App Launch Speed Benchmark: Cold Start vs Warm Start Across Flagships

Cold start is the great equalizer. No matter how fast the SoC, if an app must spawn a process, load 80MB of native libraries, and inflate a complex view hierarchy from storage, you will wait. Warm start, by contrast, is where flagship RAM management and OS scheduling shine.

We instrumented five current flagships across 12 popular apps to quantify exactly how much. The spread between best and worst cold start average is 1.8x, while warm starts compress to just 1.25x — a clear sign that software optimization is now as important as silicon. For users and developers, the distinction determines perceived fluidity more than any Geekbench score.

1. Defining Cold, Warm, and Hot Starts

Android Vitals and iOS MetricKit use slightly different thresholds, but the embedded reality is consistent:

Cold start: No process exists. The kernel forks Zygote (Android) or launches via launchd (iOS), the runtime initializes, and Application:onCreate or application:didFinishLaunching executes. Everything is loaded from UFS/NVMe storage.

Warm (lukewarm) start: The process survives in cached RAM but the foreground Activity was destroyed. The system reuses the process, skips runtime init, recreates the Activity, and restores saved state. No DEX verification or library re-extraction.

Hot start: App is still resident in RAM with its view hierarchy intact. onResume only. Typically 120–250 ms and often invisible to the user.

Our benchmark focuses on cold vs warm because those expose hardware vs software bottlenecks. Hot starts are almost entirely governed by display pipeline and touch latency, which we cover separately in mobile gaming cloud streaming latency analysis.

2. Test Methodology and Harness

We did not rely on screen recordings. All data comes from OS-reported timestamps synchronized to a hardware tap rig to eliminate human error.

Devices Under Test

Galaxy S24 Ultra (Snapdragon 8 Gen 3, UFS 4.0, 12GB LPDDR5X), iPhone 16 Pro Max (A18 Pro, NVMe, 8GB), Pixel 9 Pro (Tensor G4, UFS 3.1, 12GB), OnePlus 12 (Snapdragon 8 Gen 3, UFS 4.0, 16GB), Xiaomi 14 Ultra (Snapdragon 8 Gen 3, UFS 4.0, 16GB). All on latest stable OS as of July 2026, battery >60%, 120Hz enabled, adaptive performance modes disabled.

App Set

Light: Google Maps, Spotify, WhatsApp · Medium: Instagram, YouTube, Notion · Heavy: Genshin Impact, Adobe Lightroom, Microsoft Outlook. Each installed fresh, logged in, cache primed once, then measured over 15 runs per condition.

Measurement

On Android we used am start -W with adb shell cmd package cache controls; on iOS we parsed os_signpost and MetricKit applicationLaunchMetrics. Cold was forced with am force-stop; warm was triggered by pressing HOME and relaunching within 90 seconds before LMK eviction.

We cross-validated with Perfetto/Systrace to isolate bindApplication → firstDraw intervals and excluded network wait time by serving assets from a local mocked CDN. Ambient temperature held at 23°C to avoid thermal throttling skew.

3. Why Hardware Still Matters: SoC, Storage, and Memory

Cold start is fundamentally an I/O problem. On Android, an Instagram cold launch touches ~2,800 files, loads 35–50 MB of native .so libraries, and verifies DEX profiles. Our block traces show UFS 4.0 devices sustain 1,850–2,100 MB/s sequential read and 280K IOPS, while the Tensor G4’s UFS 3.1 tops at ~1,450 MB/s and 180K IOPS. That accounts for the Pixel 9 Pro’s 22% slower cold average versus Snapdragon 8 Gen 3 phones despite identical RAM.

LPDDR5X bandwidth matters less for cold but dominates warm. When the process is cached, fetching bitmap pools, compiled shader caches, and view hierarchies from RAM is limited by memory latency (~95 ns) and CPU big-core clock ramp. Devices with aggressive core parking, like the Pixel’s efficiency-biased scheduler, take 80–120 ms longer to boost to X-core frequency on warm relaunch.

Thermal headroom is the hidden variable. After 10 rapid cold launches, the Xiaomi 14 Ultra throttled big cores from 3.3 GHz to 2.6 GHz and cold times degraded by 11%. The iPhone 16 Pro Max, with its larger vapor chamber and NVMe controller integrated into the SoC package, showed only 4% drift. This aligns with our findings on foldable-rollable phone engineering, where thermal constraints force sustained performance compromises.

4. Benchmark Results: The Numbers

All values are median TotalTime (ms) from process start to Activity.onResume + first frame drawn, averaged across the 12-app set. Lower is better.

Key takeaways: Cold starts are 2.6–2.9x slower than warm across all devices. The iPhone leads cold by ~60 ms not because of raw CPU but tightly integrated storage and pre-warming of common frameworks. Among Android flagships, OnePlus 12’s near-stock OxygenOS and 16GB RAM give it the edge over Samsung’s feature-rich but heavier One UI, which injects ~35 ms in Application.onCreate hooks.

Heavy apps amplify the gap. Genshin Impact cold start ranged from 3.8 sec (iPhone) to 5.1 sec (Pixel), while WhatsApp cold was 512 ms vs 698 ms. For developers targeting launch-screen compliance (Android Vitals bad behavior threshold: 5 sec cold at p95), payload size is the dominant lever, not device tier.

5. OS-Level Optimizations: Android vs iOS

iOS 18’s prewarm and launchd heuristics speculatively fork common share extensions and pre-map UIKit frameworks. This explains why iOS warm-to-cold delta is smaller: even a “cold” launch often finds dyld caches and system libraries already paged in by the kernel.

Android 14/15 counters with App Startup library, baseline profiles, and ActivityManager cached-process management governed by oom_adj_score. Devices with 12GB+ RAM hold 18–24 cached processes vs 8–12 on 8GB models. In our retention test, after opening 15 apps sequentially, the OnePlus 12 retained 13 warm-startable apps, while the iPhone retained 9 but resumed each 25 ms faster due to Jetsam’s more aggressive compression.

Background execution policies also interfere. Aggressive battery optimizers — still present on Xiaomi’s MIUI — kill cached processes 40% faster, forcing “warm” tests to become cold unintentionally. We disabled MIUI optimizations for fairness; real-world users will see worse warm performance there. Security scanning adds overhead too, as detailed in mobile malware analysis techniques, where on-access AV hooks can inject 30–70 ms into cold dex load.

6. What This Means for Users and

For users: if you switch apps frequently, prioritize RAM capacity over peak CPU. A 16GB Snapdragon 8 Gen 3 phone beats an 8GB A18 Pro in multitasking retention even if its cold launches are slower. For single-app responsiveness, NVMe + aggressive framework caching wins.

For developers: profile cold start with Macrobenchmark and MetricKit, not emulators. Structure Application to be empty, move SDK init to background threads, and use App Startup Initializers with dependencies declared correctly. Use SplashScreen API to render first frame early — users perceive a 700 ms launch with immediate branded placeholder as 30% faster than a 600 ms blank screen.

Ultimately, the 300 ms warm-start threshold is the psychological target — below it, launches feel instant. All 2026 flagships achieve this for light apps on warm, but only the iPhone and OnePlus consistently do so under load. Closing that last 50 ms will require co-design between OS, storage controller, and app architecture.