March 2026 — 10 edge models, ~100 devices, 913 benchmarks on real silicon

Edge AI Benchmarks

913 benchmarks across ~100 real devices — phones, tablets, MCUs, automotive, XR headsets. 10 edge models across 8 domains, profiled on Snapdragon NPUs, Apple ANE, LiteRT, and bare-metal Cortex-M. Every model sub-60 μs on the best phone. The numbers below aren't projections. They're 19 hours of real silicon.

Benchmarks

~100

Devices

0 μs

Fastest (NPU)

0 μs

Fastest (LiteRT)

Passed

Models

Edge Models

10 models, all sub-1 MB, all sub-60μs on the best Snapdragon NPU. Every model profiled across 72 phone/tablet/IoT devices. Fastest per model highlighted.

Scroll horizontally to see all columns

Model	Accuracy	Snapdragon	ANE	LiteRT	Energy	Size
Wake WordVoice	98.1%	32 μs	118 μs	349 μs	25.7 μJ	229 KB
ECG ArrhythmiaCardiac	~100%	36 μs	41 μs	15 μs	2.1 μJ	101 KB
TS Anomaly DetectorTime Series	100%	39 μs	38 μs	7 μs	1.0 μJ	36 KB
NanoVision CIFAR-10Vision	84.9%	41 μs	83 μs	36 μs	5.4 μJ	83 KB
FinSense ENFinance	83.9%	42 μs	60 μs	—	7.5 μJ	632 KB
Speech EmotionAudio	62.2%	45 μs	109 μs	83 μs	12.5 μJ	229 KB
Bearing AnomalyIndustrial	100%	46 μs	51 μs	32 μs	4.3 μJ	40 KB
HAR ActivityIMU	95.6%	54 μs	49 μs	35 μs	5.1 μJ	29 KB
Audio EventAudio	87.5%	55 μs	142 μs	127 μs	16.6 μJ	230 KB
Fall DetectionIMU	97.5%	58 μs	60 μs	57 μs	7.4 μJ	29 KB

Qualcomm AI Hub · 72 devicesCoreML · M4 Max ANEGoogle LiteRT · XNNPACK

Honest note: Speech Emotion (62.2%) and FinSense edge (83.9%) are the weakest performers — audio emotion is a hard problem at sub-1 MB, and the FinSense edge variant trades accuracy for 180× size reduction. Distillation continues. We publish the real numbers, not just the good ones.

Domain Models

Purpose-built models for legal and tax work. They understand your documents, classify them accurately, and run on your infrastructure — so client data never leaves your network.

LawSense

Legal

99.3%

Accuracy

0.992 macro F1

F1 Score

704 MB

Model Size

Drop 10,000 documents from discovery. LawSense sorts them — contract, pleading, correspondence, regulation, opinion — at 99.3% accuracy. On the firm’s own hardware. Client files never leave the building.

DeBERTa-v3GPU Inference

TaxSense

Tax

86%

Accuracy

0.86 macro F1

F1 Score

<5 MB

Model Size

Reads tax rulings and classifies them: deductions, credits, penalties, procedure. Then shows you exactly which phrases drove the decision. Not a black box — it explains itself. That matters when the IRS asks.

DeBERTa-v3GPU Inference

Multi-Platform Latency

Side-by-side across Snapdragon NPU, Apple ANE, and Google LiteRT. Plus 29 MCU/SBC boards via Edge Impulse. LiteRT dominates signal models. Snapdragon leads audio.

Energy per Inference

Measured via IOReport on M4 Max. Microjoules per inference — real power, not estimates.

CoreML/ANE vs LiteRT/XNNPACK (μJ)

Best: TS Anomaly at 1.01 μJ (LiteRT)

Inferences per AA Battery

AA = 9,360 J. Best energy path per model.

TS Anomaly9.3B

ECG4.5B

Bearing2.2B

HAR1.8B

NanoVision1.7B

Fall1.3B

FinSense1.3B

Speech Emotion746M

Audio Event563M

Wake Word365M

Sustained Throughput

Inferences per second over 3-second burst. Peak: 134,775 inf/s (TS Anomaly, LiteRT).

Distillation: 420 MB → 1.46 MB

288x compression from BERT teacher to NanoCNN student. The sub-2MB Pareto frontier.

BERT TeacherDistilBERTNanoCNN FP32NanoCNN INT8

Platform Strengths

Four platform families, distinct strengths.

Google LiteRT

XNNPACK on M4 Max CPU

7 μs

134K inf/s · 1.01 μJ

Fastest on most models. Dominates 1D signals (TS, ECG, HAR, Bearing, Vision). Best throughput and energy. Runs anywhere with a CPU.

Apple M4 Max ANE

CoreML Neural Engine

38 μs

26K inf/s · 3.91 μJ

Dedicated neural hardware for sustained always-on inference. Best on wake word and audio. Real power via IOReport.

Snapdragon NPU

72 devices · Pixel 3 to X2 Elite

32 μs

30× faster than ResNet-50

All 10 models sub-60μs on latest phones. 72 real devices profiled via QAI Hub. Thermally invisible — 0.0°C delta at max throughput.

vs. Standard Models

Same hardware, same NPU, same measurement. ResNet-50, EfficientNet-B0, MobileNetV2 — all three baselines profiled on 19 Snapdragon devices. Our models are 5-30× faster and 20-3,350× smaller.

Snapdragon NPU Latency Comparison

ResNet-50 (97.4 MB)964 μs Standard baseline

EfficientNet-B0 (20.2 MB)655 μs "Efficient" baseline

MobileNetV2 (13.3 MB)294 μs Mobile baseline

Fall Detection (29 KB)58 μs 5-17× faster

TS Anomaly (36 KB)39 μs 8-25× faster

Wake Word (229 KB)32 μs 9-30× faster

All on same Snapdragon NPU via Qualcomm AI Hub. 57/57 baseline jobs completed, zero failures. Our models: 5-30\u00d7 faster and 20-3,350\u00d7 smaller than every standard baseline.

Thermal Impact

At maximum sustained throughput (10,000+ inferences/sec for 60 seconds), our models produce zero measurable temperature increase on the host device.

Bearing Anomaly

603,654 inferences at 10,061/s

0.0°C

TS Anomaly

226,172 inferences at 3,770/s

0.0°C

60-second sustained burst on MacBook Pro M4 Max. No throttling detected.

FinSense Live

Real market headlines scored for sentiment. Demo — production model runs on-device at 42 μs.

Live headlines via Finnhub · Demo sentiment

Device Compatibility

913 benchmarks across 101 real devices. Phones, tablets, MCUs, automotive, XR. Search your device — see the exact latency. We publish every result, including the 67 failures.

913

Benchmarks

101

Devices

93.8%

Pass Rate

39 μs

Fastest (NPU)

<1 ms

Fastest (MCU)

Failures

Wake Word

Voice · 229 KB

97/101

NPU best: 32 μs

Throughput: 31,250/s

MCU RAM: 322 KB

CR2032: 2 days

ECG Arrhythmia

Cardiac · 101 KB

71/72

NPU best: 36 μs

Throughput: 27,777/s

TS Anomaly

Time Series · 36 KB

97/101

NPU best: 39 μs

Throughput: 25,641/s

MCU RAM: 16 KB

CR2032: 2 days

NanoVision

Vision · 83 KB

98/101

NPU best: 41 μs

Throughput: 24,390/s

MCU RAM: 258 KB

CR2032: 2 days

FinSense EN

Finance · 632 KB

71/101

NPU best: 42 μs

Throughput: 23,809/s

Speech Emotion

Audio · 229 KB

71/72

NPU best: 45 μs

Throughput: 22,222/s

Bearing Anomaly

Industrial · 40 KB

98/101

NPU best: 46 μs

Throughput: 21,739/s

MCU RAM: 82 KB

CR2032: 2 days

HAR Activity

IMU · 29 KB

98/101

NPU best: 54 μs

Throughput: 18,518/s

MCU RAM: 259 KB

CR2032: 2 days

Audio Event

Audio · 230 KB

57/72

NPU best: 55 μs

Throughput: 18,181/s

Fall Detection

IMU · 29 KB

98/101

NPU best: 58 μs

Throughput: 17,241/s

MCU RAM: 403 KB

CR2032: 2 days

923 results

Scroll horizontally to see all columns

Device	Chip	Model	Latency	Energy	Thru/s	Status	Source
Snapdragon X2 Elite CRD	Snapdragon X2 Elite	Wake Word	32μs	—	31,250	PASS	QAI
Snapdragon X2 Elite CRD	Snapdragon X2 Elite	ECG Arrhythmia	36μs	—	27,777	PASS	QAI
Google Pixel 10 Pro XL	Snapdragon 8s Elite	TS Anomaly	39μs	—	25,641	PASS	QAI
Snapdragon X2 Elite CRD	Snapdragon X2 Elite	TS Anomaly	40μs	—	25,000	PASS	QAI
Snapdragon X2 Elite CRD	Snapdragon X2 Elite	NanoVision	41μs	—	24,390	PASS	QAI
Snapdragon X Plus 8-Core CRD	Snapdragon X Plus	Wake Word	42μs	—	23,809	PASS	QAI
Snapdragon X2 Elite CRD	Snapdragon X2 Elite	FinSense EN	42μs	—	23,809	PASS	QAI
Google Pixel 9 Pro	Tensor G4	TS Anomaly	43μs	—	23,255	PASS	QAI
Samsung Galaxy S25+	Snapdragon 8 Elite	TS Anomaly	43μs	—	23,255	PASS	QAI
Snapdragon X2 Elite CRD	Snapdragon X2 Elite	Speech Emotion	45μs	—	22,222	PASS	QAI
Snapdragon 8 Elite QRD	Snapdragon 8 Elite	Wake Word	46μs	—	21,739	PASS	QAI
Samsung Galaxy S25 (Family)	Unknown	NanoVision	46μs	—	21,739	PASS	QAI
Snapdragon 8 Elite Gen 5 QRD	Snapdragon 8 Elite	NanoVision	46μs	—	21,739	PASS	QAI
Google Pixel 10	Snapdragon 8s Elite	Bearing Anomaly	46μs	—	21,739	PASS	QAI
Google Pixel 10 Pro XL	Snapdragon 8s Elite	Bearing Anomaly	46μs	—	21,739	PASS	QAI
Samsung Galaxy S25+	Snapdragon 8 Elite	Wake Word	47μs	—	21,276	PASS	QAI
Samsung Galaxy S25+	Snapdragon 8 Elite	NanoVision	47μs	—	21,276	PASS	QAI
Snapdragon X Plus 8-Core CRD	Snapdragon X Plus	NanoVision	47μs	—	21,276	PASS	QAI
Samsung Galaxy S25	Unknown	Wake Word	48μs	—	20,833	PASS	QAI
Snapdragon 8 Elite QRD	Snapdragon 8 Elite	TS Anomaly	48μs	—	20,833	PASS	QAI
Samsung Galaxy S25	Unknown	NanoVision	48μs	—	20,833	PASS	QAI
Samsung Galaxy S25+	Snapdragon 8 Elite	FinSense EN	48μs	—	20,833	PASS	QAI
Samsung Galaxy S24	Snapdragon 8 Gen 3	Wake Word	49μs	—	20,408	PASS	QAI
Samsung Galaxy S25 (Family)	Unknown	ECG Arrhythmia	49μs	—	20,408	PASS	QAI
Google Pixel 6 (Family)	Tensor G1	TS Anomaly	49μs	—	20,408	PASS	QAI
Google Pixel 7 (Family)	Tensor G2	TS Anomaly	49μs	—	20,408	PASS	QAI
Snapdragon X Plus 8-Core CRD	Snapdragon X Plus	TS Anomaly	49μs	—	20,408	PASS	QAI
Samsung Galaxy S25	Unknown	FinSense EN	49μs	—	20,408	PASS	QAI
Samsung Galaxy S25 (Family)	Unknown	Wake Word	50μs	—	20,000	PASS	QAI
Snapdragon 8 Elite Gen 5 QRD	Snapdragon 8 Elite	Wake Word	50μs	—	20,000	PASS	QAI
Samsung Galaxy S25	Unknown	TS Anomaly	50μs	—	20,000	PASS	QAI
Samsung Galaxy S25 Ultra	Snapdragon 8 Elite	NanoVision	50μs	—	20,000	PASS	QAI
Samsung Galaxy S25 Ultra	Snapdragon 8 Elite	Wake Word	51μs	—	19,607	PASS	QAI
Samsung Galaxy S25 Ultra	Snapdragon 8 Elite	ECG Arrhythmia	51μs	—	19,607	PASS	QAI
Google Pixel 10	Snapdragon 8s Elite	TS Anomaly	51μs	—	19,607	PASS	QAI
Snapdragon 7 Gen 4 QRD	Snapdragon 7 Gen 4	Bearing Anomaly	51μs	—	19,607	PASS	QAI
Samsung Galaxy S25+	Snapdragon 8 Elite	ECG Arrhythmia	52μs	—	19,230	PASS	QAI
Snapdragon 8 Elite Gen 5 QRD	Snapdragon 8 Elite	ECG Arrhythmia	52μs	—	19,230	PASS	QAI
Google Pixel 6	Tensor G1	TS Anomaly	52μs	—	19,230	PASS	QAI
Samsung Galaxy S25 Ultra	Snapdragon 8 Elite	TS Anomaly	52μs	—	19,230	PASS	QAI
Snapdragon X Plus 8-Core CRD	Snapdragon X Plus	FinSense EN	52μs	—	19,230	PASS	QAI
Samsung Galaxy S24 (Family)	Snapdragon 8 Gen 3	Wake Word	53μs	—	18,867	PASS	QAI
Samsung Galaxy S25	Unknown	ECG Arrhythmia	53μs	—	18,867	PASS	QAI
Google Pixel 7	Tensor G2	TS Anomaly	53μs	—	18,867	PASS	QAI
Samsung Galaxy S25 (Family)	Unknown	TS Anomaly	53μs	—	18,867	PASS	QAI
Snapdragon 8 Elite QRD	Snapdragon 8 Elite	NanoVision	53μs	—	18,867	PASS	QAI
Samsung Galaxy S25 (Family)	Unknown	FinSense EN	53μs	—	18,867	PASS	QAI
Samsung Galaxy S25 (Family)	Unknown	Speech Emotion	53μs	—	18,867	PASS	QAI
Samsung Galaxy S25+	Snapdragon 8 Elite	Speech Emotion	53μs	—	18,867	PASS	QAI
Google Pixel 8	Tensor G3	Bearing Anomaly	53μs	—	18,867	PASS	QAI
Snapdragon 8 Elite Gen 5 QRD	Snapdragon 8 Elite	TS Anomaly	54μs	—	18,518	PASS	QAI
Snapdragon 8 Elite Gen 5 QRD	Snapdragon 8 Elite	Speech Emotion	54μs	—	18,518	PASS	QAI
Snapdragon X2 Elite CRD	Snapdragon X2 Elite	HAR Activity	54μs	—	18,518	PASS	QAI
Samsung Galaxy S24+	Snapdragon 8 Gen 3	Wake Word	55μs	—	18,181	PASS	QAI
Google Pixel 10 Pro XL	Snapdragon 8s Elite	ECG Arrhythmia	55μs	—	18,181	PASS	QAI
Snapdragon 8 Elite Gen 5 QRD	Snapdragon 8 Elite	FinSense EN	55μs	—	18,181	PASS	QAI
Samsung Galaxy S25 Ultra	Snapdragon 8 Elite	Speech Emotion	55μs	—	18,181	PASS	QAI
Samsung Galaxy S25	Unknown	Audio Event	55μs	—	18,181	PASS	QAI
Google Pixel 10	Snapdragon 8s Elite	ECG Arrhythmia	56μs	—	17,857	PASS	QAI
Google Pixel 7 Pro	Tensor G2	ECG Arrhythmia	56μs	—	17,857	PASS	QAI
Snapdragon X Plus 8-Core CRD	Snapdragon X Plus	ECG Arrhythmia	56μs	—	17,857	PASS	QAI
Samsung Galaxy S25 Ultra	Snapdragon 8 Elite	FinSense EN	56μs	—	17,857	PASS	QAI
Samsung Galaxy S25	Unknown	Speech Emotion	56μs	—	17,857	PASS	QAI
Snapdragon X Plus 8-Core CRD	Snapdragon X Plus	Speech Emotion	56μs	—	17,857	PASS	QAI
Samsung Galaxy S25 (Family)	Unknown	Audio Event	56μs	—	17,857	PASS	QAI
Samsung Galaxy S24 Ultra	Snapdragon 8 Gen 3	Wake Word	57μs	—	17,543	PASS	QAI
Snapdragon 8 Elite QRD	Snapdragon 8 Elite	ECG Arrhythmia	57μs	—	17,543	PASS	QAI
Google Pixel 7 Pro	Tensor G2	TS Anomaly	57μs	—	17,543	PASS	QAI
Samsung Galaxy S24	Snapdragon 8 Gen 3	TS Anomaly	57μs	—	17,543	PASS	QAI
Samsung Galaxy S24 Ultra	Snapdragon 8 Gen 3	TS Anomaly	57μs	—	17,543	PASS	QAI
Google Pixel 8 (Family)	Tensor G3	Bearing Anomaly	57μs	—	17,543	PASS	QAI
Samsung Galaxy S24 (Family)	Snapdragon 8 Gen 3	ECG Arrhythmia	58μs	—	17,241	PASS	QAI
Samsung Galaxy S24 Ultra	Snapdragon 8 Gen 3	ECG Arrhythmia	58μs	—	17,241	PASS	QAI
Samsung Galaxy S24 Ultra	Snapdragon 8 Gen 3	NanoVision	58μs	—	17,241	PASS	QAI
Snapdragon X2 Elite CRD	Snapdragon X2 Elite	Fall Detection	58μs	—	17,241	PASS	QAI
Samsung Galaxy S24 (Family)	Snapdragon 8 Gen 3	TS Anomaly	59μs	—	16,949	PASS	QAI
Samsung Galaxy S24+	Snapdragon 8 Gen 3	TS Anomaly	59μs	—	16,949	PASS	QAI
Samsung Galaxy S24	Snapdragon 8 Gen 3	NanoVision	59μs	—	16,949	PASS	QAI
Samsung Galaxy S24 (Family)	Snapdragon 8 Gen 3	NanoVision	59μs	—	16,949	PASS	QAI
Google Pixel 8 Pro	Tensor G3	Bearing Anomaly	59μs	—	16,949	PASS	QAI
Samsung Galaxy S24+	Snapdragon 8 Gen 3	Audio Event	59μs	—	16,949	PASS	QAI
Google Pixel 7 (Family)	Tensor G2	ECG Arrhythmia	60μs	—	16,666	PASS	QAI
Samsung Galaxy S24+	Snapdragon 8 Gen 3	ECG Arrhythmia	60μs	—	16,666	PASS	QAI
Google Pixel 9	Tensor G4	TS Anomaly	60μs	—	16,666	PASS	QAI
Samsung Galaxy S24+	Snapdragon 8 Gen 3	Speech Emotion	60μs	—	16,666	PASS	QAI
Samsung Galaxy S25 (Family)	Unknown	HAR Activity	60μs	—	16,666	PASS	QAI
Samsung Galaxy S25+	Snapdragon 8 Elite	HAR Activity	60μs	—	16,666	PASS	QAI
Samsung Galaxy S24	Snapdragon 8 Gen 3	Audio Event	60μs	—	16,666	PASS	QAI
Samsung Galaxy S24 (Family)	Snapdragon 8 Gen 3	Speech Emotion	61μs	—	16,393	PASS	QAI
Samsung Galaxy S25 Ultra	Snapdragon 8 Elite	HAR Activity	61μs	—	16,393	PASS	QAI
Samsung Galaxy S23	Snapdragon 8 Gen 2	NanoVision	62μs	—	16,129	PASS	QAI
Samsung Galaxy S24+	Snapdragon 8 Gen 3	NanoVision	62μs	—	16,129	PASS	QAI
Samsung Galaxy S24+	Snapdragon 8 Gen 3	FinSense EN	62μs	—	16,129	PASS	QAI
Samsung Galaxy S24	Snapdragon 8 Gen 3	Speech Emotion	62μs	—	16,129	PASS	QAI
Snapdragon 8 Elite QRD	Snapdragon 8 Elite	Speech Emotion	62μs	—	16,129	PASS	QAI
Google Pixel 6 (Family)	Tensor G1	Bearing Anomaly	62μs	—	16,129	PASS	QAI
Samsung Galaxy S24 (Family)	Snapdragon 8 Gen 3	FinSense EN	63μs	—	15,873	PASS	QAI
Snapdragon 8 Elite QRD	Snapdragon 8 Elite	FinSense EN	63μs	—	15,873	PASS	QAI
Samsung Galaxy S21	Snapdragon 888	Bearing Anomaly	63μs	—	15,873	PASS	QAI
Samsung Galaxy S21 Ultra	Snapdragon 888	Bearing Anomaly	63μs	—	15,873	PASS	QAI

Showing 100 of 923 results. Use search to filter.

Methodology

Qualcomm AI Hub

ONNX models profiled on real devices via cloud-hosted NPU profiling across 72 phones, tablets, XR headsets, and automotive platforms. Median latency reported.

Edge Impulse

TFLite models deployed and profiled on 29 MCU/MPU boards from Cortex-M4F to Jetson Orin. Includes BrainChip neuromorphic, NXP NPU, and Synaptics edge accelerators.

Apple CoreML / ANE

ONNX → PyTorch trace → CoreML via coremltools. Benchmarked natively in Swift on M4 Max Neural Engine. 3-second sustained burst.

Google AI Edge LiteRT

ONNX → TFLite via onnx2tf. Benchmarked with XNNPACK delegate on M4 Max CPU. 3-second sustained burst.

Power Measurement

Real power via Apple IOReport (no sudo). CPU+GPU channels sampled during inference. Energy = (power × time) / count. Idle baseline subtracted.

Failures

We publish every failure. Galaxy Tab A8 (2021) fails on QAI (device farm issue). Memryx MX3 and Qualcomm RB3 Gen 2 DK fail on Edge Impulse (platform bugs, not model issues).

Keep exploring

There’s more.

flagship systems

Systems

AURE 14B, FinSense, LawSense, ConstantSense — the models behind the numbers.

Explore

published papers

Research

Papers, proofs, and negative results. Published with DOIs.

Explore

Start a Project

Discovery call, architecture review, deployment. We’ll tell you what fits.

Explore