Packaging: APK, AAB and the size budget that actually matters
Split by ABI, split by density, drop the formats nobody uses, and stop compressing what you already compressed. A worked example, from 410 MB to 197 MB.
Read the articleIndependent · since March 2025
Native Android games, measured
Notes for people who write Android games in C++ and Rust: what the NDK changed this month, where a frame spends its sixteen milliseconds, and which numbers survive a hot phone.
Latest news
All 16 items-
Android 17 preview raises the memory limits for games
The per-app limit on low-memory devices moves from 512 MB to 768 MB. The OOM killer still ignores your intent, and that is the part that matters.
-
NDK r29 preview: what changes for game builds
The preview drops the legacy toolchain files, moves to LLVM 19 and finally makes the C++ standard library choice explicit.
-
The average frame is fine and the game still stutters
Report the 95th and 99th percentile, not the mean. A 16.6 ms average with a 40 ms tail is a game that hitches twice a second.
-
The asset half of APK size
Compression settings saved 12 MB. Wrong texture formats saved 180 MB. The order matters.
-
Gamepad to photon: measuring input latency end to end
A high-speed camera and a Bluetooth controller with a known LED give a number you can act on, without special hardware.
-
The memory type you pick is a promise about the next frame
DEVICE_LOCAL is not automatically faster, and HOST_VISIBLE is not automatically slow. The difference shows up two frames later.
Articles
All six-
Memory on mobile: budgets, allocators and the OOM killer
A 4 GB phone does not give your game 4 GB. Where the numbers come from, how to measure a real budget, and why the low-memory killer is not the thing that kills you.
-
Power and thermal budget: designing for a phone that gets hot
Sustained performance is a battery and heat problem before it is a GPU problem. How to find your device's real budget and spend it deliberately.
-
Profiling one frame: where 16.6 ms actually goes
A frame budget broken into CPU, GPU and presentation, measured with Perfetto and a Vulkan timestamp query, on a phone that costs less than a monitor.
For beginners
If you have written Java or Kotlin apps and want to move into native game code, read these four in order. They assume C++ but not Vulkan, and every number in them can be reproduced on a mid-range phone.
- Get a first Vulkan frame on screenCMake, NativeActivity, swapchain, and the four errors everyone meets.
- Make the audio not feel lateAAudio and Oboe, buffer sizing, and what the last 20 ms are made of.
- Measure a frame properlyPerfetto, timestamps, percentiles, and four false leads to ignore.
- Stay inside the memory budgetWhat a 4 GB phone gives you, and how to enforce it in the build.
Digests
All digestsOctober digest: heat, memory and the packaging bill
Power, memory and download size, and why all three decide the second ten minutes.
July digest: input latency, asset sizes and a frame under the microscope
Two measurements and two articles on the same question: what is actually slow?
May digest: page sizes, baseline profiles and the first Vulkan frame
Spring's three news items and the walkthrough that followed them.
Follow without an account
The RSS feed carries every news item. There is no newsletter, no email address and no tracking pixel.