Unfortunately TF doesn't support runtime code dispatch according to CPU capacity. Hence recompiling TF with SIMD enabled is expected to bring performance boost.
The exception is when TF is built against MKL, MKL is able to dynamically select code according to the CPU.
Source, Data, Basic Environment
- Git Repo: https://salsa.debian.org/science-team/tensorflow
- Commit: b7252d13d59ebecae7928daa25af3a5ab1289bae
- Data: https://storage.googleapis.com/download.tensorflow.org/models/inception5h.zip
- OS: Debian Experimental (2018 Sept)
Benchmark Procedure
- compile the benchmark program with
sh debian/tests/tf_tool_benchmark_model.sh - download the pretrained model and unzip to the parent directory.
./tf_benchmark_model --graph=../tensorflow_inception_graph.pb
Results
| CPU | Compiler | Average Timing (Generic -O2) (µs) |
Average Timing (Opt march=native) (µs) |
Boost |
|---|---|---|---|---|
| I5-7440HQ (4C4T @ 2.8~3.8GHz, AVX2) | GCC (Debian 8.2.0-6) | 66953.4 ± 3017 | 24661.1 ± 1769 | 2.71x |
| I5-7440HQ (4C4T @ 2.8~3.8GHz, AVX2) | Clang 7.0.0-+rc2-1~exp3 (tags/RELEASE_700/rc2) | 49932.8 ± 2550 | 26324.7 ± 1743 | 1.90x |
lumin AT debian.org, CC-BY-SA 4.0