Measured on-device inference. Every number reproducible; every failure recorded.
Every measurement in the corpus, successes and failures together. Sort by any column; a missing value always sorts last, because “not measured” is not a small number.
Generative rows carry TTFT and tok/s instead of a single latency, and their p50 ms cell is deliberately empty: prefill and decode are different workloads, and one averaged number would describe neither. See methodology.
| Model | Recipe | Target | Outcome | p50 ms | TTFT ms | tok/s | cv | size MiB | peak RSS | first run ms | cosine | fallback (run) |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| distilbert-base-uncased | unknown | Apple M2 unknown | success | 11.37 | — | — | 0.2% | 64 | 187 | — | 0.9721 | 0.0% |
| distilbert-base-uncased | unknown | Apple M2 unknown | success | 11.47 | — | — | 0.2% | 64 | 188 | — | 0.9932 | 0.0% |
| distilbert-base-uncased | unknown | Apple M2 unknown | success | 32.12 | — | — | 0.1% | 253 | 488 | — | 1.0000 | 0.0% |
| distilbert-base-uncased | unknown | Apple M2 unknown | success | 32.12 | — | — | 0.1% | 253 | 488 | — | 1.0000 | 0.0% |
| distilbert-base-uncased | unknown | Apple M2 unknown | success | 33.38 | — | — | 0.2% | 253 | 488 | — | 1.0000 | 0.0% |
| distilbert-base-uncased | unknown | Apple M2 unknown | success | 35.58 | — | — | 0.2% | 127 | 494 | — | 1.0000 | 100.0% |
| distilbert-base-uncased | unknown | Apple M2 unknown | success | 35.65 | — | — | 0.2% | 127 | 518 | — | 1.0000 | 0.0% |
| distilbert-base-uncased | unknown | Apple M2 unknown | success | 36.66 | — | — | 0.2% | 253 | 1174 | — | 1.0000 | 39.5% |
| distilbert-base-uncased | unknown | Apple M2 unknown | success | 39.41 | — | — | 1.5% | 253 | 1177 | — | 1.0000 | 33.7% |
| distilbert-base-uncased | unknown | Apple M2 unknown | success | 51.66 | — | — | 0.2% | 253 | 1308 | — | 1.0000 | 22.0% |
| distilbert-base-uncased | unknown | Apple M2 unknown | lowering failure | — | — | — | — | — | — | — | — | — |
| bart-base | ALL | sm8750 ALL 3P | success | 2.00 | — | — | 3.8% | — | 205 | — | — | 0.0% |
| bart-base | ALL | sm8650 ALL 3P | success | 2.66 | — | — | 2.6% | — | 128 | — | — | 0.0% |
| bart-base | ALL | sm8550 ALL 3P | success | 3.74 | — | — | 1.7% | — | 93 | — | — | 0.0% |
| bart-base | ort-cpu-int8-dyn | Apple M2 CPU | success | 11.36 | — | — | 0.2% | 78 | 216 | 12 | 0.9805 | 0.0% |
| bart-base | unknown | Apple M2 unknown | success | 11.37 | — | — | 0.2% | 78 | 216 | — | 0.9805 | 0.0% |
| bart-base | unknown | Apple M2 unknown | success | 11.46 | — | — | 0.2% | 79 | 218 | — | 0.9908 | 0.0% |
| bart-base | ort-cpu-int8-perchan | Apple M2 CPU | success | 11.53 | — | — | 2.6% | 79 | 218 | 12 | 0.9908 | 0.0% |
| bart-base | ort-cpu-fp32 | Apple M2 CPU | success | 31.96 | — | — | 0.1% | 313 | 682 | 33 | 1.0000 | 0.0% |
| bart-base | ort-cpu-4thread | Apple M2 CPU | success | 32.00 | — | — | 0.1% | 313 | 681 | 33 | 1.0000 | 0.0% |
| bart-base | unknown | Apple M2 unknown | success | 32.05 | — | — | 0.1% | 313 | 681 | — | 1.0000 | 0.0% |
| bart-base | unknown | Apple M2 unknown | success | 32.11 | — | — | 0.1% | 313 | 681 | — | 1.0000 | 0.0% |
| bart-base | unknown | Apple M2 unknown | success | 33.07 | — | — | 0.2% | 313 | 681 | — | 1.0000 | 0.0% |
| bart-base | ort-cpu-dynamic | Apple M2 CPU | success | 33.10 | — | — | 0.3% | 313 | 681 | 34 | 1.0000 | 0.0% |
| bart-base | ort-coreml-fp32 | Apple M2 CoreML | success | 38.68 | — | — | 1.1% | 313 | 1582 | 48 | 1.0000 | 38.5% |
| bart-base | unknown | Apple M2 unknown | success | 38.97 | — | — | 0.6% | 313 | 1620 | — | 1.0000 | 31.8% |
| bart-base | unknown | Apple M2 unknown | success | 41.53 | — | — | 2.0% | 313 | 1583 | — | 1.0000 | 34.6% |
| bart-base | ort-coreml-ane | Apple M2 CoreML | success | 46.43 | — | — | 7.2% | 313 | 1580 | 50 | 1.0000 | 38.5% |
| bart-base | unknown | Apple M2 unknown | success | 53.42 | — | — | 0.5% | 313 | 1581 | — | 1.0000 | 16.3% |
| bart-base | ort-coreml-dynamic | Apple M2 CoreML | success | 58.02 | — | — | 13.8% | 313 | 1581 | 197 | 1.0000 | 17.2% |
| bart-base | ALL | google-tensor-g2 ALL 3P | success | 72.75 | — | — | 4.3% | — | 243 | — | — | 100.0% |
| bart-base | ALL | google-tensor-g4 ALL 3P | success | 93.00 | — | — | 5.6% | — | 265 | — | — | 100.0% |
| bart-base | ALL | sm8350 ALL 3P | success | 108.20 | — | — | 12.2% | — | 284 | — | — | 100.0% |
| bart-base | ALL | samsung-exynos-1280 ALL 3P | success | 224.50 | — | — | 4.2% | — | 283 | — | — | 100.0% |
| bart-base | ALL | sdm845 ALL 3P | success | 226.91 | — | — | 4.8% | — | 244 | — | — | 100.0% |
| bart-base | unknown | Apple M2 unknown | lowering failure | — | — | — | — | 156 | — | — | — | — |
| bart-base | ort-coreml-fp16 | Apple M2 CoreML | lowering failure | — | — | — | — | 156 | — | — | — | — |
| bart-base | unknown | Apple M2 unknown | lowering failure | — | — | — | — | — | — | — | — | — |
| bart-base | ort-coreml-mlprogram | Apple M2 CoreML | lowering failure | — | — | — | — | — | — | — | — | — |
| bart-base | ort-cpu-fp16 | Apple M2 CPU | lowering failure | — | — | — | — | 156 | — | — | — | — |
| bart-base | unknown | Apple M2 unknown | lowering failure | — | — | — | — | 156 | — | — | — | — |
| mobilenet_v2_1.0_224 | ALL | sm8650 ALL 3P | success | 0.35 | — | — | 7.9% | — | 160 | 716 | — | 0.0% |
| mobilenet_v2_1.0_224 | ALL | sm8650 ALL 3P | success | 0.38 | — | — | 13.1% | — | 142 | 553 | — | 0.0% |
| mobilenet_v2_1.0_224 | ALL | google-tensor-g4 ALL 3P | success | 4.71 | — | — | 29.2% | — | 151 | 1115 | — | 0.0% |
| mobilenet_v2_1.0_224 | ALL | google-tensor-g4 ALL 3P | success | 6.61 | — | — | 14.2% | — | 153 | 1275 | — | 0.0% |
| vit-base-patch16-224-in21k | ALL | sc8480xp ALL 3P | success | 4.42 | — | — | 1.3% | — | 34 | 11678 | — | 0.0% |
| vit-base-patch16-224-in21k | ALL | sm8750 ALL 3P | success | 6.26 | — | — | 1.6% | — | 230 | 5753 | — | 0.0% |
| vit-base-patch16-224-in21k | unknown | sm8750 unknown 3P | success | 6.26 | — | — | 1.5% | — | 214 | — | — | 0.0% |
| vit-base-patch16-224-in21k | ALL | sm8650 ALL 3P | success | 6.72 | — | — | 0.6% | — | 265 | 3538 | — | 0.0% |
| vit-base-patch16-224-in21k | unknown | sm8650 unknown 3P | success | 7.68 | — | — | 0.4% | — | 164 | — | — | 0.0% |
| vit-base-patch16-224-in21k | unknown | sm8650 unknown 3P | success | 7.68 | — | — | 0.7% | — | 163 | — | — | 0.0% |
| vit-base-patch16-224-in21k | ALL | sm8650 ALL 3P | success | 7.68 | — | — | 0.5% | — | 165 | 7728 | — | 0.0% |
| vit-base-patch16-224-in21k | unknown | sm8650 unknown 3P | success | 7.73 | — | — | 0.5% | — | 164 | — | — | 0.0% |
| vit-base-patch16-224-in21k | unknown | sm8650 unknown 3P | success | 7.75 | — | — | 0.5% | — | 167 | — | — | 0.0% |
| vit-base-patch16-224-in21k | ALL | sc8340xp ALL 3P | success | 10.70 | — | — | 0.8% | — | 196 | 11453 | — | 0.0% |
| vit-base-patch16-224-in21k | unknown | sm8550 unknown 3P | success | 10.94 | — | — | 0.9% | — | 253 | — | — | 0.0% |
| vit-base-patch16-224-in21k | unknown | sm8550 unknown 3P | success | 10.94 | — | — | 0.9% | — | 260 | — | — | 0.0% |
| vit-base-patch16-224-in21k | ALL | sm8550 ALL 3P | success | 11.01 | — | — | 1.2% | — | 249 | 7381 | — | 0.0% |
| vit-base-patch16-224-in21k | ALL | sc8380xp ALL 3P | success | 11.51 | — | — | 3.3% | — | 197 | 12710 | — | 0.0% |
| vit-base-patch16-224-in21k | ALL | qualcomm-sa8775p ALL 3P | success | 13.75 | — | — | 2.9% | — | 100 | 28509 | — | 0.0% |
| vit-base-patch16-224-in21k | ALL | qualcomm-sa8295p ALL 3P | success | 16.72 | — | — | 1.5% | — | 99 | 15170 | — | 0.0% |
| vit-base-patch16-224-in21k | ALL | qcs9075 ALL 3P | success | 17.28 | — | — | 2.2% | — | 16 | 10010 | — | 0.0% |
| vit-base-patch16-224-in21k | ort-cpu-int8-dyn | Apple M2 CPU | success | 42.70 | — | — | 0.2% | 83 | 200 | 44 | 0.9842 | 0.0% |
| vit-base-patch16-224-in21k | unknown | Apple M2 unknown | success | 42.84 | — | — | 0.1% | 83 | 200 | — | 0.9842 | 0.0% |
| vit-base-patch16-224-in21k | ort-cpu-int8-perchan | Apple M2 CPU | success | 43.04 | — | — | 0.2% | 83 | 200 | 44 | 0.9945 | 0.0% |
| vit-base-patch16-224-in21k | unknown | Apple M2 unknown | success | 43.11 | — | — | 0.1% | 83 | 200 | — | 0.9945 | 0.0% |
| vit-base-patch16-224-in21k | ALL | sm8650 ALL 3P | success | 56.17 | — | — | 0.4% | — | 457 | 43438 | — | 0.0% |
| vit-base-patch16-224-in21k | unknown | Apple M2 unknown | success | 71.54 | — | — | 2.5% | 327 | 671 | — | 1.0000 | 20.4% |
| vit-base-patch16-224-in21k | ort-coreml-ane | Apple M2 CoreML | success | 71.90 | — | — | 2.8% | 327 | 686 | 89 | 1.0000 | 21.9% |
| vit-base-patch16-224-in21k | ort-coreml-fp32 | Apple M2 CoreML | success | 73.60 | — | — | 1.6% | 327 | 670 | 88 | 1.0000 | 22.8% |
| vit-base-patch16-224-in21k | unknown | Apple M2 unknown | success | 75.63 | — | — | 3.7% | 327 | 604 | — | 1.0000 | 24.4% |
| vit-base-patch16-224-in21k | ort-cpu-dynamic | Apple M2 CPU | success | 104.59 | — | — | 0.4% | 327 | 432 | 105 | 1.0000 | 0.0% |
| vit-base-patch16-224-in21k | ort-cpu-fp32 | Apple M2 CPU | success | 104.77 | — | — | 0.1% | 327 | 445 | 106 | 1.0000 | 0.0% |
| vit-base-patch16-224-in21k | ort-cpu-4thread | Apple M2 CPU | success | 104.80 | — | — | 0.3% | 327 | 445 | 106 | 1.0000 | 0.0% |
| vit-base-patch16-224-in21k | unknown | Apple M2 unknown | success | 105.53 | — | — | 0.1% | 327 | 465 | — | 1.0000 | 0.0% |
| vit-base-patch16-224-in21k | unknown | Apple M2 unknown | success | 105.67 | — | — | 0.1% | 327 | 444 | — | 1.0000 | 0.0% |
| vit-base-patch16-224-in21k | unknown | Apple M2 unknown | success | 105.87 | — | — | 0.1% | 327 | 431 | — | 1.0000 | 0.0% |
| vit-base-patch16-224-in21k | ort-coreml-fp16 | Apple M2 CoreML | success | 109.80 | — | — | 0.2% | 164 | 772 | 112 | 1.0000 | 100.0% |
| vit-base-patch16-224-in21k | ort-cpu-fp16 | Apple M2 CPU | success | 109.80 | — | — | 0.1% | 164 | 771 | 112 | 1.0000 | 0.0% |
| vit-base-patch16-224-in21k | unknown | Apple M2 unknown | success | 110.14 | — | — | 0.1% | 164 | 720 | — | 1.0000 | 0.0% |
| vit-base-patch16-224-in21k | unknown | Apple M2 unknown | success | 110.42 | — | — | 0.1% | 164 | 749 | — | 1.0000 | 100.0% |
| vit-base-patch16-224-in21k | ALL | google-tensor-g4 ALL 3P | success | 146.70 | — | — | 9.6% | — | 254 | 1319 | — | 94.8% |
| vit-base-patch16-224-in21k | unknown | Apple M2 unknown | success | 179.36 | — | — | 0.9% | 327 | 969 | — | 1.0000 | 46.2% |
| vit-base-patch16-224-in21k | ort-coreml-dynamic | Apple M2 CoreML | success | 181.08 | — | — | 0.5% | 327 | 956 | 234 | 1.0000 | 46.9% |
| vit-base-patch16-224-in21k | ALL | google-tensor-g5 ALL 3P | success | 264.26 | — | — | 4.6% | — | 447 | 776 | — | 100.0% |
| vit-base-patch16-224-in21k | unknown | sm8650 unknown 3P | success | 267.62 | — | — | 9.7% | — | 453 | — | — | 0.0% |
| vit-base-patch16-224-in21k | unknown | google-tensor-g4 unknown 3P | success | 303.76 | — | — | 2.9% | — | 446 | — | — | 100.0% |
| vit-base-patch16-224-in21k | ALL | google-tensor-g4 ALL 3P | success | 306.09 | — | — | 9.0% | — | 424 | 155 | — | 100.0% |
| vit-base-patch16-224-in21k | ALL | google-tensor-g4 ALL 3P | success | 308.80 | — | — | 5.2% | — | 443 | 613 | — | 100.0% |
| vit-base-patch16-224-in21k | unknown | google-tensor-g2 unknown 3P | success | 328.49 | — | — | 18.5% | — | 430 | — | — | 100.0% |
| vit-base-patch16-224-in21k | unknown | samsung-exynos-990 unknown 3P | success | 351.44 | — | — | 6.1% | — | 462 | — | — | 100.0% |
| vit-base-patch16-224-in21k | ALL | sm7750 ALL 3P | success | 354.83 | — | — | 0.6% | — | 461 | 450 | — | 100.0% |
| vit-base-patch16-224-in21k | unknown | sm8350 unknown 3P | success | 380.07 | — | — | 6.7% | — | 458 | — | — | 100.0% |
| vit-base-patch16-224-in21k | ALL | sm7325 ALL 3P | success | 390.35 | — | — | 9.8% | — | 442 | 1075 | — | 100.0% |
| vit-base-patch16-224-in21k | unknown | google-tensor-g3 unknown 3P | success | 444.12 | — | — | 16.4% | — | 442 | — | — | 100.0% |
| vit-base-patch16-224-in21k | ALL | sm8250-ab ALL 3P | success | 463.10 | — | — | 7.5% | — | 486 | 1344 | — | 100.0% |
| vit-base-patch16-224-in21k | ALL | sm8150 ALL 3P | success | 465.26 | — | — | 4.8% | — | 438 | 1040 | — | 100.0% |
| vit-base-patch16-224-in21k | ALL | qcs6490 ALL 3P | success | 708.96 | — | — | 0.5% | — | 464 | 667 | — | 100.0% |
| vit-base-patch16-224-in21k | unknown | sm7250 unknown 3P | success | 750.98 | — | — | 2.2% | — | 420 | — | — | 100.0% |
| vit-base-patch16-224-in21k | unknown | samsung-exynos-1280 unknown 3P | success | 780.79 | — | — | 2.1% | — | 451 | — | — | 100.0% |
| vit-base-patch16-224-in21k | unknown | sdm845 unknown 3P | success | 820.37 | — | — | 4.2% | — | 429 | — | — | 100.0% |
| vit-base-patch16-224-in21k | ALL | sm6150-ac ALL 3P | success | 894.80 | — | — | 6.3% | — | 463 | 890 | — | 100.0% |
| vit-base-patch16-224-in21k | ALL | samsung-exynos-1330 ALL 3P | success | 987.58 | — | — | 12.1% | — | 468 | 1669 | — | 100.0% |
| vit-base-patch16-224-in21k | ort-coreml-mlprogram | Apple M2 CoreML | lowering failure | — | — | — | — | 327 | — | — | — | — |
| vit-base-patch16-224-in21k | unknown | Apple M2 unknown | lowering failure | — | — | — | — | 327 | — | — | — | — |
| toxic-comment-model | ALL | sm8750 ALL 3P | success | 2.26 | — | — | 3.5% | — | 233 | — | — | 0.0% |
| toxic-comment-model | ALL | sm8650 ALL 3P | success | 2.62 | — | — | 1.5% | — | 129 | — | — | 0.0% |
| toxic-comment-model | ALL | sm8550 ALL 3P | success | 3.74 | — | — | 1.1% | — | 184 | — | — | 0.0% |
| toxic-comment-model | ort-cpu-int8-dyn | Apple M2 CPU | success | 11.35 | — | — | 0.3% | 64 | 187 | 12 | 1.0000 | 0.0% |
| toxic-comment-model | unknown | Apple M2 unknown | success | 11.37 | — | — | 0.1% | 64 | 188 | — | 1.0000 | 0.0% |
| toxic-comment-model | ort-cpu-int8-perchan | Apple M2 CPU | success | 11.43 | — | — | 0.1% | 64 | 188 | 12 | 1.0000 | 0.0% |
| toxic-comment-model | unknown | Apple M2 unknown | success | 11.52 | — | — | 0.2% | 64 | 188 | — | 1.0000 | 0.0% |
| toxic-comment-model | ort-cpu-fp32 | Apple M2 CPU | success | 31.97 | — | — | 0.7% | 255 | 492 | 33 | 1.0000 | 0.0% |
| toxic-comment-model | unknown | Apple M2 unknown | success | 32.00 | — | — | 0.3% | 255 | 491 | — | 1.0000 | 0.0% |
| toxic-comment-model | ort-cpu-4thread | Apple M2 CPU | success | 32.00 | — | — | 0.3% | 255 | 491 | 33 | 1.0000 | 0.0% |
| toxic-comment-model | unknown | Apple M2 unknown | success | 32.09 | — | — | 0.1% | 255 | 490 | — | 1.0000 | 0.0% |
| toxic-comment-model | ort-cpu-dynamic | Apple M2 CPU | success | 33.16 | — | — | 0.2% | 256 | 491 | 34 | 1.0000 | 0.0% |
| toxic-comment-model | unknown | Apple M2 unknown | success | 33.37 | — | — | 0.2% | 256 | 491 | — | 1.0000 | 0.0% |
| toxic-comment-model | ort-coreml-fp16 | Apple M2 CoreML | success | 35.50 | — | — | 0.2% | 128 | 470 | 36 | 1.0000 | 100.0% |
| toxic-comment-model | unknown | Apple M2 unknown | success | 35.54 | — | — | 0.2% | 128 | 523 | — | 1.0000 | 0.0% |
| toxic-comment-model | ort-cpu-fp16 | Apple M2 CPU | success | 35.55 | — | — | 0.3% | 128 | 498 | 36 | 1.0000 | 0.0% |
| toxic-comment-model | unknown | Apple M2 unknown | success | 35.65 | — | — | 0.2% | 128 | 495 | — | 1.0000 | 100.0% |
| toxic-comment-model | ort-coreml-ane | Apple M2 CoreML | success | 37.92 | — | — | 4.2% | 255 | 1183 | 47 | 1.0000 | 39.6% |
| toxic-comment-model | ort-coreml-fp32 | Apple M2 CoreML | success | 38.45 | — | — | 4.7% | 255 | 1185 | 45 | 1.0000 | 40.3% |
| toxic-comment-model | unknown | Apple M2 unknown | success | 39.51 | — | — | 3.6% | 255 | 1186 | — | 1.0000 | 37.1% |
| toxic-comment-model | unknown | Apple M2 unknown | success | 41.23 | — | — | 4.4% | 255 | 1184 | — | 1.0000 | 38.8% |
| toxic-comment-model | ort-coreml-dynamic | Apple M2 CoreML | success | 51.75 | — | — | 0.5% | 256 | 1316 | 147 | 1.0000 | 19.3% |
| toxic-comment-model | unknown | Apple M2 unknown | success | 53.93 | — | — | 0.8% | 256 | 1267 | — | 1.0000 | 22.8% |
| toxic-comment-model | ALL | google-tensor-g2 ALL 3P | success | 91.83 | — | — | 9.1% | — | 255 | — | — | 100.0% |
| toxic-comment-model | ALL | google-tensor-g4 ALL 3P | success | 91.88 | — | — | 6.3% | — | 261 | — | — | 100.0% |
| toxic-comment-model | ALL | sm8350 ALL 3P | success | 109.25 | — | — | 11.2% | — | 264 | — | — | 100.0% |
| toxic-comment-model | ALL | sdm845 ALL 3P | success | 220.64 | — | — | 5.0% | — | 329 | — | — | 100.0% |
| toxic-comment-model | ALL | samsung-exynos-1280 ALL 3P | success | 236.29 | — | — | 6.3% | — | 282 | — | — | 100.0% |
| toxic-comment-model | unknown | Apple M2 unknown | lowering failure | — | — | — | — | — | — | — | — | — |
| toxic-comment-model | ort-coreml-mlprogram | Apple M2 CoreML | lowering failure | — | — | — | — | — | — | — | — | — |
| Llama-3.2-1B-Instruct | unknown | Apple M2 unknown | success | — | 118.3 | 12.48 | 5.7% | 5717 | 5810 | — | — | 0.0% |
| Llama-3.2-1B-Instruct | ort-cpu-dynamic | Apple M2 CPU | success | — | 124.1 | 12.09 | 29.7% | 5717 | 5810 | — | — | 0.0% |
| Llama-3.2-1B-Instruct | ort-coreml-dynamic | Apple M2 CoreML | success | — | 473.0 | 5.26 | 20.4% | 5717 | 7253 | — | — | 16.4% |
| Llama-3.2-1B-Instruct | unknown | Apple M2 unknown | success | — | 390.0 | 6.39 | 3.4% | 5717 | 8000 | — | — | 15.0% |
| resnet-50 | ALL | sm8650 ALL 3P | success | 0.65 | — | — | 6.7% | — | 257 | 652 | — | 0.0% |
| resnet-50 | ALL | sm8650 ALL 3P | success | 1.43 | — | — | 2.7% | — | 96 | 1824 | — | 0.0% |
| resnet-50 | ALL | sm8650 ALL 3P | success | 1.46 | — | — | 4.5% | — | 137 | 1338 | — | 0.0% |
| resnet-50 | ALL | sc8380xp ALL 3P | success | 1.87 | — | — | 8.9% | — | 77 | 5239 | — | 0.0% |
| resnet-50 | ALL | google-tensor-g5 ALL 3P | success | 62.36 | — | — | 4.7% | — | 207 | 147 | — | 100.0% |
| resnet-50 | ALL | google-tensor-g4 ALL 3P | success | 78.22 | — | — | 4.0% | — | 246 | 118 | — | 100.0% |
| resnet-50 | ALL | samsung-exynos-1330 ALL 3P | success | 234.65 | — | — | 0.8% | — | 340 | 379 | — | 100.0% |
| clip-vit-base-patch32 | unknown | sm8750 unknown 3P | success | 2.42 | — | — | 4.7% | — | 223 | — | — | 0.0% |
| clip-vit-base-patch32 | ALL | sm8750 ALL 3P | success | 2.43 | — | — | 4.3% | — | 226 | 3209 | — | 0.0% |
| clip-vit-base-patch32 | ALL | sm8650 ALL 3P | success | 2.79 | — | — | 1.9% | — | 115 | 4641 | — | 0.0% |
| clip-vit-base-patch32 | unknown | sm8650 unknown 3P | success | 2.79 | — | — | 2.7% | — | 113 | — | — | 0.0% |
| clip-vit-base-patch32 | unknown | sm8650 unknown 3P | success | 2.80 | — | — | 1.9% | — | 112 | — | — | 0.0% |
| clip-vit-base-patch32 | unknown | sm8550 unknown 3P | success | 3.63 | — | — | 1.3% | — | 248 | — | — | 0.0% |
| clip-vit-base-patch32 | ALL | sm8550 ALL 3P | success | 3.64 | — | — | 1.3% | — | 247 | 4631 | — | 0.0% |
| clip-vit-base-patch32 | ort-cpu-int8-dyn | Apple M2 CPU | success | 12.02 | — | — | 0.3% | 84 | 179 | 13 | 0.9923 | 0.0% |
| clip-vit-base-patch32 | unknown | Apple M2 unknown | success | 12.02 | — | — | 0.1% | 84 | 262 | — | 0.9923 | 0.0% |
| clip-vit-base-patch32 | ort-cpu-int8-perchan | Apple M2 CPU | success | 12.09 | — | — | 0.2% | 84 | 180 | 13 | 0.9910 | 0.0% |
| clip-vit-base-patch32 | unknown | Apple M2 unknown | success | 12.12 | — | — | 0.3% | 84 | 181 | — | 0.9910 | 0.0% |
| clip-vit-base-patch32 | unknown | Apple M2 unknown | success | 18.03 | — | — | 3.5% | 334 | 642 | — | 0.9999 | 8.2% |
| clip-vit-base-patch32 | unknown | Apple M2 unknown | success | 18.37 | — | — | 2.2% | 334 | 612 | — | 0.9999 | 9.6% |
| clip-vit-base-patch32 | ort-coreml-fp32 | Apple M2 CoreML | success | 18.40 | — | — | 3.4% | 334 | 624 | 25 | 0.9999 | 8.4% |
| clip-vit-base-patch32 | ort-coreml-ane | Apple M2 CoreML | success | 20.07 | — | — | 5.8% | 334 | 592 | 26 | 0.9999 | 8.3% |
| clip-vit-base-patch32 | ort-cpu-4thread | Apple M2 CPU | success | 28.29 | — | — | 0.3% | 334 | 436 | 29 | 1.0000 | 0.0% |
| clip-vit-base-patch32 | ort-cpu-fp32 | Apple M2 CPU | success | 28.32 | — | — | 0.2% | 334 | 436 | 29 | 1.0000 | 0.0% |
| clip-vit-base-patch32 | unknown | Apple M2 unknown | success | 28.35 | — | — | 0.1% | 334 | 436 | — | 1.0000 | 0.0% |
| clip-vit-base-patch32 | unknown | Apple M2 unknown | success | 28.35 | — | — | 0.1% | 334 | 436 | — | 1.0000 | 0.0% |
| clip-vit-base-patch32 | ort-cpu-dynamic | Apple M2 CPU | success | 28.52 | — | — | 0.6% | 334 | 434 | 29 | 1.0000 | 0.0% |
| clip-vit-base-patch32 | unknown | Apple M2 unknown | success | 28.60 | — | — | 0.1% | 334 | 474 | — | 1.0000 | 0.0% |
| clip-vit-base-patch32 | ort-coreml-dynamic | Apple M2 CoreML | success | 45.84 | — | — | 8.6% | 334 | 688 | 87 | 0.9999 | 36.0% |
| clip-vit-base-patch32 | unknown | Apple M2 unknown | success | 46.24 | — | — | 5.7% | 334 | 680 | — | 0.9999 | 31.7% |
| clip-vit-base-patch32 | unknown | google-tensor-g2 unknown 3P | success | 64.59 | — | — | 2.2% | — | 422 | — | — | 100.0% |
| clip-vit-base-patch32 | ALL | google-tensor-g4 ALL 3P | success | 76.50 | — | — | 4.7% | — | 428 | 584 | — | 100.0% |
| clip-vit-base-patch32 | unknown | google-tensor-g4 unknown 3P | success | 84.76 | — | — | 6.1% | — | 435 | — | — | 100.0% |
| clip-vit-base-patch32 | unknown | samsung-exynos-990 unknown 3P | success | 93.95 | — | — | 10.5% | — | 443 | — | — | 100.0% |
| clip-vit-base-patch32 | unknown | google-tensor-g3 unknown 3P | success | 96.61 | — | — | 5.2% | — | 425 | — | — | 100.0% |
| clip-vit-base-patch32 | unknown | sm8350 unknown 3P | success | 96.83 | — | — | 14.1% | — | 441 | — | — | 100.0% |
| clip-vit-base-patch32 | unknown | samsung-exynos-1280 unknown 3P | success | 178.66 | — | — | 7.2% | — | 434 | — | — | 100.0% |
| clip-vit-base-patch32 | unknown | sm7250 unknown 3P | success | 188.76 | — | — | 1.3% | — | 403 | — | — | 100.0% |
| clip-vit-base-patch32 | unknown | sdm845 unknown 3P | success | 204.64 | — | — | 4.3% | — | 440 | — | — | 100.0% |
| clip-vit-base-patch32 | ort-coreml-mlprogram | Apple M2 CoreML | lowering failure | — | — | — | — | 334 | — | — | — | — |
| clip-vit-base-patch32 | ort-coreml-fp16 | Apple M2 CoreML | lowering failure | — | — | — | — | 167 | — | — | — | — |
| clip-vit-base-patch32 | ort-cpu-fp16 | Apple M2 CPU | lowering failure | — | — | — | — | 167 | — | — | — | — |
| clip-vit-base-patch32 | unknown | Apple M2 unknown | lowering failure | — | — | — | — | 167 | — | — | — | — |
| clip-vit-base-patch32 | unknown | Apple M2 unknown | lowering failure | — | — | — | — | 167 | — | — | — | — |
| clip-vit-base-patch32 | unknown | Apple M2 unknown | lowering failure | — | — | — | — | 334 | — | — | — | — |
| all-MiniLM-L6-v2 | ALL | sc8480xp ALL 3P | success | 0.93 | — | — | 3.6% | — | 30 | 2927 | — | 0.0% |
| all-MiniLM-L6-v2 | ALL | sm8750 ALL 3P | success | 1.23 | — | — | 14.0% | — | 199 | 1090 | — | 0.0% |
| all-MiniLM-L6-v2 | ALL | sm8650 ALL 3P | success | 1.52 | — | — | 1.9% | — | 162 | 1879 | — | 0.0% |
| all-MiniLM-L6-v2 | ALL | sm8650 ALL 3P | success | 1.52 | — | — | 2.7% | — | 159 | — | — | 0.0% |
| all-MiniLM-L6-v2 | ALL | sm8550 ALL 3P | success | 2.18 | — | — | 3.7% | — | 101 | 1710 | — | 0.0% |
| all-MiniLM-L6-v2 | ALL | sc8340xp ALL 3P | success | 2.19 | — | — | 1.3% | — | 48 | 3783 | — | 0.0% |
| all-MiniLM-L6-v2 | ALL | sc8380xp ALL 3P | success | 2.24 | — | — | 1.7% | — | 48 | 3819 | — | 0.0% |
| all-MiniLM-L6-v2 | ALL | qcs9075 ALL 3P | success | 2.53 | — | — | 0.7% | — | 12 | 2484 | — | 0.0% |
| all-MiniLM-L6-v2 | ALL | qualcomm-sa8775p ALL 3P | success | 2.77 | — | — | 6.5% | — | 88 | 6093 | — | 0.0% |
| all-MiniLM-L6-v2 | ALL | qualcomm-sa8295p ALL 3P | success | 3.64 | — | — | 1.1% | — | 143 | 3473 | — | 0.0% |
| all-MiniLM-L6-v2 | ort-cpu-int8-dyn | Apple M2 CPU | success | 4.46 | — | — | 0.3% | 22 | 124 | 5 | 0.9785 | 0.0% |
| all-MiniLM-L6-v2 | unknown | Apple M2 unknown | success | 4.47 | — | — | 0.3% | 22 | 124 | — | 0.9785 | 0.0% |
| all-MiniLM-L6-v2 | ort-cpu-int8-perchan | Apple M2 CPU | success | 4.49 | — | — | 0.3% | 22 | 124 | 5 | 0.9924 | 0.0% |
| all-MiniLM-L6-v2 | unknown | Apple M2 unknown | success | 4.49 | — | — | 0.2% | 22 | 126 | — | 0.9924 | 0.0% |
| all-MiniLM-L6-v2 | ort-cpu-4thread | Apple M2 CPU | success | 9.59 | — | — | 0.7% | 86 | 228 | 10 | 1.0000 | 0.0% |
| all-MiniLM-L6-v2 | ort-cpu-fp32 | Apple M2 CPU | success | 9.59 | — | — | 0.2% | 86 | 228 | 10 | 1.0000 | 0.0% |
| all-MiniLM-L6-v2 | unknown | Apple M2 unknown | success | 9.62 | — | — | 0.2% | 86 | 228 | — | 1.0000 | 0.0% |
| all-MiniLM-L6-v2 | unknown | Apple M2 unknown | success | 9.90 | — | — | 5.8% | 86 | 233 | — | 1.0000 | 0.0% |
| all-MiniLM-L6-v2 | ort-cpu-dynamic | Apple M2 CPU | success | 10.38 | — | — | 0.3% | 86 | 223 | 11 | 1.0000 | 0.0% |
| all-MiniLM-L6-v2 | unknown | Apple M2 unknown | success | 10.38 | — | — | 0.2% | 86 | 222 | — | 1.0000 | 0.0% |
| all-MiniLM-L6-v2 | ort-cpu-fp16 | Apple M2 CPU | success | 11.37 | — | — | 0.4% | 43 | 202 | 12 | 1.0000 | 0.0% |
| all-MiniLM-L6-v2 | ort-coreml-fp16 | Apple M2 CoreML | success | 11.37 | — | — | 0.1% | 43 | 217 | 12 | 1.0000 | 100.0% |
| all-MiniLM-L6-v2 | unknown | Apple M2 unknown | success | 11.40 | — | — | 0.1% | 43 | 230 | — | 1.0000 | 100.0% |
| all-MiniLM-L6-v2 | unknown | Apple M2 unknown | success | 11.73 | — | — | 3.0% | 43 | 241 | — | 1.0000 | 0.0% |
| all-MiniLM-L6-v2 | ort-coreml-fp32 | Apple M2 CoreML | success | 18.99 | — | — | 4.5% | 86 | 630 | 24 | 1.0000 | 28.8% |
| all-MiniLM-L6-v2 | unknown | Apple M2 unknown | success | 20.11 | — | — | 0.6% | 86 | 631 | — | 1.0000 | 29.5% |
| all-MiniLM-L6-v2 | ort-coreml-ane | Apple M2 CoreML | success | 20.16 | — | — | 1.2% | 86 | 600 | 25 | 1.0000 | 31.1% |
| all-MiniLM-L6-v2 | ALL | google-tensor-g2 ALL 3P | success | 20.84 | — | — | 4.6% | — | 150 | — | — | 100.0% |
| all-MiniLM-L6-v2 | unknown | Apple M2 unknown | success | 20.94 | — | — | 0.3% | 86 | 598 | — | 1.0000 | 30.0% |
| all-MiniLM-L6-v2 | ALL | google-tensor-g5 ALL 3P | success | 21.14 | — | — | 5.8% | — | 132 | 179 | — | 100.0% |
| all-MiniLM-L6-v2 | ort-coreml-dynamic | Apple M2 CoreML | success | 21.44 | — | — | 0.6% | 86 | 627 | 87 | 1.0000 | 12.4% |
| all-MiniLM-L6-v2 | unknown | Apple M2 unknown | success | 21.81 | — | — | 0.2% | 86 | 626 | — | 1.0000 | 10.4% |
| all-MiniLM-L6-v2 | ALL | google-tensor-g4 ALL 3P | success | 22.73 | — | — | 8.2% | — | 136 | 170 | — | 100.0% |
| all-MiniLM-L6-v2 | ALL | sm7325 ALL 3P | success | 29.67 | — | — | 3.4% | — | 133 | 372 | — | 100.0% |
| all-MiniLM-L6-v2 | ALL | sm8350 ALL 3P | success | 30.00 | — | — | 5.9% | — | 148 | — | — | 100.0% |
| all-MiniLM-L6-v2 | ALL | sm8250-ab ALL 3P | success | 30.94 | — | — | 26.3% | — | 173 | 345 | — | 100.0% |
| all-MiniLM-L6-v2 | ALL | sm7750 ALL 3P | success | 31.49 | — | — | 4.4% | — | 207 | 160 | — | 100.0% |
| all-MiniLM-L6-v2 | ALL | sm8150 ALL 3P | success | 33.69 | — | — | 4.7% | — | 107 | 244 | — | 100.0% |
| all-MiniLM-L6-v2 | ALL | samsung-exynos-1330 ALL 3P | success | 60.42 | — | — | 7.6% | — | 184 | 562 | — | 100.0% |
| all-MiniLM-L6-v2 | ALL | sdm845 ALL 3P | success | 62.63 | — | — | 5.4% | — | 146 | — | — | 100.0% |
| all-MiniLM-L6-v2 | ALL | qcs6490 ALL 3P | success | 64.04 | — | — | 4.0% | — | 85 | 198 | — | 100.0% |
| all-MiniLM-L6-v2 | ALL | sm6150-ac ALL 3P | success | 64.40 | — | — | 7.5% | — | 227 | 328 | — | 100.0% |
| all-MiniLM-L6-v2 | ALL | samsung-exynos-1280 ALL 3P | success | 65.03 | — | — | 17.0% | — | 141 | — | — | 100.0% |
| all-MiniLM-L6-v2 | unknown | Apple M2 unknown | lowering failure | — | — | — | — | — | — | — | — | — |
| all-MiniLM-L6-v2 | ort-coreml-mlprogram | Apple M2 CoreML | lowering failure | — | — | — | — | — | — | — | — | — |
Every latency figure above describes the second run onward. The first one also loads the model, initialises the delegate and compiles kernels — and a user opening an app pays that, once, every cold start.
The ratio is not a constant offset that could be subtracted. On the same phone, forcing a model onto the CPU loaded far faster than letting it reach the NPU, so part of an accelerator's headline win is bought with startup cost the headline never shows.
| Model | Device | cold load ms | first run ms | steady ms | ratio |
|---|---|---|---|---|---|
| all-MiniLM-L6-v2 | sc8480xp 3P | 2927 | — | 0.93 | 3,157× |
| resnet-50 | sc8380xp 3P | 5239 | — | 1.87 | 2,807× |
| vit-base-patch16-224-in21k | sc8480xp 3P | 11678 | — | 4.42 | 2,639× |
| all-MiniLM-L6-v2 | qualcomm-sa8775p 3P | 6093 | — | 2.77 | 2,196× |
| vit-base-patch16-224-in21k | qualcomm-sa8775p 3P | 28509 | — | 13.75 | 2,073× |
| mobilenet_v2_1.0_224 | sm8650 3P | 716 | — | 0.35 | 2,018× |
| all-MiniLM-L6-v2 | sc8340xp 3P | 3783 | — | 2.19 | 1,730× |
| all-MiniLM-L6-v2 | sc8380xp 3P | 3819 | — | 2.24 | 1,706× |
| clip-vit-base-patch32 | sm8650 3P | 4641 | — | 2.79 | 1,665× |
| mobilenet_v2_1.0_224 | sm8650 3P | 553 | — | 0.38 | 1,475× |
| clip-vit-base-patch32 | sm8750 3P | 3209 | — | 2.43 | 1,321× |
| resnet-50 | sm8650 3P | 1824 | — | 1.43 | 1,278× |
The question a delegate never answers for you. Enabling an accelerator can make a model slower with no error and no warning — the silent-fallback failure mode. Both fallback columns are shown because they disagree, and the disagreement is the finding: see methodology.
Each row compares an accelerator against the CPU of the same device. Comparing across devices would measure the two devices, not the delegate.
| Model | Task | Device | CPU fp32 | accelerated | verdict | FLOP fb (as authored) | time fb (as run) |
|---|---|---|---|---|---|---|---|
| toxic-comment-model | classify | Apple M2 | 31.97 | 11.37 | 2.81× faster | 0.0% | 0.0% |
| bart-base | embed | Apple M2 | 31.96 | 11.37 | 2.81× faster | 0.0% | 0.0% |
| vit-base-patch16-224-in21k | vision | Apple M2 | 104.59 | 42.84 | 2.44× faster | 0.0% | 0.0% |
| clip-vit-base-patch32 | unknown | Apple M2 | 28.29 | 12.02 | 2.35× faster | 0.0% | 0.0% |
| all-MiniLM-L6-v2 | embed | Apple M2 | 9.59 | 4.47 | 2.15× faster | 0.0% | 0.0% |