Google's LiteRT QNN Accelerator Achieves 100x CPU Speedup on Snapdragon Devices

Google’s New LiteRT Accelerator Supercharges AI Workloads on Snapdragon-powered Android Devices

Google introduced Qualcomm AI Engine Direct (QNN) to accelerate LiteRT models on Snapdragon 8 SoCs, achieving up to 100x faster execution than CPU-based processing. Benchmarks show 64 of 72 tested models achieved full NPU delegation on Snapdragon 8 Elite Gen 5 devices.

Why This Matters

Modern Android GPUs struggle with concurrent AI tasks like text-to-image generation and live camera segmentation, causing dropped frames and jitter. Neural Processing Units (NPUs), however, offer specialized acceleration with lower power consumption. QNN’s full model delegation and optimized kernels enable previously impossible real-time AI experiences, such as instant vision processing with <0.12s time-to-first-token.

Key Insights

• “100x CPU speedup, 10x GPU speedup across 64/72 models” (Google benchmarks, 2025)
• “Full model delegation” enables optimal NPU utilization for LLMs like Gemma
• Google optimized Apple’s FastVLM-0.5B with int8 weight and int16 activation quantization for QNN

Practical Applications

• Use Case: Real-time vision apps (e.g., live scene interpretation at 1024×1024 resolution)
• Pitfall: Over-reliance on GPU for complex AI pipelines risks performance degradation and thermal throttling

References:

• https://www.infoq.com/news/2025/11/litert-snapdragon-accelerator/