The GPU Backend Problem: Why Gemma 4 Performance on Android Varies 10x

Running Gemma 4 on Android faces a critical silent failure: when the GPU backend (OpenCL) isn't available on a device—common on budget hardware—the engine defaults to CPU without warning, delivering just 2-5 tokens/second instead of GPU's 52 tokens/second. This 10x performance gap means developers optimizing blind to their actual backend can waste effort on the wrong bottleneck. Properly detecting and initializing the backend is essential for shipping responsive on-device AI applications.