滴滴云A100 40G+TensorFlow1.15.2 +Ubuntu 18.04 性能测试

2024-04-06 13:01•javascript•阅读 443

今天拿到了滴滴云内测版A100，跑了一下 TensorFlow基准测试，现在把结果记录一下！

运行环境

平台为：滴滴云

系统为：Ubuntu 18.04

显卡为：A100-SXM4-40GB

Python版本： 3.6

TensorFlow版本：1.15.2 NV编译版

系统环境：

测试方法

TensorFlow benchmarks测试方法：

https://github.com/tensorflow/benchmarks

resnet50_v1.5

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

Step    Img/sec total_loss


1       images/sec: 602.4 +/- 0.0 (jitter = 0.0)        7.847


10      images/sec: 606.8 +/- 1.2 (jitter = 5.4)        8.053


20      images/sec: 606.3 +/- 0.8 (jitter = 4.4)        8.102


30      images/sec: 605.8 +/- 0.8 (jitter = 3.8)        8.117


40      images/sec: 606.2 +/- 0.7 (jitter = 3.8)        7.893


50      images/sec: 606.1 +/- 0.5 (jitter = 3.0)        7.919


60      images/sec: 606.2 +/- 0.5 (jitter = 2.9)        8.104


70      images/sec: 606.6 +/- 0.5 (jitter = 2.9)        7.985


80      images/sec: 606.6 +/- 0.4 (jitter = 2.8)        7.805


90      images/sec: 606.6 +/- 0.4 (jitter = 2.8)        7.973


100     images/sec: 606.7 +/- 0.4 (jitter = 2.8)        7.644


----------------------------------------------------------------


total images/sec: 606.23


----------------------------------------------------------------

--use_fp16

python tf_cnn_benchmarks.py --num_gpus=1 --batch_size=64 --model=resnet50_v1.5 --use_fp16

Step    Img/sec total_loss


1       images/sec: 1327.1 +/- 0.0 (jitter = 0.0)       7.972


10      images/sec: 1321.2 +/- 5.7 (jitter = 27.6)      7.885


20      images/sec: 1323.5 +/- 4.4 (jitter = 25.9)      8.073


30      images/sec: 1323.6 +/- 3.7 (jitter = 27.3)      7.934


40      images/sec: 1322.1 +/- 3.3 (jitter = 32.9)      8.102


50      images/sec: 1321.4 +/- 3.0 (jitter = 27.7)      7.876


60      images/sec: 1322.2 +/- 2.8 (jitter = 32.3)      7.883


70      images/sec: 1322.3 +/- 2.5 (jitter = 32.6)      7.962


80      images/sec: 1324.0 +/- 2.4 (jitter = 32.2)      8.049


90      images/sec: 1324.2 +/- 2.2 (jitter = 31.2)      7.909


100     images/sec: 1325.1 +/- 2.1 (jitter = 29.6)      7.874


----------------------------------------------------------------


total images/sec: 1322.76


----------------------------------------------------------------

Resnet50 BS64

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

Step    Img/sec total_loss


1       images/sec: 653.5 +/- 0.0 (jitter = 0.0)        8.219


10      images/sec: 646.2 +/- 2.0 (jitter = 6.0)        7.879


20      images/sec: 646.1 +/- 1.4 (jitter = 7.2)        7.909


30      images/sec: 646.0 +/- 1.2 (jitter = 6.0)        7.820


40      images/sec: 646.2 +/- 1.0 (jitter = 6.3)        8.006


50      images/sec: 646.0 +/- 1.0 (jitter = 8.6)        7.769


60      images/sec: 646.0 +/- 0.9 (jitter = 8.6)        8.114


70      images/sec: 645.7 +/- 0.9 (jitter = 9.5)        7.811


80      images/sec: 645.8 +/- 0.8 (jitter = 9.5)        7.979


90      images/sec: 645.8 +/- 0.8 (jitter = 8.0)        8.095


100     images/sec: 645.8 +/- 0.7 (jitter = 6.4)        8.038


----------------------------------------------------------------


total images/sec: 645.26


----------------------------------------------------------------

--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: 1300.1 +/- 0.0 (jitter = 0.0)       8.101


10      images/sec: 1310.1 +/- 7.5 (jitter = 7.4)       7.758


20      images/sec: 1309.7 +/- 8.0 (jitter = 42.3)      7.912


30      images/sec: 1315.0 +/- 5.9 (jitter = 32.1)      7.776


40      images/sec: 1315.5 +/- 4.7 (jitter = 28.2)      7.918


50      images/sec: 1317.5 +/- 3.9 (jitter = 27.7)      7.895


60      images/sec: 1316.5 +/- 3.4 (jitter = 18.6)      7.711


70      images/sec: 1317.3 +/- 3.1 (jitter = 16.1)      8.008


80      images/sec: 1316.9 +/- 2.8 (jitter = 11.4)      7.777


90      images/sec: 1317.7 +/- 2.6 (jitter = 11.8)      7.808


100     images/sec: 1317.1 +/- 2.4 (jitter = 9.9)       8.036


----------------------------------------------------------------


total images/sec: 1315.11


----------------------------------------------------------------

AlexNet BS512

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

Step    Img/sec total_loss


1       images/sec: 8294.2 +/- 0.0 (jitter = 0.0)       nan


10      images/sec: 8290.2 +/- 1.6 (jitter = 5.3)       nan


20      images/sec: 8290.6 +/- 1.0 (jitter = 3.7)       nan


30      images/sec: 8290.8 +/- 0.7 (jitter = 2.8)       nan


40      images/sec: 8291.3 +/- 0.6 (jitter = 2.7)       nan


50      images/sec: 8289.8 +/- 1.4 (jitter = 2.9)       nan


60      images/sec: 8290.2 +/- 1.2 (jitter = 2.9)       nan


70      images/sec: 8290.4 +/- 1.3 (jitter = 3.6)       nan


80      images/sec: 8291.1 +/- 1.1 (jitter = 3.5)       nan


90      images/sec: 8291.9 +/- 1.0 (jitter = 4.4)       nan


100     images/sec: 8291.9 +/- 1.1 (jitter = 5.2)       nan


----------------------------------------------------------------


total images/sec: 8282.46


----------------------------------------------------------------

--use_fp16

python tf_cnn_benchmarks.py --num_gpus=1 --batch_size=512 --model=alexnet --use_fp16

Step    Img/sec total_loss


1       images/sec: 10618.6 +/- 0.0 (jitter = 0.0)      7.250


10      images/sec: 10607.7 +/- 4.4 (jitter = 16.3)     7.251


20      images/sec: 10602.5 +/- 3.0 (jitter = 13.1)     7.251


30      images/sec: 10604.1 +/- 2.3 (jitter = 11.2)     7.251


40      images/sec: 10601.0 +/- 2.5 (jitter = 13.4)     7.251


50      images/sec: 10601.7 +/- 2.5 (jitter = 13.8)     7.251


60      images/sec: 10603.0 +/- 2.2 (jitter = 14.0)     7.250


70      images/sec: 10605.1 +/- 2.1 (jitter = 12.5)     7.251


80      images/sec: 10605.4 +/- 1.9 (jitter = 12.2)     7.251


90      images/sec: 10605.4 +/- 1.7 (jitter = 12.1)     7.251


100     images/sec: 10605.8 +/- 1.7 (jitter = 12.3)     7.251


----------------------------------------------------------------


total images/sec: 10587.67


----------------------------------------------------------------

Inception v3 BS64

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

Step    Img/sec total_loss


1       images/sec: 436.8 +/- 0.0 (jitter = 0.0)        7.276


10      images/sec: 437.9 +/- 1.2 (jitter = 0.8)        7.337


20      images/sec: 437.8 +/- 1.0 (jitter = 2.2)        7.269


30      images/sec: 437.9 +/- 0.8 (jitter = 2.2)        7.422


40      images/sec: 437.9 +/- 0.6 (jitter = 3.5)        7.299


50      images/sec: 438.6 +/- 0.6 (jitter = 4.1)        7.277


60      images/sec: 439.2 +/- 0.5 (jitter = 3.7)        7.363


70      images/sec: 439.5 +/- 0.5 (jitter = 4.8)        7.347


80      images/sec: 440.3 +/- 0.5 (jitter = 5.3)        7.410


90      images/sec: 440.3 +/- 0.5 (jitter = 5.2)        7.325


100     images/sec: 440.3 +/- 0.4 (jitter = 5.0)        7.346


----------------------------------------------------------------


total images/sec: 440.01


----------------------------------------------------------------

--use_fp16

python tf_cnn_benchmarks.py --num_gpus=1 --batch_size=64 --model=inception3 --use_fp16

Step    Img/sec total_loss


1       images/sec: 901.5 +/- 0.0 (jitter = 0.0)        7.305


10      images/sec: 945.5 +/- 7.0 (jitter = 5.0)        7.354


20      images/sec: 945.6 +/- 4.9 (jitter = 7.1)        7.330


30      images/sec: 945.3 +/- 3.9 (jitter = 6.9)        7.382


40      images/sec: 946.3 +/- 3.2 (jitter = 7.3)        7.278


50      images/sec: 946.6 +/- 2.8 (jitter = 7.5)        7.373


60      images/sec: 946.3 +/- 2.5 (jitter = 7.6)        7.299


70      images/sec: 946.8 +/- 2.3 (jitter = 7.5)        7.323


80      images/sec: 946.5 +/- 2.1 (jitter = 7.6)        7.317


90      images/sec: 946.6 +/- 2.0 (jitter = 7.6)        7.357


100     images/sec: 947.2 +/- 1.8 (jitter = 7.3)        7.327


----------------------------------------------------------------


total images/sec: 946.03


----------------------------------------------------------------

VGG16 BS64

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

Step    Img/sec total_loss


1       images/sec: 442.1 +/- 0.0 (jitter = 0.0)        7.321


10      images/sec: 442.4 +/- 0.1 (jitter = 0.4)        7.315


20      images/sec: 442.4 +/- 0.1 (jitter = 0.3)        7.269


30      images/sec: 442.4 +/- 0.0 (jitter = 0.2)        7.271


40      images/sec: 442.4 +/- 0.0 (jitter = 0.2)        7.282


50      images/sec: 442.4 +/- 0.0 (jitter = 0.2)        7.291


60      images/sec: 442.4 +/- 0.0 (jitter = 0.2)        7.250


70      images/sec: 442.4 +/- 0.1 (jitter = 0.2)        7.278


80      images/sec: 442.4 +/- 0.0 (jitter = 0.2)        7.274


90      images/sec: 442.4 +/- 0.0 (jitter = 0.2)        7.286


100     images/sec: 442.4 +/- 0.0 (jitter = 0.2)        7.283


----------------------------------------------------------------


total images/sec: 442.20


----------------------------------------------------------------

--use_fp16

python tf_cnn_benchmarks.py --num_gpus=1 --batch_size=64 --model=vgg16 --use_fp16

Step    Img/sec total_loss


1       images/sec: 687.4 +/- 0.0 (jitter = 0.0)        7.279


10      images/sec: 688.2 +/- 0.2 (jitter = 0.5)        7.255


20      images/sec: 688.0 +/- 0.1 (jitter = 0.5)        7.283


30      images/sec: 688.0 +/- 0.1 (jitter = 0.7)        7.254


40      images/sec: 687.9 +/- 0.1 (jitter = 0.7)        7.283


50      images/sec: 687.8 +/- 0.1 (jitter = 0.7)        7.249


60      images/sec: 687.7 +/- 0.1 (jitter = 0.8)        7.294


70      images/sec: 687.6 +/- 0.1 (jitter = 0.9)        7.278


80      images/sec: 687.6 +/- 0.1 (jitter = 0.9)        7.268


90      images/sec: 687.7 +/- 0.1 (jitter = 0.9)        7.264


100     images/sec: 687.6 +/- 0.1 (jitter = 0.9)        7.268


----------------------------------------------------------------


total images/sec: 687.07


----------------------------------------------------------------

GoogLeNet BS128

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

Step    Img/sec total_loss


1       images/sec: 1577.4 +/- 0.0 (jitter = 0.0)       7.104


10      images/sec: 1565.9 +/- 4.1 (jitter = 12.5)      7.105


20      images/sec: 1561.7 +/- 3.1 (jitter = 20.4)      7.094


30      images/sec: 1562.3 +/- 2.5 (jitter = 15.1)      7.087


40      images/sec: 1561.5 +/- 2.2 (jitter = 16.1)      7.067


50      images/sec: 1561.6 +/- 2.0 (jitter = 15.6)      7.091


60      images/sec: 1561.5 +/- 1.8 (jitter = 15.7)      7.049


70      images/sec: 1560.3 +/- 1.9 (jitter = 15.3)      7.074


80      images/sec: 1558.8 +/- 1.9 (jitter = 17.2)      7.077


90      images/sec: 1558.2 +/- 1.8 (jitter = 17.2)      7.079


100     images/sec: 1557.5 +/- 1.8 (jitter = 17.6)      7.066


----------------------------------------------------------------


total images/sec: 1556.06


----------------------------------------------------------------

--use_fp16

python tf_cnn_benchmarks.py --num_gpus=1 --batch_size=128 --model=googlenet --use_fp16

Step    Img/sec total_loss


1       images/sec: 2690.1 +/- 0.0 (jitter = 0.0)       7.173


10      images/sec: 2675.3 +/- 13.9 (jitter = 35.5)     7.068


20      images/sec: 2682.4 +/- 9.9 (jitter = 55.4)      7.086


30      images/sec: 2686.6 +/- 8.3 (jitter = 36.6)      7.075


40      images/sec: 2687.8 +/- 6.9 (jitter = 30.6)      7.084


50      images/sec: 2686.7 +/- 6.0 (jitter = 36.4)      7.076


60      images/sec: 2687.5 +/- 5.4 (jitter = 36.4)      7.075


70      images/sec: 2681.0 +/- 6.8 (jitter = 41.6)      7.075


80      images/sec: 2683.2 +/- 6.1 (jitter = 34.0)      7.065


90      images/sec: 2684.1 +/- 5.6 (jitter = 35.6)      7.092


100     images/sec: 2683.9 +/- 5.2 (jitter = 36.1)      7.052


----------------------------------------------------------------


total images/sec: 2680.27


----------------------------------------------------------------

ResNet152 BS32

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

Step    Img/sec total_loss


1       images/sec: 225.6 +/- 0.0 (jitter = 0.0)        9.060


10      images/sec: 228.3 +/- 1.0 (jitter = 2.0)        8.594


20      images/sec: 228.3 +/- 0.6 (jitter = 2.0)        8.635


30      images/sec: 228.2 +/- 0.5 (jitter = 2.5)        8.719


40      images/sec: 227.9 +/- 0.5 (jitter = 2.8)        8.599


50      images/sec: 228.1 +/- 0.5 (jitter = 2.9)        8.791


60      images/sec: 228.3 +/- 0.4 (jitter = 3.6)        8.668


70      images/sec: 228.3 +/- 0.4 (jitter = 3.3)        9.072


80      images/sec: 228.3 +/- 0.4 (jitter = 3.5)        8.874


90      images/sec: 228.4 +/- 0.3 (jitter = 3.7)        9.030


100     images/sec: 228.4 +/- 0.3 (jitter = 3.7)        8.839


----------------------------------------------------------------


total images/sec: 228.29


----------------------------------------------------------------

--use_fp16

python tf_cnn_benchmarks.py --num_gpus=1 --batch_size=32 --model=resnet152 --use_fp16

Step    Img/sec total_loss


1       images/sec: 392.9 +/- 0.0 (jitter = 0.0)        9.147


10      images/sec: 397.9 +/- 2.8 (jitter = 6.0)        9.000


20      images/sec: 399.0 +/- 2.1 (jitter = 8.6)        8.842


30      images/sec: 393.7 +/- 2.9 (jitter = 14.7)       8.813


40      images/sec: 394.4 +/- 2.3 (jitter = 15.2)       8.984


50      images/sec: 394.9 +/- 2.0 (jitter = 13.9)       8.647


60      images/sec: 395.7 +/- 1.8 (jitter = 13.9)       8.838


70      images/sec: 396.5 +/- 1.6 (jitter = 15.3)       8.941


80      images/sec: 395.9 +/- 1.4 (jitter = 13.4)       8.913


90      images/sec: 396.2 +/- 1.3 (jitter = 14.1)       8.807


100     images/sec: 395.7 +/- 1.3 (jitter = 14.5)       8.729


----------------------------------------------------------------


total images/sec: 395.34


----------------------------------------------------------------

性能对比

A100 和V100 和 2080ti 性能对比：

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

上一篇 »从源码安装tensorflow1.8-gpu版本
下一篇 »Tensorflow 又要升级了。Nvidia驱动升级，牵一发而动全身

滴滴云A100 40G+TensorFlow1.15.2 +Ubuntu 18.04 性能测试

运行环境

测试方法

resnet50_v1.5

Resnet50 BS64

AlexNet BS512

Inception v3 BS64

VGG16 BS64

GoogLeNet BS128

ResNet152 BS32

性能对比

相关推荐

ubuntu16.04卸载tensorflow0.11版本，安装tensorflow1.1.0版本

Ubuntu使用阿里云软件源

ubuntu16.04安装tensorflow1.3

tensorflow1.x和tensorflow2.x中的tensor转换为字符串的实现

09-01 Tensorflow1基本使用

Ubuntu14.10安装TensorFlow1.0.1

tensorflow1.15-keras 多标签 xception训练与测试

TensorFlow无法载入cu*64_100.dll问题