EdgeFit Atlas

Measured on-device inference. Every number reproducible; every failure recorded.

vit-base-patch16-224-in21k

hf:google/vit-base-patch16-224-in21k · vision

Graph fingerprint

85.8M graph parameters (as exported) · 1013 graph nodes · attention: mha · norm: layernorm

Constant×324, Add×121, Unsqueeze×103, MatMul×96, Concat×52, Reshape×51, Mul×50, Transpose×49

Graph size by weight dtype: fp32 632 nodes, fp16 634 nodes, int8 902 nodes, fp32 1013 nodes. Quantization inserts quantize/dequantize pairs, so the node count grows even as the artifact shrinks.

Latency by recipe

CPUaccelerator
0250500750p50 latency, msALL · sc8480xp4.42ALL · sc8480xp: 4.42 — cv 1.3%ALL · sm87506.26ALL · sm8750: 6.26 — cv 1.6%unknown · sm87506.26unknown · sm8750: 6.26 — cv 1.5%ALL · sm86507.68ALL · sm8650: 7.68 — cv 0.5%unknown · sm86507.70unknown · sm8650: 7.70 — cv 0.4%, 2 repeats spanning 7.68–7.73 msunknown · sm86507.75unknown · sm8650: 7.75 — cv 0.5%ALL · sc8340xp10.70ALL · sc8340xp: 10.70 — cv 0.8%unknown · sm855010.94unknown · sm8550: 10.94 — cv 0.9%unknown · sm855010.94unknown · sm8550: 10.94 — cv 0.9%ALL · sm855011.01ALL · sm8550: 11.01 — cv 1.2%ALL · sc8380xp11.51ALL · sc8380xp: 11.51 — cv 3.3%ALL · qualcomm-sa8775p13.75ALL · qualcomm-sa8775p: 13.75 — cv 2.9%ALL · qualcomm-sa8295p16.72ALL · qualcomm-sa8295p: 16.72 — cv 1.5%ALL · qcs907517.28ALL · qcs9075: 17.28 — cv 2.2%ALL · sm865031.45ALL · sm8650: 31.45 — cv 0.6%, 2 repeats spanning 6.72–56.17 msort-cpu-int8-dyn · Apple M242.70ort-cpu-int8-dyn · Apple M2: 42.70 — cv 0.2%, 83 MiBort-cpu-int8-perchan · Apple M243.04ort-cpu-int8-perchan · Apple M2: 43.04 — cv 0.2%, 83 MiBort-coreml-ane · Apple M271.90ort-coreml-ane · Apple M2: 71.90 — cv 2.8%, 327 MiBort-coreml-fp32 · Apple M273.60ort-coreml-fp32 · Apple M2: 73.60 — cv 1.6%, 327 MiBort-cpu-dynamic · Apple M2104.59ort-cpu-dynamic · Apple M2: 104.59 — cv 0.4%, 327 MiBort-cpu-fp32 · Apple M2104.77ort-cpu-fp32 · Apple M2: 104.77 — cv 0.1%, 327 MiBort-cpu-4thread · Apple M2104.80ort-cpu-4thread · Apple M2: 104.80 — cv 0.3%, 327 MiBunknown · Apple M2105.60unknown · Apple M2: 105.60 — cv 0.1%, 83 MiB, 10 repeats spanning 42.84–179.36 msort-coreml-fp16 · Apple M2109.80ort-coreml-fp16 · Apple M2: 109.80 — cv 0.2%, 164 MiBort-cpu-fp16 · Apple M2109.80ort-cpu-fp16 · Apple M2: 109.80 — cv 0.1%, 164 MiBunknown · sm8650137.65unknown · sm8650: 137.65 — cv 0.7%, 2 repeats spanning 7.68–267.62 msort-coreml-dynamic · Apple M2181.08ort-coreml-dynamic · Apple M2: 181.08 — cv 0.5%, 327 MiBALL · google-tensor-g5264.26ALL · google-tensor-g5: 264.26 — cv 4.6%unknown · google-tensor-g4303.76unknown · google-tensor-g4: 303.76 — cv 2.9%ALL · google-tensor-g4306.09ALL · google-tensor-g4: 306.09 — cv 9.6%, 3 repeats spanning 146.70–308.80 msunknown · google-tensor-g2328.49unknown · google-tensor-g2: 328.49 — cv 18.5%unknown · samsung-exynos-990351.44unknown · samsung-exynos-990: 351.44 — cv 6.1%ALL · sm7750354.83ALL · sm7750: 354.83 — cv 0.6%unknown · sm8350380.07unknown · sm8350: 380.07 — cv 6.7%ALL · sm7325390.35ALL · sm7325: 390.35 — cv 9.8%unknown · google-tensor-g3444.12unknown · google-tensor-g3: 444.12 — cv 16.4%ALL · sm8250-ab463.10ALL · sm8250-ab: 463.10 — cv 7.5%ALL · sm8150465.26ALL · sm8150: 465.26 — cv 4.8%ALL · qcs6490708.96ALL · qcs6490: 708.96 — cv 0.5%unknown · sm7250750.98unknown · sm7250: 750.98 — cv 2.2%unknown · samsung-exynos-1280780.79unknown · samsung-exynos-1280: 780.79 — cv 2.1%unknown · sdm845820.37unknown · sdm845: 820.37 — cv 4.2%ALL · sm6150-ac894.80ALL · sm6150-ac: 894.80 — cv 6.3%ALL · samsung-exynos-1330987.58ALL · samsung-exynos-1330: 987.58 — cv 12.1%
Median of 10–97 timed runs after 3 discarded warmups. Hover a bar for variance and artifact size.

Speed against numerics

CPUaccelerator
0.9850.990.995150100150p50 latency, ms · lower is bettercosine vs fp32 referenceort-cpu-int8-dyn · Apple M2 — cv 0.2%, 83 MiBort-cpu-int8-perchan · Apple M2 — cv 0.2%, 83 MiBort-coreml-ane · Apple M2 — cv 2.8%, 327 MiBort-coreml-fp32 · Apple M2 — cv 1.6%, 327 MiBort-cpu-dynamic · Apple M2 — cv 0.4%, 327 MiBort-cpu-fp32 · Apple M2 — cv 0.1%, 327 MiBort-cpu-4thread · Apple M2 — cv 0.3%, 327 MiBunknown · Apple M2 — cv 0.1%, 83 MiB, 10 repeats spanning 42.84–179.36 msort-coreml-fp16 · Apple M2 — cv 0.2%, 164 MiBort-cpu-fp16 · Apple M2 — cv 0.1%, 164 MiBort-coreml-dynamic · Apple M2 — cv 0.5%, 327 MiB
Marker area scales with artifact size. Cosine is a numerics check against the fp32 PyTorch reference, not task accuracy — a model can hold cosine 0.999 and still fail on the slice that matters.

Repeatability

Recipes measured more than once on this unit, in separate sessions. This is the weakest useful form of the check — the real one runs the same recipe on two physical units of the same SKU, and we have one machine. Agreement here cannot catch a defect in the methodology that both runs share; it can only catch drift.

Recipesessionsfastest msslowest msspread
unknown · sm865027.687.730.64%
ALL · sm865026.7256.17735.23%
unknown · Apple M21042.84179.36318.70%
unknown · sm865027.68267.623384.58%
ALL · google-tensor-g43146.70308.80110.50%

Every recipe measured

RecipeTargetOutcomep50 mscvsize MiBcosineFLOP fb (auth)time fb (run)partitions
ALLALLsuccess4.421.3%0.0%
ALLALLsuccess6.261.6%0.0%
unknownunknownsuccess6.261.5%0.0%
ALLALLsuccess6.720.6%0.0%
unknownunknownsuccess7.680.4%0.0%
unknownunknownsuccess7.680.7%0.0%
ALLALLsuccess7.680.5%0.0%
unknownunknownsuccess7.730.5%0.0%
unknownunknownsuccess7.750.5%0.0%
ALLALLsuccess10.700.8%0.0%
unknownunknownsuccess10.940.9%0.0%
unknownunknownsuccess10.940.9%0.0%
ALLALLsuccess11.011.2%0.0%
ALLALLsuccess11.513.3%0.0%
ALLALLsuccess13.752.9%0.0%
ALLALLsuccess16.721.5%0.0%
ALLALLsuccess17.282.2%0.0%
ort-cpu-int8-dynCPUsuccess42.700.2%830.98420.0%0.0%
unknownunknownsuccess42.840.1%830.98420.0%0.0%
ort-cpu-int8-perchanCPUsuccess43.040.2%830.99450.0%0.0%
unknownunknownsuccess43.110.1%830.99450.0%0.0%
ALLALLsuccess56.170.4%0.0%
unknownunknownsuccess71.542.5%3271.000099.1%20.4%62
ort-coreml-aneCoreMLsuccess71.902.8%3271.000099.1%21.9%62
ort-coreml-fp32CoreMLsuccess73.601.6%3271.000099.1%22.8%62
unknownunknownsuccess75.633.7%3271.000099.1%24.4%62
ort-cpu-dynamicCPUsuccess104.590.4%3271.00000.0%
ort-cpu-fp32CPUsuccess104.770.1%3271.00000.0%0.0%
ort-cpu-4threadCPUsuccess104.800.3%3271.00000.0%0.0%
unknownunknownsuccess105.530.1%3271.00000.0%0.0%
unknownunknownsuccess105.670.1%3271.00000.0%0.0%
unknownunknownsuccess105.870.1%3271.00000.0%
ort-coreml-fp16CoreMLsuccess109.800.2%1641.000099.9%100.0%
ort-cpu-fp16CPUsuccess109.800.1%1641.00000.0%0.0%
unknownunknownsuccess110.140.1%1641.00000.0%0.0%
unknownunknownsuccess110.420.1%1641.000099.9%100.0%
ALLALLsuccess146.709.6%94.8%
unknownunknownsuccess179.360.9%3271.000046.2%86
ort-coreml-dynamicCoreMLsuccess181.080.5%3271.000046.9%86
ALLALLsuccess264.264.6%100.0%
unknownunknownsuccess267.629.7%0.0%
unknownunknownsuccess303.762.9%100.0%
ALLALLsuccess306.099.0%100.0%
ALLALLsuccess308.805.2%100.0%
unknownunknownsuccess328.4918.5%100.0%
unknownunknownsuccess351.446.1%100.0%
ALLALLsuccess354.830.6%100.0%
unknownunknownsuccess380.076.7%100.0%
ALLALLsuccess390.359.8%100.0%
unknownunknownsuccess444.1216.4%100.0%
ALLALLsuccess463.107.5%100.0%
ALLALLsuccess465.264.8%100.0%
ALLALLsuccess708.960.5%100.0%
unknownunknownsuccess750.982.2%100.0%
unknownunknownsuccess780.792.1%100.0%
unknownunknownsuccess820.374.2%100.0%
ALLALLsuccess894.806.3%100.0%
ALLALLsuccess987.5812.1%100.0%
ort-coreml-mlprogramCoreMLlowering failure327
unknownunknownlowering failure327

Recorded failures

Kept rather than hidden. A recipe that cannot lower is a fact about the toolchain, and it is what teaches the static filter not to propose it again.

ort-coreml-mlprogram lowering failure
session creation failed: [ONNXRuntimeError] : 1 : FAIL : Failed to create MLModel, error: Failed to build the model execution plan using a model architecture file '/private/var/folders/g4/qr8j7fw12hv68l1g1pyv13t40000gn/T/onnxruntime-B179AF2A-DE7B-4C2D-B6F2-A7B06936E9F5-72551-000001D812654D61.mlmodelc/model.mil' with error code: -7.
unknown lowering failure
session creation failed: [ONNXRuntimeError] : 1 : FAIL : Failed to create MLModel, error: Failed to build the model execution plan using a model architecture file '/private/var/folders/g4/qr8j7fw12hv68l1g1pyv13t40000gn/T/onnxruntime-F76607A2-1516-4C94-9800-6C0B9DC1DC25-26589-0000003B2F81E921.mlmodelc/model.mil' with error code: -7.

Reproduce

Commands for every row on this page
uv run edgefit measure-remote --model hf:google/vit-base-patch16-224-in21k --device "Snapdragon X2 Elite CRD" --compute-unit all
uv run edgefit measure-remote --model hf:google/vit-base-patch16-224-in21k --device "Snapdragon 8 Elite QRD" --compute-unit all
uv run edgefit measure-remote --model hf:google/vit-base-patch16-224-in21k --device "Snapdragon 8 Elite QRD" --compute-unit unknown
uv run edgefit measure-remote --model hf:google/vit-base-patch16-224-in21k --device "Samsung Galaxy S24 (Family)" --compute-unit all
uv run edgefit measure-remote --model hf:google/vit-base-patch16-224-in21k --device "Samsung Galaxy S24" --compute-unit unknown
uv run edgefit measure-remote --model hf:google/vit-base-patch16-224-in21k --device "Samsung Galaxy S24 (Family)" --compute-unit unknown
uv run edgefit measure-remote --model hf:google/vit-base-patch16-224-in21k --device "Samsung Galaxy S24" --compute-unit all
uv run edgefit measure-remote --model hf:google/vit-base-patch16-224-in21k --device "Samsung Galaxy S24+" --compute-unit unknown
uv run edgefit measure-remote --model hf:google/vit-base-patch16-224-in21k --device "Samsung Galaxy S24 Ultra" --compute-unit unknown
uv run edgefit measure-remote --model hf:google/vit-base-patch16-224-in21k --device "Snapdragon X Plus 8-Core CRD" --compute-unit all
uv run edgefit measure-remote --model hf:google/vit-base-patch16-224-in21k --device "Samsung Galaxy S23" --compute-unit unknown
uv run edgefit measure-remote --model hf:google/vit-base-patch16-224-in21k --device "Samsung Galaxy S23 Ultra" --compute-unit unknown
uv run edgefit measure-remote --model hf:google/vit-base-patch16-224-in21k --device "Samsung Galaxy S23" --compute-unit all
uv run edgefit measure-remote --model hf:google/vit-base-patch16-224-in21k --device "Snapdragon X Elite CRD" --compute-unit all
uv run edgefit measure-remote --model hf:google/vit-base-patch16-224-in21k --device "SA8775P ADP" --compute-unit all
uv run edgefit measure-remote --model hf:google/vit-base-patch16-224-in21k --device "SA8295P ADP" --compute-unit all
uv run edgefit measure-remote --model hf:google/vit-base-patch16-224-in21k --device "Dragonwing IQ-9075 EVK" --compute-unit all
uv run edgefit measure --model hf:google/vit-base-patch16-224-in21k --recipe recipes/ort_cpu_int8_dynamic.yaml
# recipe c77ebc6856100da6 is no longer in the library
uv run edgefit measure --model hf:google/vit-base-patch16-224-in21k --recipe recipes/ort_cpu_int8_perchannel.yaml
# recipe c7282d73d5b0c996 is no longer in the library
uv run edgefit measure-remote --model hf:google/vit-base-patch16-224-in21k --device "Samsung Galaxy S24 (Family)" --compute-unit all
# recipe 229ec344c452b098 is no longer in the library
uv run edgefit measure --model hf:google/vit-base-patch16-224-in21k --recipe recipes/ort_coreml_ane.yaml
uv run edgefit measure --model hf:google/vit-base-patch16-224-in21k --recipe recipes/ort_coreml_fp32.yaml
# recipe 5d7949732ada84a4 is no longer in the library
uv run edgefit measure --model hf:google/vit-base-patch16-224-in21k --recipe recipes/ort_cpu_dynamic.yaml
uv run edgefit measure --model hf:google/vit-base-patch16-224-in21k --recipe recipes/ort_cpu_fp32.yaml
uv run edgefit measure --model hf:google/vit-base-patch16-224-in21k --recipe recipes/ort_cpu_4thread.yaml
# recipe 6d549e6069d0883f is no longer in the library
# recipe b6fda8be500c01da is no longer in the library
# recipe 5195071c42cf5b4a is no longer in the library
uv run edgefit measure --model hf:google/vit-base-patch16-224-in21k --recipe recipes/ort_coreml_fp16.yaml
uv run edgefit measure --model hf:google/vit-base-patch16-224-in21k --recipe recipes/ort_cpu_fp16.yaml
# recipe 30bf417fcb4ac0aa is no longer in the library
# recipe f49f992bae98c992 is no longer in the library
uv run edgefit measure-remote --model hf:google/vit-base-patch16-224-in21k --device "Google Pixel 9" --compute-unit all
# recipe 4607718be4410e77 is no longer in the library
uv run edgefit measure --model hf:google/vit-base-patch16-224-in21k --recipe recipes/ort_coreml_dynamic.yaml
uv run edgefit measure-remote --model hf:google/vit-base-patch16-224-in21k --device "Google Pixel 10" --compute-unit all
uv run edgefit measure-remote --model hf:google/vit-base-patch16-224-in21k --device "Samsung Galaxy S24 (Family)" --compute-unit unknown
uv run edgefit measure-remote --model hf:google/vit-base-patch16-224-in21k --device "Google Pixel 9" --compute-unit unknown
uv run edgefit measure-remote --model hf:google/vit-base-patch16-224-in21k --device "Google Pixel 9" --compute-unit all
uv run edgefit measure-remote --model hf:google/vit-base-patch16-224-in21k --device "Google Pixel 9" --compute-unit all
uv run edgefit measure-remote --model hf:google/vit-base-patch16-224-in21k --device "Google Pixel 7" --compute-unit unknown
uv run edgefit measure-remote --model hf:google/vit-base-patch16-224-in21k --device "Samsung Galaxy Note 20 (Intl)" --compute-unit unknown
uv run edgefit measure-remote --model hf:google/vit-base-patch16-224-in21k --device "Snapdragon 7 Gen 4 QRD" --compute-unit all
uv run edgefit measure-remote --model hf:google/vit-base-patch16-224-in21k --device "Samsung Galaxy S21" --compute-unit unknown
uv run edgefit measure-remote --model hf:google/vit-base-patch16-224-in21k --device "Samsung Galaxy A73 5G" --compute-unit all
uv run edgefit measure-remote --model hf:google/vit-base-patch16-224-in21k --device "Google Pixel 8" --compute-unit unknown
uv run edgefit measure-remote --model hf:google/vit-base-patch16-224-in21k --device "Samsung Galaxy Tab S7" --compute-unit all
uv run edgefit measure-remote --model hf:google/vit-base-patch16-224-in21k --device "Google Pixel 4" --compute-unit all
uv run edgefit measure-remote --model hf:google/vit-base-patch16-224-in21k --device "Dragonwing RB3 Gen 2 Vision Kit" --compute-unit all
uv run edgefit measure-remote --model hf:google/vit-base-patch16-224-in21k --device "Google Pixel 5" --compute-unit unknown
uv run edgefit measure-remote --model hf:google/vit-base-patch16-224-in21k --device "Samsung Galaxy A53 5G" --compute-unit unknown
uv run edgefit measure-remote --model hf:google/vit-base-patch16-224-in21k --device "Google Pixel 3" --compute-unit unknown
uv run edgefit measure-remote --model hf:google/vit-base-patch16-224-in21k --device "Xiaomi Redmi Note 10 5G" --compute-unit all
uv run edgefit measure-remote --model hf:google/vit-base-patch16-224-in21k --device "Samsung Galaxy A14 5G" --compute-unit all
uv run edgefit measure --model hf:google/vit-base-patch16-224-in21k --recipe recipes/ort_coreml_mlprogram.yaml
# recipe f89aae4d09e71940 is no longer in the library