_sumsq",
   "S": 1,
   "n": 390,
   "elements": 390,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0002839169465005398,
   "ci_lo": 0.0001594587229192257,
   "ci_hi": 0.0008917916566133499,
   "device_s": 1.8750084564089775e-05,
   "elements_per_s": 1373641.146845944,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3253.567488
  },
  {
   "kernel": "rolling_sumsq",
   "S": 1,
   "n": 3900,
   "elements": 3900,
   "backend": "cpu",
   "threads": 1,
   "median_s": 7.4999406933784485e-06,
   "ci_lo": 7.33276829123497e-06,
   "ci_hi": 7.542315870523453e-06,
   "device_s": 0.0,
   "elements_per_s": 520004111.9582764,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3431.530496
  },
  {
   "kernel": "rolling_sumsq",
   "S": 1,
   "n": 3900,
   "elements": 3900,
   "backend": "cpu",
   "threads": 18,
   "median_s": 7.4999406933784485e-06,
   "ci_lo": 7.415656000375748e-06,
   "ci_hi": 8.250121027231216e-06,
   "device_s": 0.0,
   "elements_per_s": 520004111.9582764,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3255.468032
  },
  {
   "kernel": "rolling_sumsq",
   "S": 1,
   "n": 3900,
   "elements": 3900,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0002416251227259636,
   "ci_lo": 0.0001499168574810028,
   "ci_hi": 0.0002560419961810112,
   "device_s": 4.958361387252808e-06,
   "elements_per_s": 16140705.718018986,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "rolling_sumsq",
   "S": 1,
   "n": 39000,
   "elements": 39000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 6.93327747285366e-05,
   "ci_lo": 6.899982690811157e-05,
   "ci_hi": 7.020775228738785e-05,
   "device_s": 0.0,
   "elements_per_s": 562504531.9864868,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3355.8528
  },
  {
   "kernel": "rolling_sumsq",
   "S": 1,
   "n": 39000,
   "elements": 39000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 7.025012746453285e-05,
   "ci_lo": 6.979191675782204e-05,
   "ci_hi": 8.300039917230606e-05,
   "device_s": 0.0,
   "elements_per_s": 555159135.0448426,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 781.1072
  },
  {
   "kernel": "rolling_sumsq",
   "S": 1,
   "n": 39000,
   "elements": 39000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.00016458332538604736,
   "ci_lo": 0.00015858327969908714,
   "ci_hi": 0.00017341692000627518,
   "device_s": 1.9583385437726974e-05,
   "elements_per_s": 236962036.75871435,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "rolling_sumsq",
   "S": 1,
   "n": 390000,
   "elements": 390000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.0006914162077009678,
   "ci_lo": 0.000687750056385994,
   "ci_hi": 0.0007007499225437641,
   "device_s": 0.0,
   "elements_per_s": 564059672.9671572,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3432.742912
  },
  {
   "kernel": "rolling_sumsq",
   "S": 1,
   "n": 390000,
   "elements": 390000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.000718916766345501,
   "ci_lo": 0.0007105418480932713,
   "ci_hi": 0.0012256246991455555,
   "device_s": 0.0,
   "elements_per_s": 542482827.2993312,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "rolling_sumsq",
   "S": 1,
   "n": 390000,
   "elements": 390000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0007903752848505974,
   "ci_lo": 0.000574291218072176,
   "ci_hi": 0.0009351670742034912,
   "device_s": 6.225006654858589e-05,
   "elements_per_s": 493436481.97921664,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "rolling_sumsq",
   "S": 10,
   "n": 390,
   "elements": 3900,
   "backend": "cpu",
   "threads": 1,
   "median_s": 7.1660615503787994e-06,
   "ci_lo": 7.000286132097244e-06,
   "ci_hi": 7.166992872953415e-06,
   "device_s": 0.0,
   "elements_per_s": 544231998.6483852,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3432.742912
  },
  {
   "kernel": "rolling_sumsq",
   "S": 10,
   "n": 390,
   "elements": 3900,
   "backend": "cpu",
   "threads": 18,
   "median_s": 1.800013706088066e-05,
   "ci_lo": 1.5374738723039627e-05,
   "ci_hi": 2.3958273231983185e-05,
   "device_s": 0.0,
   "elements_per_s": 216665016.87233216,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3432.742912
  },
  {
   "kernel": "rolling_sumsq",
   "S": 10,
   "n": 390,
   "elements": 3900,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0002032076008617878,
   "ci_lo": 0.00014958297833800316,
   "ci_hi": 0.0006746253930032253,
   "device_s": 1.4499993994832039e-05,
   "elements_per_s": 19192195.486095995,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 17.170432
  },
  {
   "kernel": "rolling_sumsq",
   "S": 10,
   "n": 3900,
   "elements": 39000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 6.929086521267891e-05,
   "ci_lo": 6.908271461725235e-05,
   "ci_hi": 6.937514990568161e-05,
   "device_s": 0.0,
   "elements_per_s": 562844754.2153615,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "rolling_sumsq",
   "S": 10,
   "n": 3900,
   "elements": 39000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 4.487484693527222e-05,
   "ci_lo": 3.987457603216171e-05,
   "ci_hi": 5.0374772399663925e-05,
   "device_s": 0.0,
   "elements_per_s": 869083744.313465,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "rolling_sumsq",
   "S": 10,
   "n": 3900,
   "elements": 39000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.00016070762649178505,
   "ci_lo": 0.00014404207468032837,
   "ci_hi": 0.00017450004816055298,
   "device_s": 1.9625062122941017e-05,
   "elements_per_s": 242676722.01601195,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "rolling_sumsq",
   "S": 10,
   "n": 39000,
   "elements": 390000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.0007095411419868469,
   "ci_lo": 0.0007064170204102993,
   "ci_hi": 0.000905290711671114,
   "device_s": 0.0,
   "elements_per_s": 549651002.4886332,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3432.742912
  },
  {
   "kernel": "rolling_sumsq",
   "S": 10,
   "n": 39000,
   "elements": 390000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.000132707878947258,
   "ci_lo": 0.00013195769861340523,
   "ci_hi": 0.00013837497681379318,
   "device_s": 0.0,
   "elements_per_s": 2938785572.4451556,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3358.88384
  },
  {
   "kernel": "rolling_sumsq",
   "S": 10,
   "n": 39000,
   "elements": 390000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0007357923313975334,
   "ci_lo": 0.0007097921334207058,
   "ci_hi": 0.0007738331332802773,
   "device_s": 9.062490426003933e-05,
   "elements_per_s": 530040859.8432253,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3431.46496
  },
  {
   "kernel": "rolling_sumsq",
   "S": 10,
   "n": 390000,
   "elements": 3900000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.006913874763995409,
   "ci_lo": 0.00687399972230196,
   "ci_hi": 0.006919416133314371,
   "device_s": 0.0,
   "elements_per_s": 564083113.033748,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3368.943616
  },
  {
   "kernel": "rolling_sumsq",
   "S": 10,
   "n": 390000,
   "elements": 3900000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.0017658751457929611,
   "ci_lo": 0.0013864999637007713,
   "ci_hi": 0.0021367911249399185,
   "device_s": 0.0,
   "elements_per_s": 2208536662.0009346,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "rolling_sumsq",
   "S": 10,
   "n": 390000,
   "elements": 3900000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.002544207964092493,
   "ci_lo": 0.0024965000338852406,
   "ci_hi": 0.0026084999553859234,
   "device_s": 0.00076149997767061,
   "elements_per_s": 1532893558.6407974,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3253.567488
  },
  {
   "kernel": "rolling_sumsq",
   "S": 100,
   "n": 390,
   "elements": 39000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 6.991717964410782e-05,
   "ci_lo": 6.962474435567856e-05,
   "ci_hi": 7.033301517367363e-05,
   "device_s": 0.0,
   "elements_per_s": 557802820.4014759,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3368.943616
  },
  {
   "kernel": "rolling_sumsq",
   "S": 100,
   "n": 390,
   "elements": 39000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.00017983326688408852,
   "ci_lo": 0.00014462508261203766,
   "ci_hi": 0.0002232920378446579,
   "device_s": 0.0,
   "elements_per_s": 216867550.01307648,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "rolling_sumsq",
   "S": 100,
   "n": 390,
   "elements": 39000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0002676667645573616,
   "ci_lo": 0.00017749983817338943,
   "ci_hi": 0.0007679169066250324,
   "device_s": 9.541632607579231e-06,
   "elements_per_s": 145703558.1705259,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "rolling_sumsq",
   "S": 100,
   "n": 3900,
   "elements": 390000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.000806666910648346,
   "ci_lo": 0.0007096659392118454,
   "ci_hi": 0.0011533750221133232,
   "device_s": 0.0,
   "elements_per_s": 483470928.151179,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "rolling_sumsq",
   "S": 100,
   "n": 3900,
   "elements": 390000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.00020312517881393433,
   "ci_lo": 0.00012504169717431068,
   "ci_hi": 0.0002259998582303524,
   "device_s": 0.0,
   "elements_per_s": 1919998309.797161,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 781.484032
  },
  {
   "kernel": "rolling_sumsq",
   "S": 100,
   "n": 3900,
   "elements": 390000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0006820829585194588,
   "ci_lo": 0.0003593750298023224,
   "ci_hi": 0.0008333330042660236,
   "device_s": 9.058334399014711e-05,
   "elements_per_s": 571777956.2277012,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3431.530496
  },
  {
   "kernel": "rolling_sumsq",
   "S": 100,
   "n": 39000,
   "elements": 3900000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.007338124793022871,
   "ci_lo": 0.007089000195264816,
   "ci_hi": 0.008123625069856644,
   "device_s": 0.0,
   "elements_per_s": 531470928.88201374,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "rolling_sumsq",
   "S": 100,
   "n": 39000,
   "elements": 3900000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.0010874588042497635,
   "ci_lo": 0.0009993752464652061,
   "ci_hi": 0.0013045421801507473,
   "device_s": 0.0,
   "elements_per_s": 3586342751.3381577,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "rolling_sumsq",
   "S": 100,
   "n": 39000,
   "elements": 3900000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.002226833254098892,
   "ci_lo": 0.0021508331410586834,
   "ci_hi": 0.002632542047649622,
   "device_s": 0.00043350004125386477,
   "elements_per_s": 1751365978.0413911,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3431.87456
  },
  {
   "kernel": "rolling_sumsq",
   "S": 100,
   "n": 390000,
   "elements": 39000000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.06962170917540789,
   "ci_lo": 0.06953629199415445,
   "ci_hi": 0.06967033399268985,
   "device_s": 0.0,
   "elements_per_s": 560170103.0025239,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3222.339584
  },
  {
   "kernel": "rolling_sumsq",
   "S": 100,
   "n": 390000,
   "elements": 39000000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.009413291700184345,
   "ci_lo": 0.009286124724894762,
   "ci_hi": 0.010024874936789274,
   "device_s": 0.0,
   "elements_per_s": 4143077814.027185,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3355.8528
  },
  {
   "kernel": "rolling_sumsq",
   "S": 100,
   "n": 390000,
   "elements": 39000000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.014852333813905716,
   "ci_lo": 0.014340042136609554,
   "ci_hi": 0.01528970804065466,
   "device_s": 0.0023322500055655837,
   "elements_per_s": 2625849949.823083,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3255.451648
  },
  {
   "kernel": "rolling_sumsq",
   "S": 1000,
   "n": 390,
   "elements": 390000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.0007188748568296432,
   "ci_lo": 0.0007158750668168068,
   "ci_hi": 0.0007281671278178692,
   "device_s": 0.0,
   "elements_per_s": 542514453.3778304,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3431.563264
  },
  {
   "kernel": "rolling_sumsq",
   "S": 1000,
   "n": 390,
   "elements": 390000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.0002710837870836258,
   "ci_lo": 0.0002649589441716671,
   "ci_hi": 0.00028412509709596634,
   "device_s": 0.0,
   "elements_per_s": 1438669586.977882,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3431.858176
  },
  {
   "kernel": "rolling_sumsq",
   "S": 1000,
   "n": 390,
   "elements": 390000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0009075421839952469,
   "ci_lo": 0.000882917083799839,
   "ci_hi": 0.000995416659861803,
   "device_s": 2.3708329536020756e-05,
   "elements_per_s": 429732090.5603684,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "rolling_sumsq",
   "S": 1000,
   "n": 3900,
   "elements": 3900000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.00695441709831357,
   "ci_lo": 0.006937959231436253,
   "ci_hi": 0.006972583010792732,
   "device_s": 0.0,
   "elements_per_s": 560794664.0050883,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3432.742912
  },
  {
   "kernel": "rolling_sumsq",
   "S": 1000,
   "n": 3900,
   "elements": 3900000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.0010273326188325882,
   "ci_lo": 0.0008542081341147423,
   "ci_hi": 0.0010719997808337212,
   "device_s": 0.0,
   "elements_per_s": 3796238850.5017724,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 556.466176
  },
  {
   "kernel": "rolling_sumsq",
   "S": 1000,
   "n": 3900,
   "elements": 3900000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.002144791651517153,
   "ci_lo": 0.0021287081763148308,
   "ci_hi": 0.0034982911311089993,
   "device_s": 0.0007578750373795629,
   "elements_per_s": 1818358439.264379,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3431.497728
  },
  {
   "kernel": "rolling_sumsq",
   "S": 1000,
   "n": 39000,
   "elements": 39000000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.07026320835575461,
   "ci_lo": 0.0701742498204112,
   "ci_hi": 0.07056499971076846,
   "device_s": 0.0,
   "elements_per_s": 555055781.1498778,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "rolling_sumsq",
   "S": 1000,
   "n": 39000,
   "elements": 39000000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.00831229193136096,
   "ci_lo": 0.007511000148952007,
   "ci_hi": 0.008729041088372469,
   "device_s": 0.0,
   "elements_per_s": 4691846764.05062,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3368.894464
  },
  {
   "kernel": "rolling_sumsq",
   "S": 1000,
   "n": 39000,
   "elements": 39000000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.014728832989931107,
   "ci_lo": 0.014530166983604431,
   "ci_hi": 0.01593674998730421,
   "device_s": 0.002570083364844322,
   "elements_per_s": 2647867623.094178,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3368.894464
  },
  {
   "kernel": "rolling_sumsq",
   "S": 10000,
   "n": 390,
   "elements": 3900000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.007047916762530804,
   "ci_lo": 0.007002290803939104,
   "ci_hi": 0.007088833022862673,
   "device_s": 0.0,
   "elements_per_s": 553355002.8192397,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3431.858176
  },
  {
   "kernel": "rolling_sumsq",
   "S": 10000,
   "n": 390,
   "elements": 3900000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.00107216602191329,
   "ci_lo": 0.001034500077366829,
   "ci_hi": 0.0011164997704327106,
   "device_s": 0.0,
   "elements_per_s": 3637496358.111046,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 2962.440192
  },
  {
   "kernel": "rolling_sumsq",
   "S": 10000,
   "n": 390,
   "elements": 3900000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0021627498790621758,
   "ci_lo": 0.002053916919976473,
   "ci_hi": 0.0022570830769836903,
   "device_s": 0.0002617916325107217,
   "elements_per_s": 1803259839.5941842,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3431.858176
  },
  {
   "kernel": "rolling_sumsq",
   "S": 10000,
   "n": 3900,
   "elements": 39000000,
   "backend": "cpu",
   "threads": 1,
   "median_s": 0.06922604190185666,
   "ci_lo": 0.06907229218631983,
   "ci_hi": 0.06939475005492568,
   "device_s": 0.0,
   "elements_per_s": 563371802.4105898,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3360.096256
  },
  {
   "kernel": "rolling_sumsq",
   "S": 10000,
   "n": 3900,
   "elements": 39000000,
   "backend": "cpu",
   "threads": 18,
   "median_s": 0.00654304213821888,
   "ci_lo": 0.006384790875017643,
   "ci_hi": 0.00675958301872015,
   "device_s": 0.0,
   "elements_per_s": 5960530159.540807,
   "reps": 5,
   "precision": "fp64",
   "peak_rss_mb": 3432.742912
  },
  {
   "kernel": "rolling_sumsq",
   "S": 10000,
   "n": 3900,
   "elements": 39000000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.014753291849046946,
   "ci_lo": 0.01457533286884427,
   "ci_hi": 0.01564762508496642,
   "device_s": 0.0025377500569447875,
   "elements_per_s": 2643477835.254739,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 2962.440192
  }
 ],
 "metal_summary": [
  {
   "kernel": "garch_nll_P1",
   "S": 1,
   "n": 390,
   "elements": 390,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0009180000051856041,
   "ci_lo": 0.00044724997133016586,
   "ci_hi": 0.0009358339011669159,
   "device_s": 0.00032679166179150343,
   "elements_per_s": 424836.59890736995,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 2962.440192,
   "speedup_vs_cpu_mt": 0.0019062691806288963,
   "speedup_vs_cpu_1t": 0.00195090773515134,
   "cpu_mt_median_s": 1.7499551177024841e-06,
   "cpu_1t_median_s": 1.7909333109855652e-06,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": true
  },
  {
   "kernel": "garch_nll_P1",
   "S": 1,
   "n": 3900,
   "elements": 3900,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.001348957885056734,
   "ci_lo": 0.0010095001198351383,
   "ci_hi": 0.0015331669710576534,
   "device_s": 0.0011987084290012717,
   "elements_per_s": 2891120.6518771155,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 2962.440192,
   "speedup_vs_cpu_mt": 0.012571176081729732,
   "speedup_vs_cpu_1t": 0.01204647092632898,
   "cpu_mt_median_s": 1.695798709988594e-05,
   "cpu_1t_median_s": 1.6250181943178177e-05,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": true
  },
  {
   "kernel": "garch_nll_P1",
   "S": 10,
   "n": 390,
   "elements": 3900,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0003432091325521469,
   "ci_lo": 0.0003193751908838749,
   "ci_hi": 0.0009956671856343746,
   "device_s": 0.00019762502051889896,
   "elements_per_s": 11363333.98531415,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 2920.169472,
   "speedup_vs_cpu_mt": 0.08534725576498299,
   "speedup_vs_cpu_1t": 0.06385848181092918,
   "cpu_mt_median_s": 2.929195761680603e-05,
   "cpu_1t_median_s": 2.1916814148426056e-05,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": true
  },
  {
   "kernel": "garch_nll_P1",
   "S": 10,
   "n": 3900,
   "elements": 39000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0014300826005637646,
   "ci_lo": 0.0012115412391722202,
   "ci_hi": 0.0016847080551087856,
   "device_s": 0.0010766249615699053,
   "elements_per_s": 27271152.019208886,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3360.11264,
   "speedup_vs_cpu_mt": 0.04836508601699924,
   "speedup_vs_cpu_1t": 0.15071999124737592,
   "cpu_mt_median_s": 6.916606798768044e-05,
   "cpu_1t_median_s": 0.0002155420370399952,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": true
  },
  {
   "kernel": "garch_nll_P1",
   "S": 100,
   "n": 390,
   "elements": 39000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0006238752976059914,
   "ci_lo": 0.0006115422584116459,
   "ci_hi": 0.0006384579464793205,
   "device_s": 0.0004310833755880594,
   "elements_per_s": 62512492.720348835,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3355.901952,
   "speedup_vs_cpu_mt": 0.2583317036906555,
   "speedup_vs_cpu_1t": 0.30722023762429446,
   "cpu_mt_median_s": 0.00016116676852107048,
   "cpu_1t_median_s": 0.0001916671171784401,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": true
  },
  {
   "kernel": "garch_nll_P1",
   "S": 100,
   "n": 3900,
   "elements": 390000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0016493331640958786,
   "ci_lo": 0.0015722080133855343,
   "ci_hi": 0.004429208114743233,
   "device_s": 0.0009533748961985111,
   "elements_per_s": 236459199.68740082,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 2962.440192,
   "speedup_vs_cpu_mt": 0.12547982504384633,
   "speedup_vs_cpu_1t": 0.9794619635248267,
   "cpu_mt_median_s": 0.00020695803686976433,
   "cpu_1t_median_s": 0.0016154590994119644,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": true
  },
  {
   "kernel": "garch_nll_P1",
   "S": 1000,
   "n": 390,
   "elements": 390000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0011531668715178967,
   "ci_lo": 0.0010962923988699913,
   "ci_hi": 0.0011774590238928795,
   "device_s": 0.00017799995839595795,
   "elements_per_s": 338199101.65009224,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3222.339584,
   "speedup_vs_cpu_mt": 0.23984627728802252,
   "speedup_vs_cpu_1t": 1.4213035256320952,
   "cpu_mt_median_s": 0.0002765827812254429,
   "cpu_1t_median_s": 0.0016390001401305199,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": true
  },
  {
   "kernel": "garch_nll_P1",
   "S": 1000,
   "n": 3900,
   "elements": 3900000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0072068748995661736,
   "ci_lo": 0.007158916909247637,
   "ci_hi": 0.009398084133863449,
   "device_s": 0.0010924999369308352,
   "elements_per_s": 541149951.171591,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3360.11264,
   "speedup_vs_cpu_mt": 0.20783391957146272,
   "speedup_vs_cpu_1t": 2.5406528003548057,
   "cpu_mt_median_s": 0.0014978330582380295,
   "cpu_1t_median_s": 0.018310166895389557,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": true
  },
  {
   "kernel": "garch_nll_P1",
   "S": 10000,
   "n": 390,
   "elements": 3900000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.00675466563552618,
   "ci_lo": 0.006540417205542326,
   "ci_hi": 0.007071124855428934,
   "device_s": 0.00046787504106760025,
   "elements_per_s": 577378690.5880197,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 2962.440192,
   "speedup_vs_cpu_mt": 0.33211607376881125,
   "speedup_vs_cpu_1t": 2.6265426009135546,
   "cpu_mt_median_s": 0.0022433330304920673,
   "cpu_1t_median_s": 0.017741417046636343,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": true
  },
  {
   "kernel": "garch_nll_P1",
   "S": 10000,
   "n": 3900,
   "elements": 39000000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.06106641702353954,
   "ci_lo": 0.05950908316299319,
   "ci_hi": 0.0616582497023046,
   "device_s": 0.004054000019095838,
   "elements_per_s": 638648899.0334327,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3355.901952,
   "speedup_vs_cpu_mt": 0.22653068193252948,
   "speedup_vs_cpu_1t": 2.6571710799488897,
   "cpu_mt_median_s": 0.01383341709151864,
   "cpu_1t_median_s": 0.16226391727104783,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": true
  },
  {
   "kernel": "garch_nll_P64",
   "S": 1,
   "n": 390,
   "elements": 390,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0005012080073356628,
   "ci_lo": 0.00047737499698996544,
   "ci_hi": 0.0006123748607933521,
   "device_s": 0.0003255000337958336,
   "elements_per_s": 778120.0505418383,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3431.46496,
   "speedup_vs_cpu_mt": 0.3436696347608925,
   "speedup_vs_cpu_1t": 0.16535078265523034,
   "cpu_mt_median_s": 0.00017224997282028198,
   "cpu_1t_median_s": 8.287513628602028e-05,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": true
  },
  {
   "kernel": "garch_nll_P64",
   "S": 1,
   "n": 3900,
   "elements": 3900,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0014235423877835274,
   "ci_lo": 0.00134983379393816,
   "ci_hi": 0.0033927499316632748,
   "device_s": 0.00122337497305125,
   "elements_per_s": 2739644.4485733556,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3431.87456,
   "speedup_vs_cpu_mt": 0.09966326838366861,
   "speedup_vs_cpu_1t": 0.5674807673058265,
   "cpu_mt_median_s": 0.00014187488704919815,
   "cpu_1t_median_s": 0.0008078329265117645,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": true
  },
  {
   "kernel": "garch_nll_P64",
   "S": 10,
   "n": 390,
   "elements": 3900,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0010179998353123665,
   "ci_lo": 0.0004632081836462021,
   "ci_hi": 0.0010751672089099884,
   "device_s": 0.00032300001475960016,
   "elements_per_s": 3831041.877136759,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3431.530496,
   "speedup_vs_cpu_mt": 0.23956310168891443,
   "speedup_vs_cpu_1t": 0.8111084478655968,
   "cpu_mt_median_s": 0.0002438751980662346,
   "cpu_1t_median_s": 0.0008257082663476467,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": true
  },
  {
   "kernel": "garch_nll_P64",
   "S": 10,
   "n": 3900,
   "elements": 39000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0013822498731315136,
   "ci_lo": 0.0008970410563051701,
   "ci_hi": 0.0033571659587323666,
   "device_s": 0.0006641249638050795,
   "elements_per_s": 28214869.654243305,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3431.530496,
   "speedup_vs_cpu_mt": 0.5610419763916696,
   "speedup_vs_cpu_1t": 5.848345499988377,
   "cpu_mt_median_s": 0.000775500200688839,
   "cpu_1t_median_s": 0.008083874825388193,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": true
  },
  {
   "kernel": "garch_nll_P64",
   "S": 100,
   "n": 390,
   "elements": 39000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0005332920700311661,
   "ci_lo": 0.0005230000242590904,
   "ci_hi": 0.0005429168231785297,
   "device_s": 0.00033650000113993883,
   "elements_per_s": 73130658.02332444,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 2962.440192,
   "speedup_vs_cpu_mt": 1.8859274420294159,
   "speedup_vs_cpu_1t": 15.650977615094181,
   "cpu_mt_median_s": 0.001005750149488449,
   "cpu_1t_median_s": 0.008346542250365019,
   "metal_wins_vs_mt": true,
   "cpu_wins_vs_mt": false
  },
  {
   "kernel": "garch_nll_P64",
   "S": 100,
   "n": 3900,
   "elements": 390000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0021704169921576977,
   "ci_lo": 0.0018634586594998837,
   "ci_hi": 0.0022297906689345837,
   "device_s": 0.0010580417001619935,
   "elements_per_s": 179688972.86059555,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3431.858176,
   "speedup_vs_cpu_mt": 2.84455693975565,
   "speedup_vs_cpu_1t": 36.920938824506244,
   "cpu_mt_median_s": 0.006173874717205763,
   "cpu_1t_median_s": 0.0801338329911232,
   "metal_wins_vs_mt": true,
   "cpu_wins_vs_mt": false
  },
  {
   "kernel": "garch_nll_P64",
   "S": 1000,
   "n": 390,
   "elements": 390000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.002336665987968445,
   "ci_lo": 0.002312499564141035,
   "ci_hi": 0.0027117920108139515,
   "device_s": 0.001452916651032865,
   "elements_per_s": 166904470.73228282,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3359.932416,
   "speedup_vs_cpu_mt": 3.632133587567199,
   "speedup_vs_cpu_1t": 41.4259215711908,
   "cpu_mt_median_s": 0.00848708301782608,
   "cpu_1t_median_s": 0.09679854195564985,
   "metal_wins_vs_mt": true,
   "cpu_wins_vs_mt": false
  },
  {
   "kernel": "garch_nll_P64",
   "S": 1000,
   "n": 3900,
   "elements": 3900000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.010272707790136337,
   "ci_lo": 0.009604125283658504,
   "ci_hi": 0.010956583078950644,
   "device_s": 0.0029698750004172325,
   "elements_per_s": 379646737.7125929,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 2962.440192,
   "speedup_vs_cpu_mt": 7.934195016316967,
   "speedup_vs_cpu_1t": 83.34503273547365,
   "cpu_mt_median_s": 0.08150566695258021,
   "cpu_1t_median_s": 0.8561791670508683,
   "metal_wins_vs_mt": true,
   "cpu_wins_vs_mt": false
  },
  {
   "kernel": "realized_fused",
   "S": 1,
   "n": 390,
   "elements": 390,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0002344171516597271,
   "ci_lo": 0.00019499985501170158,
   "ci_hi": 0.00025337468832731247,
   "device_s": 3.199989441782236e-05,
   "elements_per_s": 1663700.7882687368,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3432.742912,
   "speedup_vs_cpu_mt": 0.006398401293585508,
   "speedup_vs_cpu_1t": 0.006398401293585508,
   "cpu_mt_median_s": 1.4998950064182281e-06,
   "cpu_1t_median_s": 1.4998950064182281e-06,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": true
  },
  {
   "kernel": "realized_fused",
   "S": 1,
   "n": 3900,
   "elements": 3900,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0006333747878670692,
   "ci_lo": 0.0001420830376446247,
   "ci_hi": 0.000777125358581543,
   "device_s": 2.037500962615013e-05,
   "elements_per_s": 6157491.7011356,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 2913.878016,
   "speedup_vs_cpu_mt": 0.017629517660396336,
   "speedup_vs_cpu_1t": 0.018353696103846455,
   "cpu_mt_median_s": 1.116609200835228e-05,
   "cpu_1t_median_s": 1.1624768376350403e-05,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": true
  },
  {
   "kernel": "realized_fused",
   "S": 1,
   "n": 39000,
   "elements": 39000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.000206624623388052,
   "ci_lo": 0.0001971670426428318,
   "ci_hi": 0.00024229194968938828,
   "device_s": 4.3874955736100674e-05,
   "elements_per_s": 188748075.42543435,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3355.901952,
   "speedup_vs_cpu_mt": 0.8090362681222294,
   "speedup_vs_cpu_1t": 0.8836458781717422,
   "cpu_mt_median_s": 0.0001671668142080307,
   "cpu_1t_median_s": 0.00018258299678564072,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": true
  },
  {
   "kernel": "realized_fused",
   "S": 1,
   "n": 390000,
   "elements": 390000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0009353337809443474,
   "ci_lo": 0.00017445813864469528,
   "ci_hi": 0.0009425836615264416,
   "device_s": 1.2583332136273384e-05,
   "elements_per_s": 416963449.78178984,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 781.1072,
   "speedup_vs_cpu_mt": 2.2300860195139536,
   "speedup_vs_cpu_1t": 2.1728878719355733,
   "cpu_mt_median_s": 0.0020858747884631157,
   "cpu_1t_median_s": 0.002032375428825617,
   "metal_wins_vs_mt": true,
   "cpu_wins_vs_mt": false
  },
  {
   "kernel": "realized_fused",
   "S": 10,
   "n": 390,
   "elements": 3900,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0002327919937670231,
   "ci_lo": 0.00019591627642512321,
   "ci_hi": 0.0002467092126607895,
   "device_s": 3.124994691461325e-05,
   "elements_per_s": 16753153.477877328,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3255.451648,
   "speedup_vs_cpu_mt": 0.6099872578848089,
   "speedup_vs_cpu_1t": 0.04850805233668789,
   "cpu_mt_median_s": 0.00014200014993548393,
   "cpu_1t_median_s": 1.1292286217212677e-05,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": true
  },
  {
   "kernel": "realized_fused",
   "S": 10,
   "n": 3900,
   "elements": 39000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.00016404176130890846,
   "ci_lo": 0.00015954114496707916,
   "ci_hi": 0.0009339582175016403,
   "device_s": 1.6249949112534523e-05,
   "elements_per_s": 237744338.32467064,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3222.339584,
   "speedup_vs_cpu_mt": 0.8229631795433707,
   "speedup_vs_cpu_1t": 1.145538879915521,
   "cpu_mt_median_s": 0.000135000329464674,
   "cpu_1t_median_s": 0.00018791621550917625,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": true
  },
  {
   "kernel": "realized_fused",
   "S": 10,
   "n": 39000,
   "elements": 390000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0007184999994933605,
   "ci_lo": 0.000268666073679924,
   "ci_hi": 0.0007987502031028271,
   "device_s": 5.52500132471323e-05,
   "elements_per_s": 542797495.1635388,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3355.901952,
   "speedup_vs_cpu_mt": 0.5059149029110798,
   "speedup_vs_cpu_1t": 3.5805503293330316,
   "cpu_mt_median_s": 0.00036349985748529434,
   "cpu_1t_median_s": 0.002572625409811735,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": false
  },
  {
   "kernel": "realized_fused",
   "S": 10,
   "n": 390000,
   "elements": 3900000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0010219174437224865,
   "ci_lo": 0.001003749668598175,
   "ci_hi": 0.0010474580340087414,
   "device_s": 8.649996016174555e-05,
   "elements_per_s": 3816355248.6134977,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3222.339584,
   "speedup_vs_cpu_mt": 2.4725987229278337,
   "speedup_vs_cpu_1t": 20.444125928265052,
   "cpu_mt_median_s": 0.0025267917662858963,
   "cpu_1t_median_s": 0.02089220890775323,
   "metal_wins_vs_mt": true,
   "cpu_wins_vs_mt": false
  },
  {
   "kernel": "realized_fused",
   "S": 100,
   "n": 390,
   "elements": 39000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.00022633420303463936,
   "ci_lo": 0.00017816666513681412,
   "ci_hi": 0.00027566729113459587,
   "device_s": 3.15000070258975e-05,
   "elements_per_s": 172311561.73966,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3368.943616,
   "speedup_vs_cpu_mt": 2.4392376077309077,
   "speedup_vs_cpu_1t": 0.8267643797230321,
   "cpu_mt_median_s": 0.0005520828999578953,
   "cpu_1t_median_s": 0.0001871250569820404,
   "metal_wins_vs_mt": true,
   "cpu_wins_vs_mt": false
  },
  {
   "kernel": "realized_fused",
   "S": 100,
   "n": 3900,
   "elements": 390000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0006413329392671585,
   "ci_lo": 0.000292791984975338,
   "ci_hi": 0.0008833748288452625,
   "device_s": 5.287490785121918e-05,
   "elements_per_s": 608108481.7593294,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 490.962944,
   "speedup_vs_cpu_mt": 0.9380200573315559,
   "speedup_vs_cpu_1t": 3.1862033963283407,
   "cpu_mt_median_s": 0.0006015831604599953,
   "cpu_1t_median_s": 0.002043417189270258,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": false
  },
  {
   "kernel": "realized_fused",
   "S": 100,
   "n": 39000,
   "elements": 3900000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0011692089028656483,
   "ci_lo": 0.0008370000869035721,
   "ci_hi": 0.0011960407719016075,
   "device_s": 0.000259208376519382,
   "elements_per_s": 3335588696.289753,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3360.096256,
   "speedup_vs_cpu_mt": 1.8734535658542084,
   "speedup_vs_cpu_1t": 17.61073211258148,
   "cpu_mt_median_s": 0.0021904585883021355,
   "cpu_1t_median_s": 0.020590624772012234,
   "metal_wins_vs_mt": true,
   "cpu_wins_vs_mt": false
  },
  {
   "kernel": "realized_fused",
   "S": 100,
   "n": 390000,
   "elements": 39000000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.008720249868929386,
   "ci_lo": 0.007397250272333622,
   "ci_hi": 0.008947291877120733,
   "device_s": 0.0021910000359639525,
   "elements_per_s": 4472348910.43187,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 486.178816,
   "speedup_vs_cpu_mt": 2.1550509398559075,
   "speedup_vs_cpu_1t": 26.67203779822428,
   "cpu_mt_median_s": 0.01879258267581463,
   "cpu_1t_median_s": 0.2325868341140449,
   "metal_wins_vs_mt": true,
   "cpu_wins_vs_mt": false
  },
  {
   "kernel": "realized_fused",
   "S": 1000,
   "n": 390,
   "elements": 390000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.00033845799043774605,
   "ci_lo": 0.00033012498170137405,
   "ci_hi": 0.00035137496888637543,
   "device_s": 0.00013987498823553324,
   "elements_per_s": 1152284806.4410944,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3431.87456,
   "speedup_vs_cpu_mt": 2.4815989367571367,
   "speedup_vs_cpu_1t": 6.605815916448483,
   "cpu_mt_median_s": 0.0008399169892072678,
   "cpu_1t_median_s": 0.002235791180282831,
   "metal_wins_vs_mt": true,
   "cpu_wins_vs_mt": false
  },
  {
   "kernel": "realized_fused",
   "S": 1000,
   "n": 3900,
   "elements": 3900000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0008750832639634609,
   "ci_lo": 0.0008294158615171909,
   "ci_hi": 0.0015468327328562737,
   "device_s": 0.0002547500189393759,
   "elements_per_s": 4456718761.065055,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3432.742912,
   "speedup_vs_cpu_mt": 1.9909537272506195,
   "speedup_vs_cpu_1t": 24.27716502583243,
   "cpu_mt_median_s": 0.0017422502860426903,
   "cpu_1t_median_s": 0.021244540810585022,
   "metal_wins_vs_mt": true,
   "cpu_wins_vs_mt": false
  },
  {
   "kernel": "realized_fused",
   "S": 1000,
   "n": 39000,
   "elements": 39000000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.005891874898225069,
   "ci_lo": 0.005617166869342327,
   "ci_hi": 0.00844891695305705,
   "device_s": 0.0009469999931752682,
   "elements_per_s": 6619285146.693929,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3369.04192,
   "speedup_vs_cpu_mt": 3.4364769430726474,
   "speedup_vs_cpu_1t": 40.52558334364075,
   "cpu_mt_median_s": 0.02024729223921895,
   "cpu_1t_median_s": 0.23877166723832488,
   "metal_wins_vs_mt": true,
   "cpu_wins_vs_mt": false
  },
  {
   "kernel": "realized_fused",
   "S": 10000,
   "n": 390,
   "elements": 3900000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0022132499143481255,
   "ci_lo": 0.0018930002115666866,
   "ci_hi": 0.0032122922129929066,
   "device_s": 0.001203041640110314,
   "elements_per_s": 1762114605.6380522,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3431.87456,
   "speedup_vs_cpu_mt": 0.9605408719443508,
   "speedup_vs_cpu_1t": 9.595994502745471,
   "cpu_mt_median_s": 0.002125917002558708,
   "cpu_1t_median_s": 0.021238334011286497,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": false
  },
  {
   "kernel": "realized_fused",
   "S": 10000,
   "n": 3900,
   "elements": 39000000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0078027499839663506,
   "ci_lo": 0.006132957991212606,
   "ci_hi": 0.008690916933119297,
   "device_s": 0.0019207500154152513,
   "elements_per_s": 4998237811.046104,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 577.1264,
   "speedup_vs_cpu_mt": 2.0113315140749037,
   "speedup_vs_cpu_1t": 29.947449069536802,
   "cpu_mt_median_s": 0.01569391693919897,
   "cpu_1t_median_s": 0.2336724577471614,
   "metal_wins_vs_mt": true,
   "cpu_wins_vs_mt": false
  },
  {
   "kernel": "regression_stats",
   "S": 1,
   "n": 390,
   "elements": 390,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.000894167460501194,
   "ci_lo": 0.00017508305609226227,
   "ci_hi": 0.000988542102277279,
   "device_s": 7.087492849677801e-05,
   "elements_per_s": 436159.91100973333,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3368.910848,
   "speedup_vs_cpu_mt": 0.003588669989219929,
   "speedup_vs_cpu_1t": 0.003541800115612355,
   "cpu_mt_median_s": 3.2088719308376312e-06,
   "cpu_1t_median_s": 3.1669624149799347e-06,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": true
  },
  {
   "kernel": "regression_stats",
   "S": 1,
   "n": 3900,
   "elements": 3900,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0003482499159872532,
   "ci_lo": 0.0003069578669965267,
   "ci_hi": 0.0003833337686955929,
   "device_s": 0.00019212497863918543,
   "elements_per_s": 11198854.101497471,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 2962.440192,
   "speedup_vs_cpu_mt": 0.0867420550075482,
   "speedup_vs_cpu_1t": 0.08602668142876818,
   "cpu_mt_median_s": 3.0207913368940353e-05,
   "cpu_1t_median_s": 2.9958784580230713e-05,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": true
  },
  {
   "kernel": "regression_stats",
   "S": 1,
   "n": 39000,
   "elements": 39000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0004085828550159931,
   "ci_lo": 0.0003508748486638069,
   "ci_hi": 0.00041229184716939926,
   "device_s": 0.00018866662867367268,
   "elements_per_s": 95451875.96888623,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3359.932416,
   "speedup_vs_cpu_mt": 0.9395287346496851,
   "speedup_vs_cpu_1t": 0.7399549819072855,
   "cpu_mt_median_s": 0.0003838753327727318,
   "cpu_1t_median_s": 0.00030233291909098625,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": false
  },
  {
   "kernel": "regression_stats",
   "S": 1,
   "n": 390000,
   "elements": 390000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0007901252247393131,
   "ci_lo": 0.0007615000940859318,
   "ci_hi": 0.0008232500404119492,
   "device_s": 9.650003630667925e-05,
   "elements_per_s": 493592645.5564979,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 2920.12032,
   "speedup_vs_cpu_mt": 3.851129874745179,
   "speedup_vs_cpu_1t": 3.7631692009752586,
   "cpu_mt_median_s": 0.0030428748577833176,
   "cpu_1t_median_s": 0.0029733749106526375,
   "metal_wins_vs_mt": true,
   "cpu_wins_vs_mt": false
  },
  {
   "kernel": "regression_stats",
   "S": 10,
   "n": 390,
   "elements": 3900,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0006523327901959419,
   "ci_lo": 0.0003105420619249344,
   "ci_hi": 0.0007279999554157257,
   "device_s": 9.250000584870577e-05,
   "elements_per_s": 5978543.557236411,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 2962.440192,
   "speedup_vs_cpu_mt": 0.06336187265949686,
   "speedup_vs_cpu_1t": 0.04790509353446491,
   "cpu_mt_median_s": 4.133302718400955e-05,
   "cpu_1t_median_s": 3.1250063329935074e-05,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": true
  },
  {
   "kernel": "regression_stats",
   "S": 10,
   "n": 3900,
   "elements": 39000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.00036174990236759186,
   "ci_lo": 0.0003430834040045738,
   "ci_hi": 0.0005066245794296265,
   "device_s": 0.00018737500067800283,
   "elements_per_s": 107809289.63560627,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3431.514112,
   "speedup_vs_cpu_mt": 0.2107814615911396,
   "speedup_vs_cpu_1t": 0.8286108550920896,
   "cpu_mt_median_s": 7.625017315149307e-05,
   "cpu_1t_median_s": 0.0002997498959302902,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": true
  },
  {
   "kernel": "regression_stats",
   "S": 10,
   "n": 39000,
   "elements": 390000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0006961673498153687,
   "ci_lo": 0.0006541251204907894,
   "ci_hi": 0.0008656666614115238,
   "device_s": 0.0002501250710338354,
   "elements_per_s": 560210127.7852694,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 576.421888,
   "speedup_vs_cpu_mt": 0.8078766133692931,
   "speedup_vs_cpu_1t": 4.429491347203493,
   "cpu_mt_median_s": 0.0005624173209071159,
   "cpu_1t_median_s": 0.003083667252212763,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": true
  },
  {
   "kernel": "regression_stats",
   "S": 10,
   "n": 390000,
   "elements": 3900000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.004117042291909456,
   "ci_lo": 0.002989500295370817,
   "ci_hi": 0.004763709381222725,
   "device_s": 0.000529208336956799,
   "elements_per_s": 947281986.3094499,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 2962.440192,
   "speedup_vs_cpu_mt": 1.6164011753500427,
   "speedup_vs_cpu_1t": 7.9335070336527025,
   "cpu_mt_median_s": 0.006654791999608278,
   "cpu_1t_median_s": 0.032662583980709314,
   "metal_wins_vs_mt": true,
   "cpu_wins_vs_mt": false
  },
  {
   "kernel": "regression_stats",
   "S": 100,
   "n": 390,
   "elements": 39000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0003363746218383312,
   "ci_lo": 0.00026350002735853195,
   "ci_hi": 0.0005579171702265739,
   "device_s": 0.00019220833200961351,
   "elements_per_s": 115942159.33074829,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3255.451648,
   "speedup_vs_cpu_mt": 0.5239693836444206,
   "speedup_vs_cpu_1t": 1.0743231551070866,
   "cpu_mt_median_s": 0.00017625000327825546,
   "cpu_1t_median_s": 0.00036137504503130913,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": true
  },
  {
   "kernel": "regression_stats",
   "S": 100,
   "n": 3900,
   "elements": 390000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0005946666933596134,
   "ci_lo": 0.0005814996547996998,
   "ci_hi": 0.0005976245738565922,
   "device_s": 0.00023925001733005047,
   "elements_per_s": 655829566.9741753,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 2962.440192,
   "speedup_vs_cpu_mt": 0.739769482011876,
   "speedup_vs_cpu_1t": 5.041970592864576,
   "cpu_mt_median_s": 0.0004399162717163563,
   "cpu_1t_median_s": 0.0029982919804751873,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": true
  },
  {
   "kernel": "regression_stats",
   "S": 100,
   "n": 39000,
   "elements": 3900000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0044694580137729645,
   "ci_lo": 0.0031338329426944256,
   "ci_hi": 0.004542499780654907,
   "device_s": 0.0016885417280718684,
   "elements_per_s": 872589022.6470106,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3431.46496,
   "speedup_vs_cpu_mt": 0.7100881967325159,
   "speedup_vs_cpu_1t": 7.202327484390641,
   "cpu_mt_median_s": 0.0031737093813717365,
   "cpu_1t_median_s": 0.03219050029292703,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": false
  },
  {
   "kernel": "regression_stats",
   "S": 100,
   "n": 390000,
   "elements": 39000000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.03070899983868003,
   "ci_lo": 0.027952750213444233,
   "ci_hi": 0.031758958008140326,
   "device_s": 0.005229625036008656,
   "elements_per_s": 1269986004.2617507,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 2913.845248,
   "speedup_vs_cpu_mt": 1.0993601156685116,
   "speedup_vs_cpu_1t": 9.813711025090953,
   "cpu_mt_median_s": 0.033760249614715576,
   "cpu_1t_median_s": 0.3013692502863705,
   "metal_wins_vs_mt": true,
   "cpu_wins_vs_mt": false
  },
  {
   "kernel": "regression_stats",
   "S": 1000,
   "n": 390,
   "elements": 390000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0005814996547996998,
   "ci_lo": 0.0005787923000752926,
   "ci_hi": 0.0009665000252425671,
   "device_s": 0.00020362494979053736,
   "elements_per_s": 670679675.870723,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3431.858176,
   "speedup_vs_cpu_mt": 1.1988402904959397,
   "speedup_vs_cpu_1t": 5.367372139264439,
   "cpu_mt_median_s": 0.0006971252150833607,
   "cpu_1t_median_s": 0.0031211250461637974,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": false
  },
  {
   "kernel": "regression_stats",
   "S": 1000,
   "n": 3900,
   "elements": 3900000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.003796167206019163,
   "ci_lo": 0.0033167502842843533,
   "ci_hi": 0.004350459203124046,
   "device_s": 0.000579125015065074,
   "elements_per_s": 1027352007.5238521,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 781.1072,
   "speedup_vs_cpu_mt": 0.9013804482638873,
   "speedup_vs_cpu_1t": 9.060761582722323,
   "cpu_mt_median_s": 0.003421790897846222,
   "cpu_1t_median_s": 0.03439616598188877,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": false
  },
  {
   "kernel": "regression_stats",
   "S": 1000,
   "n": 39000,
   "elements": 39000000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.026743208058178425,
   "ci_lo": 0.025766875129193068,
   "ci_hi": 0.02711204206570983,
   "device_s": 0.004475083318538964,
   "elements_per_s": 1458314197.5771036,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3360.11264,
   "speedup_vs_cpu_mt": 1.1681813469411308,
   "speedup_vs_cpu_1t": 12.11658580559542,
   "cpu_mt_median_s": 0.031240916810929775,
   "cpu_1t_median_s": 0.3240363751538098,
   "metal_wins_vs_mt": true,
   "cpu_wins_vs_mt": false
  },
  {
   "kernel": "regression_stats",
   "S": 10000,
   "n": 390,
   "elements": 3900000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.004190749954432249,
   "ci_lo": 0.00392800010740757,
   "ci_hi": 0.005979666020721197,
   "device_s": 0.0011161250295117497,
   "elements_per_s": 930621020.6779948,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3432.742912,
   "speedup_vs_cpu_mt": 0.944759231194123,
   "speedup_vs_cpu_1t": 7.312414252819052,
   "cpu_mt_median_s": 0.003959249705076218,
   "cpu_1t_median_s": 0.030644499696791172,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": false
  },
  {
   "kernel": "regression_stats",
   "S": 10000,
   "n": 3900,
   "elements": 39000000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.026574499905109406,
   "ci_lo": 0.026411292143166065,
   "ci_hi": 0.026990290731191635,
   "device_s": 0.004499250091612339,
   "elements_per_s": 1467572301.9909616,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3221.83168,
   "speedup_vs_cpu_mt": 1.4107556096287706,
   "speedup_vs_cpu_1t": 13.430701182218417,
   "cpu_mt_median_s": 0.03749012481421232,
   "cpu_1t_median_s": 0.3569141672924161,
   "metal_wins_vs_mt": true,
   "cpu_wins_vs_mt": false
  },
  {
   "kernel": "rolling_sumsq",
   "S": 1,
   "n": 390,
   "elements": 390,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0002839169465005398,
   "ci_lo": 0.0001594587229192257,
   "ci_hi": 0.0008917916566133499,
   "device_s": 1.8750084564089775e-05,
   "elements_per_s": 1373641.146845944,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3253.567488,
   "speedup_vs_cpu_mt": 0.0027898646399008046,
   "speedup_vs_cpu_1t": 0.003373751654483219,
   "cpu_mt_median_s": 7.920898497104645e-07,
   "cpu_1t_median_s": 9.578652679920197e-07,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": true
  },
  {
   "kernel": "rolling_sumsq",
   "S": 1,
   "n": 3900,
   "elements": 3900,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0002416251227259636,
   "ci_lo": 0.0001499168574810028,
   "ci_hi": 0.0002560419961810112,
   "device_s": 4.958361387252808e-06,
   "elements_per_s": 16140705.718018986,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 2962.440192,
   "speedup_vs_cpu_mt": 0.031039573239594054,
   "speedup_vs_cpu_1t": 0.031039573239594054,
   "cpu_mt_median_s": 7.4999406933784485e-06,
   "cpu_1t_median_s": 7.4999406933784485e-06,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": true
  },
  {
   "kernel": "rolling_sumsq",
   "S": 1,
   "n": 39000,
   "elements": 39000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.00016458332538604736,
   "ci_lo": 0.00015858327969908714,
   "ci_hi": 0.00017341692000627518,
   "device_s": 1.9583385437726974e-05,
   "elements_per_s": 236962036.75871435,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 2962.440192,
   "speedup_vs_cpu_mt": 0.42683623811679494,
   "speedup_vs_cpu_1t": 0.4212624490719783,
   "cpu_mt_median_s": 7.025012746453285e-05,
   "cpu_1t_median_s": 6.93327747285366e-05,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": true
  },
  {
   "kernel": "rolling_sumsq",
   "S": 1,
   "n": 390000,
   "elements": 390000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0007903752848505974,
   "ci_lo": 0.000574291218072176,
   "ci_hi": 0.0009351670742034912,
   "device_s": 6.225006654858589e-05,
   "elements_per_s": 493436481.97921664,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 2962.440192,
   "speedup_vs_cpu_mt": 0.909589128259996,
   "speedup_vs_cpu_1t": 0.8747948233625048,
   "cpu_mt_median_s": 0.000718916766345501,
   "cpu_1t_median_s": 0.0006914162077009678,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": false
  },
  {
   "kernel": "rolling_sumsq",
   "S": 10,
   "n": 390,
   "elements": 3900,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0002032076008617878,
   "ci_lo": 0.00014958297833800316,
   "ci_hi": 0.0006746253930032253,
   "device_s": 1.4499993994832039e-05,
   "elements_per_s": 19192195.486095995,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 17.170432,
   "speedup_vs_cpu_mt": 0.08858003826895974,
   "speedup_vs_cpu_1t": 0.03526473183083745,
   "cpu_mt_median_s": 1.800013706088066e-05,
   "cpu_1t_median_s": 7.1660615503787994e-06,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": true
  },
  {
   "kernel": "rolling_sumsq",
   "S": 10,
   "n": 3900,
   "elements": 39000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.00016070762649178505,
   "ci_lo": 0.00014404207468032837,
   "ci_hi": 0.00017450004816055298,
   "device_s": 1.9625062122941017e-05,
   "elements_per_s": 242676722.01601195,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 2962.440192,
   "speedup_vs_cpu_mt": 0.2792328398774908,
   "speedup_vs_cpu_1t": 0.4311610265504162,
   "cpu_mt_median_s": 4.487484693527222e-05,
   "cpu_1t_median_s": 6.929086521267891e-05,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": true
  },
  {
   "kernel": "rolling_sumsq",
   "S": 10,
   "n": 39000,
   "elements": 390000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0007357923313975334,
   "ci_lo": 0.0007097921334207058,
   "ci_hi": 0.0007738331332802773,
   "device_s": 9.062490426003933e-05,
   "elements_per_s": 530040859.8432253,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3431.46496,
   "speedup_vs_cpu_mt": 0.18036050837224432,
   "speedup_vs_cpu_1t": 0.9643225563919291,
   "cpu_mt_median_s": 0.000132707878947258,
   "cpu_1t_median_s": 0.0007095411419868469,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": true
  },
  {
   "kernel": "rolling_sumsq",
   "S": 10,
   "n": 390000,
   "elements": 3900000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.002544207964092493,
   "ci_lo": 0.0024965000338852406,
   "ci_hi": 0.0026084999553859234,
   "device_s": 0.00076149997767061,
   "elements_per_s": 1532893558.6407974,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3253.567488,
   "speedup_vs_cpu_mt": 0.6940765734230536,
   "speedup_vs_cpu_1t": 2.7174959207635196,
   "cpu_mt_median_s": 0.0017658751457929611,
   "cpu_1t_median_s": 0.006913874763995409,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": true
  },
  {
   "kernel": "rolling_sumsq",
   "S": 100,
   "n": 390,
   "elements": 39000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0002676667645573616,
   "ci_lo": 0.00017749983817338943,
   "ci_hi": 0.0007679169066250324,
   "device_s": 9.541632607579231e-06,
   "elements_per_s": 145703558.1705259,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 2962.440192,
   "speedup_vs_cpu_mt": 0.6718550477549103,
   "speedup_vs_cpu_1t": 0.2612097910613942,
   "cpu_mt_median_s": 0.00017983326688408852,
   "cpu_1t_median_s": 6.991717964410782e-05,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": false
  },
  {
   "kernel": "rolling_sumsq",
   "S": 100,
   "n": 3900,
   "elements": 390000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0006820829585194588,
   "ci_lo": 0.0003593750298023224,
   "ci_hi": 0.0008333330042660236,
   "device_s": 9.058334399014711e-05,
   "elements_per_s": 571777956.2277012,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3431.530496,
   "speedup_vs_cpu_mt": 0.2978012810272249,
   "speedup_vs_cpu_1t": 1.1826521987872431,
   "cpu_mt_median_s": 0.00020312517881393433,
   "cpu_1t_median_s": 0.000806666910648346,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": true
  },
  {
   "kernel": "rolling_sumsq",
   "S": 100,
   "n": 39000,
   "elements": 3900000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.002226833254098892,
   "ci_lo": 0.0021508331410586834,
   "ci_hi": 0.002632542047649622,
   "device_s": 0.00043350004125386477,
   "elements_per_s": 1751365978.0413911,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3431.87456,
   "speedup_vs_cpu_mt": 0.4883431672524638,
   "speedup_vs_cpu_1t": 3.2953184884928928,
   "cpu_mt_median_s": 0.0010874588042497635,
   "cpu_1t_median_s": 0.007338124793022871,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": true
  },
  {
   "kernel": "rolling_sumsq",
   "S": 100,
   "n": 390000,
   "elements": 39000000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.014852333813905716,
   "ci_lo": 0.014340042136609554,
   "ci_hi": 0.01528970804065466,
   "device_s": 0.0023322500055655837,
   "elements_per_s": 2625849949.823083,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3255.451648,
   "speedup_vs_cpu_mt": 0.6337920907333103,
   "speedup_vs_cpu_1t": 4.687593885765181,
   "cpu_mt_median_s": 0.009413291700184345,
   "cpu_1t_median_s": 0.06962170917540789,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": true
  },
  {
   "kernel": "rolling_sumsq",
   "S": 1000,
   "n": 390,
   "elements": 390000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0009075421839952469,
   "ci_lo": 0.000882917083799839,
   "ci_hi": 0.000995416659861803,
   "device_s": 2.3708329536020756e-05,
   "elements_per_s": 429732090.5603684,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 2962.440192,
   "speedup_vs_cpu_mt": 0.29870103215504695,
   "speedup_vs_cpu_1t": 0.7921117822479183,
   "cpu_mt_median_s": 0.0002710837870836258,
   "cpu_1t_median_s": 0.0007188748568296432,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": true
  },
  {
   "kernel": "rolling_sumsq",
   "S": 1000,
   "n": 3900,
   "elements": 3900000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.002144791651517153,
   "ci_lo": 0.0021287081763148308,
   "ci_hi": 0.0034982911311089993,
   "device_s": 0.0007578750373795629,
   "elements_per_s": 1818358439.264379,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3431.497728,
   "speedup_vs_cpu_mt": 0.47898947112456725,
   "speedup_vs_cpu_1t": 3.242467441252045,
   "cpu_mt_median_s": 0.0010273326188325882,
   "cpu_1t_median_s": 0.00695441709831357,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": true
  },
  {
   "kernel": "rolling_sumsq",
   "S": 1000,
   "n": 39000,
   "elements": 39000000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.014728832989931107,
   "ci_lo": 0.014530166983604431,
   "ci_hi": 0.01593674998730421,
   "device_s": 0.002570083364844322,
   "elements_per_s": 2647867623.094178,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3368.894464,
   "speedup_vs_cpu_mt": 0.5643550943271196,
   "speedup_vs_cpu_1t": 4.770453192305718,
   "cpu_mt_median_s": 0.00831229193136096,
   "cpu_1t_median_s": 0.07026320835575461,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": true
  },
  {
   "kernel": "rolling_sumsq",
   "S": 10000,
   "n": 390,
   "elements": 3900000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.0021627498790621758,
   "ci_lo": 0.002053916919976473,
   "ci_hi": 0.0022570830769836903,
   "device_s": 0.0002617916325107217,
   "elements_per_s": 1803259839.5941842,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 3431.858176,
   "speedup_vs_cpu_mt": 0.49574203299838304,
   "speedup_vs_cpu_1t": 3.2587757053011432,
   "cpu_mt_median_s": 0.00107216602191329,
   "cpu_1t_median_s": 0.007047916762530804,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": true
  },
  {
   "kernel": "rolling_sumsq",
   "S": 10000,
   "n": 3900,
   "elements": 39000000,
   "backend": "metal",
   "threads": 0,
   "median_s": 0.014753291849046946,
   "ci_lo": 0.01457533286884427,
   "ci_hi": 0.01564762508496642,
   "device_s": 0.0025377500569447875,
   "elements_per_s": 2643477835.254739,
   "reps": 5,
   "precision": "fp32",
   "peak_rss_mb": 2962.440192,
   "speedup_vs_cpu_mt": 0.4434970991671636,
   "speedup_vs_cpu_1t": 4.6922437792300995,
   "cpu_mt_median_s": 0.00654304213821888,
   "cpu_1t_median_s": 0.06922604190185666,
   "metal_wins_vs_mt": false,
   "cpu_wins_vs_mt": true
  }
 ],
 "breakeven": {
  "garch_nll_P1": {
   "smallest_metal_win_elements": null,
   "largest_cpu_win_elements": 39000000,
   "max_speedup_vs_mt": 0.33211607376881125,
   "n_cases": 10
  },
  "garch_nll_P64": {
   "smallest_metal_win_elements": 39000,
   "largest_cpu_win_elements": 39000,
   "max_speedup_vs_mt": 7.934195016316967,
   "n_cases": 8
  },
  "realized_fused": {
   "smallest_metal_win_elements": 39000,
   "largest_cpu_win_elements": 39000,
   "max_speedup_vs_mt": 3.4364769430726474,
   "n_cases": 17
  },
  "regression_stats": {
   "smallest_metal_win_elements": 390000,
   "largest_cpu_win_elements": 390000,
   "max_speedup_vs_mt": 3.851129874745179,
   "n_cases": 17
  },
  "rolling_sumsq": {
   "smallest_metal_win_elements": null,
   "largest_cpu_win_elements": 39000000,
   "max_speedup_vs_mt": 0.909589128259996,
   "n_cases": 17
  }
 },
 "shader_library": [
  {
   "backend": "metal",
   "compiled_now": false,
   "kernel": "shader_library_load",
   "note": "metallib:/Users/simon-pierreboucher/Desktop/UQO/Recherche/Recherche-UQO/finance/wp40_raid/build/macos-release/raid.metallib chunks=1 convert_s=0.000001",
   "seconds": 0.000230708,
   "t": 0.03212570771574974,
   "wall_first_call": 0.013961625
  }
 ]
}