1{2 "B": 10000,3 "seed": 0,4 "comparisons": [5 {6 "comparison": "laptop: TITAN all-resident (64 slots, preloaded) vs mlx-lm",7 "n_a": 12,8 "n_b": 12,9 "median_a": 489.12865342395037,10 "ci_a": [11 488.2082407000963,12 489.536882845776513 ],14 "median_b": 495.04061243824856,15 "ci_b": [16 494.1194679570178,17 495.890470899934618 ],19 "diff": -5.911959014298191,20 "ci_diff": [21 -7.065204357666232,22 -4.83455006205679223 ],24 "rel_diff_pct": -1.1942371728209773,25 "unit": "tok/s",26 "source": "gateB/capacity_sweep_laptop/sweep.jsonl"27 },28 {29 "comparison": "laptop: TITAN 64 slots not preloaded (sync path) vs mlx-lm",30 "n_a": 12,31 "n_b": 12,32 "median_a": 201.4454551722675,33 "ci_a": [34 200.95837279071753,35 202.6623232419196436 ],37 "median_b": 495.04061243824856,38 "ci_b": [39 494.1194679570178,40 495.890470899934641 ],42 "diff": -293.5951572659811,43 "ci_diff": [44 -294.4450157276671,45 -291.9789707435338646 ],47 "rel_diff_pct": -59.307287097097344,48 "unit": "tok/s",49 "source": "gateB/capacity_sweep_laptop/sweep.jsonl"50 },51 {52 "comparison": "M3U96a Qwen3-30B: TITAN all-resident (128 slots) vs mlx-lm",53 "n_a": 18,54 "n_b": 18,55 "median_a": 121.58726022750393,56 "ci_a": [57 121.4222850284662,58 121.753911030540759 ],60 "median_b": 121.65810731258236,61 "ci_b": [62 121.37584628881791,63 121.811266991811864 ],65 "diff": -0.07084708507842663,66 "ci_diff": [67 -0.27269819921461647,68 0.283233034349606869 ],70 "rel_diff_pct": -0.058234577738740925,71 "unit": "tok/s",72 "source": "gateC/qwen3_m3u96a/baseline.jsonl"73 },74 {75 "comparison": "laptop cold NVMe, 16 slots: transition m=8 vs no prefetch",76 "n_a": 12,77 "n_b": 12,78 "median_a": 27.904324206563125,79 "ci_a": [80 26.510834259463596,81 30.5133137918526482 ],83 "median_b": 36.36478535014914,84 "ci_b": [85 34.13753044405526,86 42.0450070628706187 ],88 "diff": -8.460461143586016,89 "ci_diff": [90 -14.521560050815605,91 -5.083125103663352592 ],93 "rel_diff_pct": -23.265533020811077,94 "unit": "tok/s",95 "source": "gateC/cold_nvme_laptop/sweep.jsonl"96 },97 {98 "comparison": "laptop cold NVMe, 16 slots: transition2 m=8 vs no prefetch",99 "n_a": 12,100 "n_b": 12,101 "median_a": 34.96599880167474,102 "ci_a": [103 33.67628976796817,104 41.49003741051476105 ],106 "median_b": 36.36478535014914,107 "ci_b": [108 34.13753044405526,109 42.04500706287061110 ],111 "diff": -1.3987865484743978,112 "ci_diff": [113 -7.324854162049796,114 5.998514358174624115 ],116 "rel_diff_pct": -3.8465414686372155,117 "unit": "tok/s",118 "source": "gateC/cold_nvme_laptop/sweep.jsonl"119 },120 {121 "comparison": "laptop cold NVMe, 16 slots: transition3 m=8 vs no prefetch",122 "n_a": 12,123 "n_b": 12,124 "median_a": 29.083938320259847,125 "ci_a": [126 27.671125828589567,127 34.70508437115666128 ],129 "median_b": 36.36478535014914,130 "ci_b": [131 34.13753044405526,132 42.04500706287061133 ],134 "diff": -7.280847029889294,135 "ci_diff": [136 -13.09924827054045,137 -1.0543321191199624138 ],139 "rel_diff_pct": -20.02169670405997,140 "unit": "tok/s",141 "source": "gateC/cold_nvme_laptop/sweep.jsonl"142 },143 {144 "comparison": "laptop cold NVMe, 16 slots: transition2 m=16 vs no prefetch",145 "n_a": 12,146 "n_b": 12,147 "median_a": 25.940903169028186,148 "ci_a": [149 25.001461847324514,150 29.583536873911555151 ],152 "median_b": 36.36478535014914,153 "ci_b": [154 34.13753044405526,155 42.04500706287061156 ],157 "diff": -10.423882181120955,158 "ci_diff": [159 -16.145654899217902,160 -5.865959852413532161 ],162 "rel_diff_pct": -28.66477027363563,163 "unit": "tok/s",164 "source": "gateC/cold_nvme_laptop/sweep.jsonl"165 },166 {167 "comparison": "laptop cold NVMe, 16 slots: prev_rank m=16 vs no prefetch",168 "n_a": 12,169 "n_b": 12,170 "median_a": 19.13073106755353,171 "ci_a": [172 18.745326302296334,173 20.005360570414403174 ],175 "median_b": 36.36478535014914,176 "ci_b": [177 34.13753044405526,178 42.04500706287061179 ],180 "diff": -17.23405428259561,181 "ci_diff": [182 -22.91427599531708,183 -14.833423635496317184 ],185 "rel_diff_pct": -47.392151821198446,186 "unit": "tok/s",187 "source": "gateC/cold_nvme_laptop/sweep.jsonl"188 },189 {190 "comparison": "laptop cold NVMe, 32 slots: transition m=8 vs no prefetch",191 "n_a": 12,192 "n_b": 12,193 "median_a": 41.32499296378025,194 "ci_a": [195 35.45132953777109,196 46.87438633199691197 ],198 "median_b": 50.55650193124784,199 "ci_b": [200 46.6179394611299,201 56.852686944161775202 ],203 "diff": -9.23150896746759,204 "ci_diff": [205 -17.486052289083403,206 -2.598222895951629207 ],208 "rel_diff_pct": -18.259785813546966,209 "unit": "tok/s",210 "source": "gateC/cold_nvme_laptop/sweep.jsonl"211 },212 {213 "comparison": "laptop cold NVMe, 32 slots: transition2 m=8 vs no prefetch",214 "n_a": 12,215 "n_b": 12,216 "median_a": 42.02425447124503,217 "ci_a": [218 38.459575220686936,219 46.75470157960534220 ],221 "median_b": 50.55650193124784,222 "ci_b": [223 46.6179394611299,224 56.852686944161775225 ],226 "diff": -8.532247460002807,227 "ci_diff": [228 -15.240124741890014,229 -2.7179076483432230 ],231 "rel_diff_pct": -16.876657074901807,232 "unit": "tok/s",233 "source": "gateC/cold_nvme_laptop/sweep.jsonl"234 },235 {236 "comparison": "laptop cold NVMe, 32 slots: transition3 m=8 vs no prefetch",237 "n_a": 12,238 "n_b": 12,239 "median_a": 41.02227504595441,240 "ci_a": [241 37.59374509091457,242 46.645720976690825243 ],244 "median_b": 50.55650193124784,245 "ci_b": [246 46.6179394611299,247 56.852686944161775248 ],249 "diff": -9.534226885293428,250 "ci_diff": [251 -16.24725536552569,252 -2.9552358624899995253 ],254 "rel_diff_pct": -18.858557299433205,255 "unit": "tok/s",256 "source": "gateC/cold_nvme_laptop/sweep.jsonl"257 },258 {259 "comparison": "laptop cold NVMe, 32 slots: transition2 m=16 vs no prefetch",260 "n_a": 12,261 "n_b": 12,262 "median_a": 40.9858914503088,263 "ci_a": [264 35.4116167818508,265 46.03429262237789266 ],267 "median_b": 50.55650193124784,268 "ci_b": [269 46.6179394611299,270 56.852686944161775271 ],272 "diff": -9.570610480939038,273 "ci_diff": [274 -17.710608210914906,275 -3.556491759775213276 ],277 "rel_diff_pct": -18.930523504086935,278 "unit": "tok/s",279 "source": "gateC/cold_nvme_laptop/sweep.jsonl"280 },281 {282 "comparison": "laptop cold NVMe, 32 slots: prev_rank m=16 vs no prefetch",283 "n_a": 12,284 "n_b": 12,285 "median_a": 28.37905550744074,286 "ci_a": [287 26.997499853533533,288 32.45499196158143289 ],290 "median_b": 50.55650193124784,291 "ci_b": [292 46.6179394611299,293 56.852686944161775294 ],295 "diff": -22.177446423807098,296 "ci_diff": [297 -28.57031143448528,298 -16.946638006695025299 ],300 "rel_diff_pct": -43.86665528000013,301 "unit": "tok/s",302 "source": "gateC/cold_nvme_laptop/sweep.jsonl"303 },304 {305 "comparison": "laptop cold NVMe, 16 slots (general): probe m=8 vs no prefetch",306 "n_a": 4,307 "n_b": 4,308 "median_a": 31.841276936782968,309 "ci_a": [310 31.596870661689568,311 32.84528995900421312 ],313 "median_b": 32.99759250772945,314 "ci_b": [315 32.636451601945915,316 33.273626374779317 ],318 "diff": -1.1563155709464823,319 "ci_diff": [320 -1.5500378457589257,321 -0.1523025487252383322 ],323 "rel_diff_pct": -3.5042422282038412,324 "unit": "tok/s",325 "source": "gateC/cold_nvme_laptop/attribution.jsonl"326 },327 {328 "comparison": "laptop cold NVMe, 16 slots (general): probe m=16 vs no prefetch",329 "n_a": 4,330 "n_b": 4,331 "median_a": 25.317370635052725,332 "ci_a": [333 24.76517223527836,334 26.203663395431743335 ],336 "median_b": 32.99759250772945,337 "ci_b": [338 32.636451601945915,339 33.273626374779340 ],341 "diff": -7.680221872676725,342 "ci_diff": [343 -8.398426644951499,344 -6.752416400732631345 ],346 "rel_diff_pct": -23.275097632888485,347 "unit": "tok/s",348 "source": "gateC/cold_nvme_laptop/attribution.jsonl"349 },350 {351 "comparison": "laptop cold NVMe, 16 slots (general): probe2 m=8 vs no prefetch",352 "n_a": 4,353 "n_b": 4,354 "median_a": 31.34039585473687,355 "ci_a": [356 30.97991967772089,357 32.02471239861806358 ],359 "median_b": 32.99759250772945,360 "ci_b": [361 32.636451601945915,362 33.273626374779363 ],364 "diff": -1.6571966529925817,365 "ci_diff": [366 -2.1243054121081393,367 -0.8894574043585379368 ],369 "rel_diff_pct": -5.022174428647772,370 "unit": "tok/s",371 "source": "gateC/cold_nvme_laptop/attribution.jsonl"372 },373 {374 "comparison": "laptop cold NVMe, 16 slots (general): probe3 m=8 vs no prefetch",375 "n_a": 4,376 "n_b": 4,377 "median_a": 30.747024815645197,378 "ci_a": [379 29.277490438874107,380 32.56493375331503381 ],382 "median_b": 32.99759250772945,383 "ci_b": [384 32.636451601945915,385 33.273626374779386 ],387 "diff": -2.2505676920842532,388 "ci_diff": [389 -3.7201020688553434,390 -0.4326587544144189391 ],392 "rel_diff_pct": -6.820399674785588,393 "unit": "tok/s",394 "source": "gateC/cold_nvme_laptop/attribution.jsonl"395 },396 {397 "comparison": "laptop cold NVMe, 16 slots (general): transition2 m=8 vs no prefetch",398 "n_a": 4,399 "n_b": 4,400 "median_a": 27.12807201818323,401 "ci_a": [402 26.024585035633084,403 28.245109293687104404 ],405 "median_b": 32.99759250772945,406 "ci_b": [407 32.636451601945915,408 33.273626374779409 ],410 "diff": -5.869520489546222,411 "ci_diff": [412 -7.147820761442922,413 -4.541891551805659414 ],415 "rel_diff_pct": -17.787723417007857,416 "unit": "tok/s",417 "source": "gateC/cold_nvme_laptop/attribution.jsonl"418 },419 {420 "comparison": "laptop cold NVMe, 16 slots (general): transition2t50 m=8 vs no prefetch",421 "n_a": 4,422 "n_b": 4,423 "median_a": 27.924338730957444,424 "ci_a": [425 26.832091171562322,426 29.41238423146309427 ],428 "median_b": 32.99759250772945,429 "ci_b": [430 32.636451601945915,431 33.273626374779432 ],433 "diff": -5.073253776772006,434 "ci_diff": [435 -6.165501336167129,436 -3.5852082762663606437 ],438 "rel_diff_pct": -15.37461793791057,439 "unit": "tok/s",440 "source": "gateC/cold_nvme_laptop/attribution.jsonl"441 },442 {443 "comparison": "laptop cold NVMe, 16 slots (general): transition2t70 m=8 vs no prefetch",444 "n_a": 4,445 "n_b": 4,446 "median_a": 27.822276415206307,447 "ci_a": [448 26.652845588205146,449 29.388223861291042450 ],451 "median_b": 32.99759250772945,452 "ci_b": [453 32.636451601945915,454 33.273626374779455 ],456 "diff": -5.175316092523143,457 "ci_diff": [458 -6.3447469195243045,459 -3.6093686464384085460 ],461 "rel_diff_pct": -15.683920247547945,462 "unit": "tok/s",463 "source": "gateC/cold_nvme_laptop/attribution.jsonl"464 },465 {466 "comparison": "laptop H4: fused up+gate vs two kernels, 64 slots resident",467 "n_a": 12,468 "n_b": 12,469 "median_a": 475.590753313536,470 "ci_a": [471 473.54619144727985,472 476.23632189642444473 ],474 "median_b": 476.71854951152625,475 "ci_b": [476 475.7441354752446,477 478.1937096893877478 ],479 "diff": -1.1277961979902216,480 "ci_diff": [481 -3.608229933155485,482 0.07714985221900861483 ],484 "rel_diff_pct": -0.23657485095678102,485 "unit": "tok/s",486 "source": "gateD/fused_upgate_laptop.jsonl"487 },488 {489 "comparison": "laptop H4: fused up+gate vs two kernels, 16 slots cold",490 "n_a": 4,491 "n_b": 4,492 "median_a": 35.94742494598419,493 "ci_a": [494 35.731574910456565,495 36.7122725819878496 ],497 "median_b": 34.98584079217508,498 "ci_b": [499 34.12741716795578,500 35.49684771325741501 ],502 "diff": 0.9615841538091061,503 "ci_diff": [504 0.42788351347727627,505 1.964214448416044506 ],507 "rel_diff_pct": 2.7484951970174563,508 "unit": "tok/s",509 "source": "gateD/fused_upgate_laptop.jsonl"510 },511 {512 "comparison": "M3U96a+M3U96b OLMoE 32 slots: odd experts remote vs local",513 "n_a": 16,514 "n_b": 16,515 "median_a": 15.099750302534616,516 "ci_a": [517 14.832823372534836,518 15.478581724182211519 ],520 "median_b": 81.29106238077739,521 "ci_b": [522 70.45778798608657,523 86.91246514746022524 ],525 "diff": -66.19131207824277,526 "ci_diff": [527 -71.76838200410603,528 -55.195792166972154529 ],530 "rel_diff_pct": -81.42507938719571,531 "unit": "tok/s",532 "source": "gateE/two_node_m3u96a_m3u96b.jsonl"533 }534 ]535}