Skip to main content
. Author manuscript; available in PMC: 2018 Apr 6.
Published in final edited form as: MobiSys. 2017 Jun;2017:7–12. doi: 10.1145/3089801.3089805

Table 2.

Comparison of the time (in seconds) required to run the forward pass of Inception model using the original TensorFlow v1.0.1, TensorFlow with RenderScript matrix multiplication and convolution operations, and TensorFlow with RenderScript matrix multiplication only.

Nexus 6 Nexus 5X
Batch size Original TF TF + RS MatMul & RS Conv2D TF + RS MatMul Original TF TF + RS MatMul & RS Conv2D TF + RS MatMul
1 0.453 1.765 0.312 0.699 2.775 0.351
2 0.718 1.757 0.370 1.235 2.782 0.471
3 0.979 1.879 0.475 1.785 2.811 0.649
4 1.246 1.841 0.575 2.335 2.853 0.839
5 1.535 1.830 0.645 2.988 2.930 1.025