diff --git a/benchmark/IntelOptimizedPaddle.md b/benchmark/IntelOptimizedPaddle.md
index 8ee7fd28c58f2a2bcb82040eb824a37062bd4e9c..94a79c3c8ea3d5a8c50ba2b7782abec21e2cff13 100644
--- a/benchmark/IntelOptimizedPaddle.md
+++ b/benchmark/IntelOptimizedPaddle.md
@@ -22,6 +22,7 @@ On each machine, we will test and compare the performance of training on single
 
 #### Training
 Test on batch size 64, 128, 256 on Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz
+Pay attetion that the speed below includes forward, backward and parameter update time. So we can not directly compare the data with the benchmark of caffe `time` [command](https://github.com/PaddlePaddle/Paddle/blob/develop/benchmark/caffe/image/run.sh#L9), which only contain forward and backward. The updating time of parameter would become very heavy when the weight size are large, especially on alexnet.
 
 Input image size - 3 * 224 * 224, Time: images/second
 
@@ -55,6 +56,16 @@ Input image size - 3 * 224 * 224, Time: images/second
 
 <img src="figs/googlenet-cpu-train.png" width="500">
 
+- Alexnet
+
+| BatchSize    | 64     | 128    | 256    |
+|--------------|--------| ------ | -------|
+| OpenBLAS     | 0.85   | 1.03   | 1.17   | 
+| MKLML        | 71.26  | 106.94 | 155.18 |
+| MKL-DNN      | 362.66 | 497.66 | 610.73 | 
+
+chart TBD
+
 #### Inference
 Test on batch size 1, 2, 4, 8, 16 on Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz
 - VGG-19