References

Tensorflow Official Benchmarks (May 2017, GitHub source): https://www.tensorflow.org/performance/benchmarks

IBM Power9 benchmark results (Nov 2017, 1.4.0): https://developer.ibm.com/linuxonpower/perfcol/perfcol-mldl/

Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour, Facebook (Jun 2017): https://research.fb.com/wp-content/uploads/2017/06/imagenet1kin1h5.pdf

POWER9 (hal000)

The following table is the result we get running with the same configuration as Tensorflow official benchmark settings.

This figure compares the result we get with Tensorflow official ones.

Green color stands for results get using floating point 16.

Red is the official Tensorflow result.

Blue stands for results get using floating point 32.

This figure shows the performance ratio of floating point 16 and 32 compare to Tensorflow official ones.

POWER8 (p8)