tensorflow - Scaling performance across multi GPUs -

-

i've been running cifar 10 model in tensorflow tutorials train across multiple gpus.

source: https://github.com/tensorflow/tensorflow/blob/r0.8/tensorflow/models/image/cifar10/cifar10_multi_gpu_train.py

gpus: nvidia 8 * m40

configuration: tensorflow 0.8.0 , cuda 7.5, cudnn 4

the result of training performance not scalable expected. pattern of graph looks amdahl's law.

the chart of scaling performance across multi gpus

is normal? if is, see main cause of that?

try gpu utilization high can (>80%). data disk not delivered fast enough, gpus idle of time.


Comments

Post a Comment

Popular posts from this blog

html - Styling progress bar with inline style -

-
i trying set progress bar color red using inline style. need in way because have access <body> part of document. style need apply: progress { color: red } progress::-moz-progress-bar { background: red;} progress::-webkit-progress-value { background: red; } progress[aria-valuenow]:before { background: red; } i tried <progress style="background: red;"> incorrect effect, progress bar color turns blue red background. guide right direction desired style? thank in advance. here live example: https://jsfiddle.net/z36d3f6l/
Read more

java - Oracle Sql developer error: could not install some modules -

-
i have downloaded fresh copy of sqldeveloper ( sqldeveloper-4.1.3.20.78-no-jre ) oracle website. when unzip , lauch gui, keep getting following error message: warning - not install modules: dynamic module config - no module providing capability org.netbeans.netbinox found. my system: windows 7 32 bit hp probook. jdk 1.6 , 1.7, 1.8 available in java/jdk subdirectory. netbeans 8.0.2 installed finally got way out after dirty hours. extracted sqldeveloper zip in nested directory. i.e: e://a/b/c/sqldeveloper. starting sqldeveloper gui above directory threw erros. solution: go in c:\users**username**\appdata\roaming : delete sql developer , sqldeveloper folders extract sql developer zip in new non nested directory. i.e: e:// start , should work now reference: https://community.oracle.com/thread/3870680?start=0&tstart=0
Read more

How to use autoclose brackets in Jupyter notebook? -

-
this may sound silly question, how make use of autoclose brackets in jupyter notebook? example, when type print( jupyter notebook auto-closes brakets print() and places cursor inside. type argument, say print(1 + 1) now cursor between second 1 , right bracket ) . key navigate right of ) ? of course can press end or → or ) achieve this, not seem save time, purpose of brackets autoclose suppose? thanks in advance.
Read more