bartekkuncer commented on a change in pull request #20856:
URL: https://github.com/apache/incubator-mxnet/pull/20856#discussion_r800441603



##########
File path: 
docs/python_docs/python/tutorials/performance/backend/dnnl/dnnl_quantization.md
##########
@@ -0,0 +1,304 @@
+<!--- Licensed to the Apache Software Foundation (ASF) under one -->
+<!--- or more contributor license agreements.  See the NOTICE file -->
+<!--- distributed with this work for additional information -->
+<!--- regarding copyright ownership.  The ASF licenses this file -->
+<!--- to you under the Apache License, Version 2.0 (the -->
+<!--- "License"); you may not use this file except in compliance -->
+<!--- with the License.  You may obtain a copy of the License at -->
+
+<!---   http://www.apache.org/licenses/LICENSE-2.0 -->
+
+<!--- Unless required by applicable law or agreed to in writing, -->
+<!--- software distributed under the License is distributed on an -->
+<!--- "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY -->
+<!--- KIND, either express or implied.  See the License for the -->
+<!--- specific language governing permissions and limitations -->
+<!--- under the License. -->
+
+## Introduction
+
+After successful model building and achieving desired accuracy on the test 
data, often the next step is to optimize inference to deploy the model to 
production. One of the key features of usable model is to have as small latency 
as possible to be able to provide services to large number of customers at the 
same time. In addition to customer satisfaction, with well optimized model, 
hardware load is reduced which also reduces energy costs needed to perform 
inference.
+
+Two main types of software optimizations can be characerized as:
+- memory-bound optimizations - main objective of these optimizations is to 
reduce memory operations (reads and writes) - it is done by e.g. chaining 
operations which can be performed one after another immediately, where input of 
every subsequent operation is the output of the previous one (example: ReLU 
activation after convolution)
+- compute-bound optimizations - these optimizations are mainly done on 
operations which require large number of CPU cycles to complete, like 
FullyConnected and Convolution - one of the methods to speedup compute-bound 
operations is to lower computation precision - this type of optimization is 
called quantization

Review comment:
       ```suggestion
   - compute-bound optimizations - these optimizations are mainly made on 
operations which require large number of CPU cycles to complete, like 
FullyConnected and Convolution. One of the methods to speedup compute-bound 
operations is to lower computation precision - this type of optimization is 
called quantization.
   ```




-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]


Reply via email to