headroom

A two-second, on-device probe that tells your Flutter app how fast an LLM of a given size will decode on this phone, right now, and whether it will fit, without downloading a model. iOS 17+ and Android 10+ (API 29).

It answers three questions an app cannot ask the OS:

• How fast? A single-stream decode estimate in tokens per second, as an interval, for any model size.
• For how long? A sustained estimate after the phone has throttled.
• Will it fit? A memory verdict against the memory the OS reports.

Every figure says where it came from: measured on this device just now, calibrated (n=…) from committed measurements of n other devices, or unknown. The types make it impossible to read an estimate as a measurement.

This is the Flutter plugin for the Headroom Swift package; the probe is the package's own C and Swift code, and the estimator is a Dart port of its maths.

Usage

import 'package:headroom/headroom.dart';

final report = await Headroom.probe();            // one to two seconds, off the UI thread
report.ceilingGBps;                               // Quantity(54.7, basis: measured) on iOS; null on Android

final estimate = report.estimate(ModelSpec.tinyLlama1_1BQ4_0, contextTokens: 1024);
estimate.peak;        // Interval [55.7, 76.7] tok/s, basis calibrated (n=2) on iOS; unknown on Android
estimate.sustained;   // Interval? — calibrated (n=2) on an iPhone; null elsewhere
estimate.fit;         // fits (2810 MB spare), basis measured

report.toJsonString();                            // the whole report, for pasting back

A model you only know by file size works too: ModelSpec.fromGgufBytes(name: '2.2 GB model', ggufBytes: 2200000000). The estimate then notes that the KV cache is not modelled.

Headroom.cancel() stops a running probe at its next stage boundary. Flutter's animation library also exports a type named Interval, so a file that uses both can import 'package:flutter/material.dart' hide Interval;.

What each platform measures

	iOS	Android
GPU ceiling	STREAM copy/scale/add/triad through Metal, 3 × 128 MB, GPU-timestamped, output verified; the triad median is the ceiling	none in this version: gpu is null and a warning says so
CPU triad	the C core's pthread STREAM triad, 3 × 64 MB, NEON, output verified; the performance cluster and all cores are tried and the better median kept, both recorded in attempts	the same C core, one all-core attempt (Android has no API for the performance cluster), recorded in attempts
Available memory	os_proc_available_memory, the process's budget, basis measured	ActivityManager.MemoryInfo.availMem, basis measured, plus lowMemory and threshold
Thermal, power	ProcessInfo.thermalState, Low Power Mode, UIDevice battery	PowerManager.currentThermalStatus folded as below, Battery Saver, BatteryManager
Device	uname identifier, iOS version and build	Build.MANUFACTURER MODEL, Build.SOC_MANUFACTURER SOC_MODEL (API 31+, else unknown), Build.VERSION.RELEASE, Build.ID
Predictions	peak and sustained intervals, basis calibrated (n=2)	unknown, with the note "no Android calibration yet"

A probe never allocates past what the OS reports as available: if the three arrays would not fit, that probe is skipped and the warning says so. In the iOS Simulator and the Android Emulator a warning says the figures are the host machine's; simulator memory is reported with basis unknown.

Android's seven thermal levels fold into the report's four; the original name is kept in conditions.platformThermalStatus:

PowerManager status	ThermalState
NONE	nominal
LIGHT, MODERATE	fair
SEVERE	serious
CRITICAL, EMERGENCY, SHUTDOWN	critical

The maths

Single-stream decode is memory-bound: each token reads every weight once plus the KV cache for every token already in context.

bytesPerToken  = tensorBytes + 2 × layers × kvHeads × headDim × contextTokens × 2   (f16 KV, llama.cpp's default)
η              = achievedGBps / measuredCeilingGBps                                 (per calibration device)
peak tok/s     = measuredCeilingGBps × η / bytesPerToken      as an interval over the ceiling's 95% CI and min…max η
sustained      = peak × sustainedFactor                        as an interval over the measured phones
required bytes = tensorBytes × 1.1 + kvBytes(context) + 150 MB    (stated assumptions; a margin under 10% is "tight")

η and the sustained factor are computed at run time from assets/calibration.json, a copy of the Swift package's file. Nothing is typed in as a constant; the package's Calibration/README.md does the arithmetic row by row.

SoC	Decode (tok/s)	Achieved GB/s	Measured ceiling GB/s	η	Source
A15 Bionic (iPhone 13)	42.8012	27.221	pending (null)	—	PocketRoofline SISO, warm start, unplugged
Apple M1 (MacBook Pro 2020)	61.1016	38.860	56.000 (median of 3 runs)	0.6939	PocketRoofline SISO; ceiling from headroom-probe on the same machine
A16 Bionic (iPhone 15 Plus)	65.1648	41.444	45.317 (median of 4 probes)	0.9145	PocketRoofline app SISO, warm start, live UI on screen, battery, airplane mode; ceiling from the Headroom demo on the same phone

η is the range 0.6939–0.9145, basis calibrated (n=2): decode used 69% of the measured ceiling on the M1 and 91% on the A16, and the estimator keeps the whole range, so every upper bound is 1.32 × what the M1 alone gave. The A16 joined after the Swift package's first out-of-sample test: a prediction for that phone, made from the M1 alone, missed by 32%. The package's README has the result.

Sustained factor	Peak → sustained	Factor	Source
A15 Bionic, llama.cpp-Metal	42.80 → 30.93 after 5 × 1024-token generations	0.7227	PocketRoofline SILO
A18 Pro (iPhone 16 Pro), MLX	40.49 → 23.67 over 20 runs	0.5846	arXiv:2603.23640
A16 Bionic (iPhone 15 Plus), llama.cpp-Metal	65.16 → 27.16 on the 5th 1024-token generation, run straight after LISO	0.4169	PocketRoofline app SILO; the step into it is unexplained

The sustained range is 0.4169–0.7227, basis calibrated (n=3). The A16 factor comes from a halving of decode speed, not yet explained, at the boundary between two benchmark regimes; it is included because it widens the range.

Every calibration device is an Apple SoC with a Metal-measured ceiling, which is why Android gets no η: a prediction there would be a borrowed constant, and the plugin says unknown instead.

Example

example/ has one Probe button, a report card (ceilings with their 95% intervals and basis tags, thermal state, power, memory), an estimate table for four model sizes at 1024 tokens of context, and Copy JSON for pasting a phone's result back into the calibration. example/integration_test/ runs the real probe on whatever device it is pointed at and checks that every figure recomputed in Dart from the raw iteration times equals what the native side reported.

The native side

tool/sync_native.sh is the single source of truth for every native file: it copies the C core into ios/Classes/CHeadroom/ and android/src/main/cpp/CHeadroom/, the Swift probe layer into ios/Classes/Core/, and calibration.json into assets/. Nothing in those places is edited by hand. When the Swift package changes, run it again and commit the result; tool/sync_native.sh --check reports drift.

The estimator layer of the Swift package (Estimate, MemoryFit, Calibration, ModelSpec) is not copied: it is ported to Dart under lib/src/, and the unit tests reproduce the package's test fixtures number for number. The bootstrap interval uses the same seeded SplitMix64 stream and the same index derivation as Swift, and a test proves it by recomputing the committed M1 run's intervals bit for bit.

• iOS is a CocoaPods pod (ios/headroom.podspec): Swift 6 language mode, the C files at -O3 in every configuration, and a module map so the core's import CHeadroom resolves unchanged. The native side exposes only the probe, on the headroom/probe method channel.
• Android builds the same C core with CMake (-O3, NEON on arm64) and a small JNI wrapper; the Kotlin side reads memory, thermal, battery and device facts and writes the same report JSON, with a Kotlin port of the statistics.

Because the C kernel is optimised in every configuration here, a debug flutter run still measures a CPU ceiling; the Swift package marks debug CPU figures unknown because a package cannot carry per-target flags.

Limitations

• GPU path only, iOS only. The ceiling and the estimate describe a Metal backend such as llama.cpp-Metal or MLX. The CPU triad is a ceiling for CPU backends and is not used in the estimate.
• No Android calibration. Android reports a CPU ceiling, memory and conditions, but no throughput prediction: there is no GPU probe and no Android row in the calibration.
• Calibrated on n = 2 devices, an M1 Mac and an A16 iPhone, whose η differ by a factor of 1.32; every interval carries that spread. The Swift package's one out-of-sample test so far, on the A16 before it joined the calibration, missed by 32%. The A15 row is pending the iPhone 13 run.
• The thermal factor is from three phones on two runtimes, presented as a range, not a curve for any particular device. Its lowest value, the A16's, is not yet explained.
• Model geometry must be verified. A spec built from file size alone omits the KV cache; the estimate says so. Only two presets ship, both read from GGUF headers with their file hashes in ModelSpec.
• Available memory on Android is system-wide (availMem), not a per-process budget like iOS's.
• Simulator and emulator figures are the host's, and the report says so. In the iOS Simulator the GPU probe is skipped: the simulator's command-buffer timestamps do not bracket the GPU work (128 MB dispatches report tens of microseconds, thousands of GB/s), so there is no honest figure to report. Probe a device.

The Swift package's README lists what was searched for and not found before this was built.

Status

• Unit tests (flutter test): 69, including the Swift fixture numbers, the A16 rows recomputed from their raw records, and the bit-for-bit statistics check against the committed M1 run and the four iPhone 15 Plus probes taken with the Swift package's demo app.
• End to end: run on the iOS Simulator and the Android Emulator through example/integration_test/probe_test.dart; those figures are the host machine's. No physical phone has been probed through the plugin yet.

License

MIT. Built on the measurements in PocketRoofline.