MobileNetV2为何仅在移动设备上比MobileNetV1运行更快？

Great question! Let's break down the key reasons behind this phenomenon, referencing the statement you came across:

例如，MobileNet V2在移动设备上比MobileNet V1更快，但在桌面GPU上略慢。

• Core Optimization for Mobile Hardware
  MobileNetV2 was purpose-built to maximize efficiency on mobile CPUs, which have tight constraints on computational power and memory bandwidth. Its signature inverted residual blocks with linear bottlenecks are tailor-made for this environment:

  • The inverted residual structure rearranges the order of pointwise and depthwise convolutions (compared to V1), cutting down the total number of operations needed to transform features. For mobile CPUs that struggle with heavy tensor computations, this reduces redundant work drastically.
  • Linear bottlenecks avoid the information loss caused by ReLU activation on low-dimensional feature maps—critical when every computation cycle is precious on mobile, as we don't waste resources reconstructing damaged features.

• Misalignment with Desktop GPU Strengths
  Desktop GPUs excel at large-scale parallelism and batch processing of big tensor operations, and MobileNetV2's design works against this strength:

  • The inverted residual blocks rely on many small, fragmented operations (like tiny pointwise convolutions and depthwise convolutions). GPUs aren't optimized for these tiny tasks—they can't fully leverage their thousands of CUDA cores when splitting work into such small chunks, leading to hardware underutilization.
  • MobileNetV1 has a more "uniform" structure with larger, contiguous operations that fit better with how GPUs are engineered. The frequent memory access and operation switching required by V2's blocks add extra overhead, negating potential gains and making it slightly slower than V1 on desktop GPUs.

内容的提问来源于stack exchange，提问作者Seongkyun Han