A Selfie and a Recording Generate a Digital Twin? The Power Consumption Barrier Behind This Is Higher Than Imagined

Zero. That's the frame rate Google's digital avatar solution achieves on mobile devices without a dedicated NPU. A selfie and a voice recording seem like simple inputs, but behind them lie multimodal alignment, real-time inference rendering, and speech synthesis powered by the Gemini Omni model. The combined computational demand of these three, estimated using a 7nm process node, requires at least 20 TOPS of integer compute power. Meanwhile, the NPU energy efficiency of current mainstream smartphone SoCs can barely provide 10 TOPS under a 5W power budget.