Google Colab V100 +TensorFlow1.15.2 性能测试

为了对比滴滴云内测版NVIDIA A100，跑了一下Google Colab V100 的 TensorFlow基准测试，现在把结果记录一下！

运行环境

平台为：Google Colab

系统为：Ubuntu 18.04

显卡为：V100-SXM2-16GB

Python版本： 3.6

TensorFlow版本：1.15.2

显卡相关：

测试方法

TensorFlow benchmarks测试方法：

https://github.com/tensorflow/benchmarks

ResNet50_v1.5 BS64

!python tf_cnn_benchmarks.py --num_gpus=1 --batch_size=64 --model=resnet50_v1.5

Step Img/sec total_loss
1   images/sec: 349.6 +/- 0.0 (jitter = 0.0)  7.848
10  images/sec: 349.9 +/- 0.2 (jitter = 0.4)  8.053
20  images/sec: 349.9 +/- 0.1 (jitter = 0.6)  8.103
30  images/sec: 350.2 +/- 0.1 (jitter = 0.6)  8.118
40  images/sec: 350.2 +/- 0.1 (jitter = 0.8)  7.894
50  images/sec: 350.3 +/- 0.1 (jitter = 0.8)  7.918
60  images/sec: 350.1 +/- 0.1 (jitter = 0.7)  8.103
70  images/sec: 350.0 +/- 0.1 (jitter = 0.8)  7.986
80  images/sec: 350.0 +/- 0.1 (jitter = 0.8)  7.808
90  images/sec: 350.0 +/- 0.1 (jitter = 0.8)  7.972
100 images/sec: 350.0 +/- 0.1 (jitter = 0.9)  7.649
----------------------------------------------------------------
total images/sec: 349.78
----------------------------------------------------------------

Resnet50 BS64

python tf_cnn_benchmarks.py --num_gpus=1 --batch_size=64 --model=resnet50

Step Img/sec total_loss
1   images/sec: 386.2 +/- 0.0 (jitter = 0.0)  8.220
10  images/sec: 384.8 +/- 0.4 (jitter = 0.7)  7.880
20  images/sec: 385.5 +/- 0.5 (jitter = 2.2)  7.910
30  images/sec: 385.7 +/- 0.4 (jitter = 2.6)  7.821
40  images/sec: 386.0 +/- 0.4 (jitter = 2.3)  8.004
50  images/sec: 386.2 +/- 0.3 (jitter = 2.4)  7.768
60  images/sec: 386.3 +/- 0.3 (jitter = 2.4)  8.118
70  images/sec: 386.1 +/- 0.3 (jitter = 2.5)  7.816
80  images/sec: 386.3 +/- 0.2 (jitter = 2.4)  7.977
90  images/sec: 386.2 +/- 0.2 (jitter = 2.5)  8.098
100 images/sec: 386.3 +/- 0.2 (jitter = 2.4)  8.045
----------------------------------------------------------------
total images/sec: 386.06
----------------------------------------------------------------

--use_fp16

python tf_cnn_benchmarks.py --num_gpus=1 --batch_size=64 --model=resnet50 --use_fp16

Step Img/sec total_loss
1   images/sec: 911.0 +/- 0.0 (jitter = 0.0)  8.103
10  images/sec: 918.1 +/- 1.2 (jitter = 3.1)  7.756
20  images/sec: 914.3 +/- 2.3 (jitter = 4.3)  7.915
30  images/sec: 914.2 +/- 2.2 (jitter = 4.2)  7.769
40  images/sec: 912.8 +/- 1.7 (jitter = 6.5)  7.915
50  images/sec: 911.7 +/- 1.5 (jitter = 7.3)  7.888
60  images/sec: 912.9 +/- 1.3 (jitter = 7.0)  7.707
70  images/sec: 911.8 +/- 1.2 (jitter = 7.6)  8.011
80  images/sec: 912.3 +/- 1.1 (jitter = 7.3)  7.779
90  images/sec: 912.9 +/- 1.0 (jitter = 6.9)  7.805
100 images/sec: 913.1 +/- 0.9 (jitter = 6.8)  8.034
----------------------------------------------------------------
total images/sec: 912.08
----------------------------------------------------------------

AlexNet BS512

python tf_cnn_benchmarks.py --num_gpus=1 --batch_size=512 --model=alexnet

Step Img/sec total_loss
1   images/sec: 4824.0 +/- 0.0 (jitter = 0.0) nan
10  images/sec: 4804.0 +/- 5.9 (jitter = 23.3)    nan
20  images/sec: 4802.3 +/- 4.3 (jitter = 24.4)    nan
30  images/sec: 4801.7 +/- 4.4 (jitter = 24.0)    nan
40  images/sec: 4804.5 +/- 3.9 (jitter = 23.0)    nan
50  images/sec: 4805.4 +/- 4.0 (jitter = 24.4)    nan
60  images/sec: 4806.7 +/- 3.5 (jitter = 24.8)    nan
70  images/sec: 4810.1 +/- 3.4 (jitter = 24.4)    nan
80  images/sec: 4810.0 +/- 3.1 (jitter = 25.7)    nan
90  images/sec: 4810.9 +/- 2.8 (jitter = 23.4)    nan
100 images/sec: 4811.5 +/- 2.7 (jitter = 23.4)    nan
----------------------------------------------------------------
total images/sec: 4808.18
----------------------------------------------------------------

Inception v3 BS64

python tf_cnn_benchmarks.py --num_gpus=1 --batch_size=64 --model=inception3

Step Img/sec total_loss
1   images/sec: 255.3 +/- 0.0 (jitter = 0.0)  7.277
10  images/sec: 254.3 +/- 0.5 (jitter = 2.2)  7.304
20  images/sec: 254.4 +/- 0.3 (jitter = 2.4)  7.292
30  images/sec: 254.3 +/- 0.3 (jitter = 2.3)  7.402
40  images/sec: 254.2 +/- 0.3 (jitter = 2.3)  7.314
50  images/sec: 254.3 +/- 0.2 (jitter = 2.3)  7.283
60  images/sec: 254.3 +/- 0.2 (jitter = 2.2)  7.363
70  images/sec: 254.3 +/- 0.2 (jitter = 2.1)  7.350
80  images/sec: 254.3 +/- 0.2 (jitter = 2.2)  7.384
90  images/sec: 254.3 +/- 0.2 (jitter = 1.9)  7.318
100 images/sec: 254.3 +/- 0.1 (jitter = 1.9)  7.376
----------------------------------------------------------------
total images/sec: 254.19
----------------------------------------------------------------

VGG16 BS64

python tf_cnn_benchmarks.py --num_gpus=1 --batch_size=64 --model=vgg16

Step Img/sec total_loss
1   images/sec: 250.0 +/- 0.0 (jitter = 0.0)  7.319
10  images/sec: 250.2 +/- 0.2 (jitter = 0.2)  7.297
20  images/sec: 250.4 +/- 0.1 (jitter = 0.5)  7.284
30  images/sec: 250.4 +/- 0.1 (jitter = 0.6)  7.274
40  images/sec: 250.4 +/- 0.1 (jitter = 0.6)  7.288
50  images/sec: 250.4 +/- 0.1 (jitter = 0.6)  7.278
60  images/sec: 250.3 +/- 0.1 (jitter = 0.6)  7.278
70  images/sec: 250.3 +/- 0.1 (jitter = 0.6)  7.266
80  images/sec: 250.3 +/- 0.1 (jitter = 0.6)  7.288
90  images/sec: 250.2 +/- 0.1 (jitter = 0.6)  7.269
100 images/sec: 250.3 +/- 0.1 (jitter = 0.6)  7.270
----------------------------------------------------------------
total images/sec: 250.19
----------------------------------------------------------------

GoogLeNet BS128

python tf_cnn_benchmarks.py --num_gpus=1 --batch_size=128 --model=googlenet

Step Img/sec total_loss
1   images/sec: 1034.6 +/- 0.0 (jitter = 0.0) 7.105
10  images/sec: 1034.2 +/- 0.9 (jitter = 1.8) 7.105
20  images/sec: 1030.9 +/- 1.8 (jitter = 2.9) 7.094
30  images/sec: 1031.0 +/- 1.3 (jitter = 4.2) 7.086
40  images/sec: 1031.6 +/- 1.0 (jitter = 3.9) 7.067
50  images/sec: 1030.6 +/- 0.9 (jitter = 5.4) 7.093
60  images/sec: 1030.4 +/- 0.8 (jitter = 5.4) 7.050
70  images/sec: 1030.6 +/- 0.8 (jitter = 5.7) 7.073
80  images/sec: 1030.3 +/- 0.7 (jitter = 5.9) 7.078
90  images/sec: 1030.3 +/- 0.6 (jitter = 5.6) 7.078
100 images/sec: 1030.0 +/- 0.6 (jitter = 5.5) 7.069
----------------------------------------------------------------
total images/sec: 1029.42
----------------------------------------------------------------

ResNet152 BS32

python tf_cnn_benchmarks.py --num_gpus=1 --batch_size=32 --model=resnet152

Step Img/sec total_loss
1   images/sec: 137.0 +/- 0.0 (jitter = 0.0)  9.023
10  images/sec: 138.0 +/- 0.4 (jitter = 1.4)  8.574
20  images/sec: 138.5 +/- 0.3 (jitter = 1.6)  8.600
30  images/sec: 138.5 +/- 0.2 (jitter = 1.6)  8.755
40  images/sec: 138.6 +/- 0.2 (jitter = 1.6)  8.624
50  images/sec: 138.5 +/- 0.2 (jitter = 1.6)  8.801
60  images/sec: 138.4 +/- 0.1 (jitter = 1.7)  8.679
70  images/sec: 138.4 +/- 0.1 (jitter = 1.8)  9.112
80  images/sec: 138.4 +/- 0.1 (jitter = 1.7)  8.872
90  images/sec: 138.4 +/- 0.1 (jitter = 1.7)  9.025
100 images/sec: 138.4 +/- 0.1 (jitter = 1.7)  8.847
----------------------------------------------------------------
total images/sec: 138.39
----------------------------------------------------------------

性能对比

A100 和V100 和 2080ti 性能对比：

https://www.tonyisstark.com/383.html

Google Colab V100 +TensorFlow1.15.2 性能测试相关推荐

Google Colab 免费GPU服务器使用教程挂载云端硬盘
一.前言二.Google Colab特征三.开始使用 3.1在谷歌云盘上创建文件夹 3.2创建Colaboratory 3.3创建完成四.设置GPU运行五.运行.py文件 5.1安装必要库 5 ...
20 种小技巧，玩转 Google Colab
选自 | amitness.com 作者 | Amit Chaudhary 转自 | 机器之心编辑 | 陈萍 Google Colab 给广大的 AI 开发者提供了免费的 GPU,你可以在上面轻松地 ...
谷歌 colab_如何在Google Colab上使用熊猫分析
谷歌 colab Recently, pandas have come up with an amazing open-source library called pandas-profiling. ...
Google Colab——谷歌免费GPU使用教程
Google Colab简介 Google Colaboratory是谷歌开放的一款研究工具,主要用于机器学习的开发和研究.这款工具现在可以免费使用.Google Colab最大的好处是给广大的AI开 ...
20种小技巧，玩转Google Colab
本文转载自机器之心. 选自amitness.com 作者:Amit Chaudhary 机器之心编译编辑:陈萍 Google Colab 给广大的 AI 开发者提供了免费的 GPU,你可以在上面轻松 ...
【亲测】跑深度学习模型：笔记本的RTX3060 6G vs Google colab免费GPU 速度比较
简单测评笔记本:thinkbook16p RTX3060标配模型是FCN,跑的小数据集, 用的tensorflow 同样的数据和模型参数,我们来看看在两块GPU上的表现如何吧: 1.本地 RTX3 ...
Google Colab大起底：与支持GPU的笔记本相比，谁更胜一筹？
全文共1463字,预计学习时长7分钟图源:unsplash 除非还生活在远古时期,否则你肯定至少听说过GoogleColaboratory 的名号.这是基于云的电脑运行环境,有了它就可以在Googl ...
Google Colab——用谷歌免费GPU跑你的深度学习代码
Google Colab简介 Google Colaboratory是谷歌开放的一款研究工具,主要用于机器学习的开发和研究.这款工具现在可以免费使用,但是不是永久免费暂时还不确定.Google Col ...
Google Colab使用教程
简介 Google Colaboratory是谷歌开放的云服务平台,提供免费的CPU.GPU和TPU服务器. 目前深度学习在图像和文本上的应用越来越多,不断有新的模型.新的算法获得更好的效果,然而,一 ...