tensorflow
Low Performance on i9-9900K processor
Hi I'm running TensorFlow benchmark on a single Motherboard Pro WS C246-ACE(Processor: i9-i9900K, DRAM: 64GB), just trying out the performance without using GPUs. But the performance is quick low. 2.5~3.5 image/sec. Even after 20000 epoch of training, the max top-1/top-5 accuracy is about 0.031/0.094. Learning_rate is increasing at a same speed, from 0 to 1.7e-3(still increasing while training). I wonder is this the expected performance or I put the wrong parameters.
tf.version=1.15.0-dev20190821 No MKL is used.
Below is my script: python tf_cnn_benchmarks.py --num_gpus=1 --batch_size=32 --model=resnet50_v1.5 --variable_update=parameter_server --device=cpu --data_format=NHWC --data_name=imagenet --data_dir=F:\ILSVRC2017_CLS-LOC --save_summaries_steps=10 --train_dir=C:\Users\C246_0710\Anaconda3\envs\tensorflow1\benchmarks-master\scripts\tf_cnn_benchmarks\training --summary_verbosity=1 --print_training_accuracy --benchmark_log_dir=C:\Users\C246_0710\Anaconda3\envs\tensorflow1\benchmarks-master\scripts\tf_cnn_benchmarks\training --save_model_steps=1000
Running the same command and version of TF (1.15.0) I get this with a Ryzen 3700X:
Running warm up
Done warm up
Step Img/sec total_loss
1 images/sec: 9.2 +/- 0.0 (jitter = 0.0) 7.791
10 images/sec: 9.3 +/- 0.0 (jitter = 0.1) 7.952
Unless there is something wrong with the data your CPU seems to be underperforming.
Did you try on Linux distro such as Ubuntu? From experience I found performance is much better on Ubuntu (18.04) than Windows 10 for these type of tasks
