From 1e127014ff8882c6a8d48d1ddcac740f18601a0b Mon Sep 17 00:00:00 2001 From: chuanqiw Date: Tue, 26 Mar 2019 10:57:59 +0800 Subject: [PATCH] Update INT8 calibration README test=release/1.3 --- python/paddle/fluid/contrib/int8_inference/README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/python/paddle/fluid/contrib/int8_inference/README.md b/python/paddle/fluid/contrib/int8_inference/README.md index 55a21ed1c5..3228610f96 100644 --- a/python/paddle/fluid/contrib/int8_inference/README.md +++ b/python/paddle/fluid/contrib/int8_inference/README.md @@ -65,7 +65,7 @@ Please note that [full ImageNet validation dataset](http://www.image-net.org/cha Notes: * The accuracy measurement requires the model with `label`. -* The INT8 theoretical speedup is 4X on Intel® Xeon® Cascadelake Server (please refer to `providing a theoretical peak compute gain of 4x int8 OPS over fp32 OPS` in [Reference](https://software.intel.com/en-us/articles/lower-numerical-precision-deep-learning-inference-and-training "Reference")). However, the actual test results at the model level will be less than 4X, and in general the average is about 2X. In addition, the calculation library optimization of batch size 1 is not as good as the large batch size. +* The INT8 theoretical speedup is 4X on Intel® Xeon® Cascadelake Server (please refer to `The theoretical peak compute gains are 4x int8 OPS over fp32 OPS.` in [Reference](https://software.intel.com/en-us/articles/lower-numerical-precision-deep-learning-inference-and-training "Reference")). Therefore, op-level gain is 4X and topology-level is smaller. ## 4. How to reproduce the results * Small dataset (Single core) -- GitLab