A mobile GPU is the graphics processing unit integrated into a smartphone or tablet system-on-chip, responsible for drawing the user interface, games, video effects and camera previews, and increasingly for general parallel computation. In mobile hardware, it shares memory, power and cooling with the CPU, NPU and modem on the same chip.
Unlike a desktop graphics card, a mobile GPU works within a power budget of only a few watts and no active cooling. Its design therefore prioritizes energy per frame and memory bandwidth savings over raw peak throughput.
How a mobile GPU renders a frame
Most mobile GPUs use tile-based rendering. The screen is divided into small tiles, often 16×16 or 32×32 pixels, and the GPU first sorts all triangles into the tiles they touch. It then renders each tile completely in fast on-chip memory and writes the finished tile to system memory only once.
This approach exists because external memory access is the most expensive operation in a phone, in both energy and time. Desktop GPUs traditionally render in immediate mode, reading and writing a large framebuffer in dedicated high-bandwidth memory, which a battery-powered device cannot afford.
The following table compares the two rendering approaches.
| Property | Tile-based (mobile) | Immediate mode (desktop) |
|---|---|---|
| Framebuffer access | On-chip tile memory | External video memory |
| Memory bandwidth need | Low | High |
| Overdraw handling | Hidden surfaces often removed per tile | Relies on depth testing and ordering |
| Cost of extra render passes | High if tiles are flushed to memory | Moderate |
| Power draw | A few watts | Tens to hundreds of watts |
Main mobile GPU families and graphics APIs
Mobile GPUs come from a small number of architecture families, each tied to particular chip vendors. Their capabilities are exposed to apps through standardized or platform-specific graphics APIs.
| Family | Found in | Notes |
|---|---|---|
| Adreno | Snapdragon SoCs | Widely used in Android flagships |
| Mali and Immortalis | Many Arm-based SoCs | Licensed to multiple chip makers |
| Apple GPU | A-series and M-series chips | Custom tile-based design, Metal only |
| PowerVR | Selected mid-range and embedded SoCs | Pioneer of tile-based deferred rendering |
| Xclipse | Some Exynos SoCs | Based on a desktop-derived architecture |
On Android, apps use Vulkan or the older OpenGL ES; on Apple devices, they use Metal. Vulkan and Metal are low-overhead APIs that give the app explicit control over render passes, memory and synchronization, which matters on tile-based hardware because a careless render pass can force tiles to be written out and read back.
Mobile GPU compared with neighboring processors
Within a mobile SoC, the GPU is one of several specialized processors. Choosing the right one for a workload has a large effect on speed and battery life.
| Processor | Best at | Example workloads |
|---|---|---|
| CPU | Branching, sequential logic | App logic, parsing, layout |
| Mobile GPU | Massively parallel floating-point work | UI composition, games, image filters |
| Mobile NPU | Low-precision neural network math | On-device AI, photo segmentation |
| ISP | Fixed camera pipeline stages | Demosaicing, noise reduction, HDR merge |
| Video codec block | Encoding and decoding streams | Video playback and recording |
Machine learning inference can run on the GPU, but a dedicated NPU usually delivers better performance per watt for supported models. The GPU remains a flexible fallback when a model uses operators the NPU does not support.
Thermal limits and sustained performance
A mobile GPU can run at its peak clock speed only briefly. As the chip heats up, the operating system and firmware apply thermal throttling, lowering clocks to keep the device’s surface temperature comfortable and protect the battery.
For this reason, sustained performance after 10 to 20 minutes of gaming is often far lower than the peak figures in short benchmarks. The size of the gap depends on the phone’s cooling design, its case, the ambient temperature and whether the device is charging at the same time.
- Frame rate caps of 30 or 60 frames per second keep heat down in long sessions.
- Dynamic resolution lowers rendering resolution when the frame time rises.
- Variable rate shading reduces work in areas where detail is less visible.
- Upscaling renders fewer pixels and reconstructs the final image.
Common mobile GPU optimization mistakes
Many performance problems on phones come from techniques that work well on desktops but conflict with tile-based architecture. The table lists frequent ones.
| Mistake | Effect | Better approach |
|---|---|---|
| Unnecessary full-screen passes | Tiles flushed to memory repeatedly | Merge passes and use subpasses |
| Heavy transparent overdraw | Each pixel shaded many times | Reduce layered translucent elements |
| Reading back GPU results each frame | Pipeline stalls | Use asynchronous readback |
| Testing only on flagships | Poor results on mid-range devices | Profile on representative hardware |
| Ignoring thermal behavior | Frame rate drops after minutes | Test long sessions and cap frame rate |
Vendor profiling tools reveal memory bandwidth, tile flushes and shader cost, which are more useful indicators on mobile than a single frames-per-second counter.
A non-technical view
A mobile GPU is the part of a phone’s chip that draws everything on the screen, from scrolling menus to 3D games. It handles many small drawing jobs at once. That makes it fast at pictures and animations.
It works like an artist who paints a mural one small section at a time at a nearby table. The artist does not walk to a distant wall for every stroke. This saves energy. That matters because the phone runs on a battery and has no fan.
A concrete example
A small studio ports a 3D racing game from consoles to phones. On a flagship device the first build runs at 60 frames per second, but after fifteen minutes the frame rate falls to about 35 and the phone becomes warm, and on mid-range devices the game stutters from the start.
Profiling shows that the post-processing chain uses five separate full-screen passes, each forcing the GPU to write tiles out to memory and read them back. The team merges them into one render pass with subpasses, reduces translucent smoke particles, adds dynamic resolution, and offers a 30 frames per second battery mode.
The revised build holds a steady 60 frames per second for a full race series on the flagship and a stable 30 on mid-range phones, while battery drain per hour of play drops noticeably.