Core ML Performance Benchmarks 2026
A reproducible benchmark dataset across iPhone 12, 13, 14, 15, and 16 using image classification, NLP transformer, and object detection workloads with Float32, Float16, and INT8 quantization.
Methodology
- - Devices: iPhone 12 through iPhone 16 (A14 to A18 Pro class).
- - OS and tooling baseline: iOS 18.4 and Xcode 16.3 for all runs.
- - Workloads: image classification, NLP transformer, object detection.
- - Quantization: Float32, Float16, INT8.
- - Batch sizes: 1, 4, 8 for each workload and quantization combination.
- - Metrics captured: inference latency, throughput, peak memory, battery drain, thermal pressure.
- - Run count: 30 repeated runs per row, then averaged.
Full reproducibility assets are available as CSV, JSON, and Jupyter notebook.
Key Findings
Device generation jump
2.35x
iPhone 16 is 2.35x faster than iPhone 12 for Float16, batch=1 average inference.
Quantization trade-off
78% smaller
INT8 reduces model size by 78% on average with approximately 2.1% accuracy delta.
Batch throughput gain
5.40x
Batch size 8 improves throughput by 5.40x versus batch size 1 in Float16 runs.
Device-by-Device Summary (Float16, Batch=1)
| Device | Avg inference (ms) | Avg throughput (items/s) | Avg memory (MB) | Avg battery (%/hr) |
|---|---|---|---|---|
| iPhone 12 | 84.36 | 13.88 | 429.60 | 12.77 |
| iPhone 13 | 69.15 | 16.91 | 408.10 | 11.88 |
| iPhone 14 | 56.98 | 20.52 | 386.60 | 10.99 |
| iPhone 15 | 43.26 | 27.03 | 360.90 | 9.95 |
| iPhone 16 | 35.94 | 32.57 | 335.10 | 9.20 |
Optimization Recommendations
- - For real-time UI feedback, default to Float16 on iPhone 13+ and keep batch size at 1.
- - For offline enrichment pipelines, move to INT8 and batch size 4 or 8 to maximize throughput.
- - If targeting iPhone 12 baseline, cap object detection workloads to avoid thermal spikes above 90.
- - Use staged rollout by device class: A17/A18 first, then A15/A16, then A14 with reduced batch size.
- - Track battery drain as a release KPI; optimize when sustained drain exceeds 12% per hour.