TensorFlow Benchmark

LeaderGPU® is a new player in the GPU computing market, intend to change the game rules. At this moment, GPU computing market is represented by several large players, such as AWS, Google Cloud, etc. However, a large player does not always mean the best market offer. The LeaderGPU® project, in comparison to AWS and Google Cloud, provides physical servers, not VPS, where hardware resources can be shared among several dozens of users. The table below compares the cost of processing 500,000 images for the Inception V3 model by different services:

Model GPU Service Number of images Time Price (per minute) Total Cost
Inception V3 8x K80 Google cloud 500000 36m 43sec € 0,0825* € 3,02
Inception V3 8x K80 AWS 500000 36m 14sec € 0,107 € 3,87
Inception V3 8x GTX 1080 LeaderGPU 500000 12m 9sec € 0,10 € 1,22

The table shows that LeaderGPU® is not only 300% faster than its competitors also cost-effective by at least 29%, in comparison to Google Cloud and AWS.

Tests were conducted on the LeaderGPU® computing systems. For the assessment of competitors’, we’ve used the results of tests of Google and AWS instances. Tests were conducted on synthetic data of the following networks models ResNet-50, ResNet-152, VGG16 and AlexNet. At the end of this article you will find the results of tests of other models. The testing of synthetic data was carried out using tf. Variable in analogy with the configuration of models for ImageNet.


LeaderGPU® tests (ltbv20 2x Nvidia® Tesla® P 100)

Testing environment:

  • Instance type:ltbv20
  • GPU:2x NVIDIA® Tesla® P100
  • OS:CentOS 7
  • CUDA / cuDNN:8.0 / 5.1
  • TensorFlow GitHub hash:b1e174e
  • Benchmark GitHub hash:9165a70
  • Date of testing:June 2017
Options InceptionV3 ResNet-50 ResNet-152 Alexnet VGG16
Batch size on GPU 64 64 32 512 32
Optimization sgd sgd sgd sgd sgd

Ltbv 20 Nvidia Tesla P100 (1, 2 GPUs)

Testing synthetic data (images/sec)

GPUs InceptionV3 ResNet-50 ResNet-152 VGG16 Alexnet
1 136.55 217.76 82.05 137.32 2807.64
2 259.14 410.88 150.41 240.61 5117.86

Other results

Testing synthetic data (images/sec)

Batch size alexnet vgg11 vgg16 vgg19 lenet googlenet
32 1516.70 373.45 240.61 203.73 14524.23 714.25
64 2480.30 472.15 274.67 230.73 28599.07 877.76
128 3486.68 540.51 288.80 243.55 44943.19 990.89
256 4440.35 464.69 -* -* 63311.75 1075.38
512 5117.86 -* -* -* 80078.57 1104.74
Batch size overfeat inceptionv3 inceptionv4 resnet50 resnet101 resnet152
32 574.13 233.99 116.45 361.06 214.15 150.41
64 1052.63 259.14 125.09 410.88 245.36 170.79
128 1509.01 269.51 -* 439.41 -* -*
256 2041.60 -* -* -* -* -*
512 2323.77 -* -* -* -* -*

* Available amount of GPU random access memory doesn't allow to launch tests on this size of a packet (batch size).

LeaderGPU® tests (GTX 1080)

Testing environment:

  • Instance type:ltbv17, 14, 16
  • GPU:GTX 1080
  • OS:CentOS 7
  • CUDA / cuDNN:8.0 / 5.1
  • TensorFlow GitHubb1e174e
  • Benchmark GitHub hash:9165a70
  • Date of testing:June 2017
Options InceptionV3 ResNet-50 ResNet-152 Alexnet VGG16
Batch size on GPU 64 64 32 512 32
Optimization sgd sgd sgd sgd sgd

Ltbv 17, 14, 16 GTX 1080 (2, 4, 8 GPUs)

Testing synthetic data (images/sec)

GPUs InceptionV3 VGG16 ResNet-50 ResNet-152 Alexnet
2 187.93 173.2 259.39 109.02 3344.11
4 345.05 276.43 485.92 192.25 6221.67
8 685.59 428.57 949.72 369.02 9405.27

Other results

Testing synthetic data (images/sec)

2x GTX 1080

Batch size alexnet vgg11 vgg16 vgg19 lenet googlenet
32 823.87 223.73 150.50 129.67 14440.58 608.46
64 1517.33 299.24 173.20 149.62 25817.36 676.81
128 2198.87 291.47 -* -* 40910.02 717.52
256 2878.43 -* -* -* 53821.73 730.47
512 3344.11 -* -* -* 66096.43 -*
Batch size overfeat inceptionv3 inceptionv4 resnet50 resnet101 resnet152
32 284.06 180.62 91.63 245.55 154.15 109.02
64 568.15 187.93 -* 259.39 -* -*
128 911.17 -* -* -* -* -*
256 1211.36 -* -* -* -* -*
512 1424.58 -* -* -* -* -*

* Available amount of GPU random access memory doesn't allow to launch tests on this size of a packet (batch size).

4x GTX 1080

Batch size alexnet vgg11 vgg16 vgg19 lenet googlenet
32 1238.14 295.30 272.03 155.75 18389.01 1110.35
64 2375.18 354.55 276.43 169.51 37465.98 1235.77
128 3889.23 321.28 -* -* 60612.34 1365.62
256 5056.10 -* -* -* 89908.56 1394.58
512 6221.67 -* -* -* 114433.39 -*
Batch size overfeat inceptionv3 inceptionv4 resnet50 resnet101 resnet152
32 450.85 328.23 166.82 447.25 276.27 192.25
64 885.37 345.05 -* 485.92 -* -*
128 1576.74 -* -* -* -* -*
256 2126.47 -* -* -* -* -*
512 2447.81 -* -* -* -* -*

* Available amount of GPU random access memory doesn't allow to launch tests on this size of a packet (batch size).

8x GTX 1080

Batch size alexnet vgg11 vgg16 vgg19 lenet googlenet
32 1347.98 381.49 347.37 333.71 27248.65 2023.19
64 2406.83 620.29 428.57 -* 51105.12 2352.15
128 4255.75 -* -* -* 93211.00 2644.26
256 6318.54 -* -* -* 145559.65 2610.21
512 9405.27 -* -* -* 206469.92 -*
Batch size overfeat inceptionv3 inceptionv4 resnet50 resnet101 resnet152
32 555.36 632.23 323.09 857.12 518.57 369.02
64 1042.12 685.59 -* 949.72 -* -*
128 1735.24 -* -* -* -* -*
256 2575.93 -* -* -* -* -*
512 3815.25 -* -* -* -* -*

* Available amount of GPU random access memory doesn't allow to launch tests on this size of a packet (batch size).

LeaderGPU® tests (GTX 1080TI)

Testing environment:

  • Instance type:ltbv21, 18, 36
  • GPU:GTX 1080TI
  • OS:CentOS 7
  • CUDA / cuDNN:8.0 / 5.1
  • TensorFlow GitHubb1e174e
  • Benchmark GitHub hash:9165a70
  • Date of testing:June 2017
Options InceptionV3 ResNet-50 ResNet-152 Alexnet VGG16
Batch size on GPU 64 64 32 512 32
Optimization sgd sgd sgd sgd sgd

Ltbv 21, 18 GTX 1080TI (2, 4, 10 GPUs)

Testing synthetic data (images/sec)

GPUs InceptionV3 VGG16 ResNet-50 ResNet-152 Alexnet
2 264.7 235.15 377.41 127.43 4596.37
4 493.14 401.68 706.95 270.35 8513.54
10 928.26 478.82 1418.60 513.37 -

Other results

Testing synthetic data (images/sec)

2x GTX 1080 TI

Batch size alexnet vgg11 vgg16 vgg19 lenet googlenet
32 880.18 287.25 190.05 169.67 13411.38 807.60
64 1743.20 385.95 235.15 198.28 28360.89 954.35
128 2808.68 457.54 - - 44453.02 1042.77
256 3777.74 - - - 67451.51 1070.28
512 4596.37 - - - 87898.53 -
Batch size overfeat inceptionv3 inceptionv4 resnet50 resnet101 resnet152
32 304.50 250.37 125.81 351.21 218.02 127.43
64 607.91 264.70 - 377.41 236.24 -
128 1162.21 - - 381.62 - -
256 1617.89 - - - - -
512 1992.50 - - - - -

4x GTX 1080 TI

Batch size alexnet vgg11 vgg16 vgg19 lenet googlenet
32 1264.01 378.39 331.08 208.39 19239.51 1487.66
64 2502.01 481.49 401.68 236.07 38818.10 1755.63
128 4539.97 541.39 - - 71457.41 1943.93
256 6787.68 - - - 111721.23 1992.45
512 8513.54 - - - 152549.70 -*
Batch size overfeat inceptionv3 inceptionv4 resnet50 resnet101 resnet152
32 475.69 451.16 228.76 648.11 383.04 270.35
64 942.19 493.14 - 706.95 422.93 -
128 1706.03 - - 722.16 - -
256 2907.18 - - - - -
512 3478.50 - - - - -

10x GTX 1080 TI

Batch size alexnet vgg11 vgg16 vgg19 lenet googlenet
32 865.89 368.50 309.07 289.88 18065.32 2200.48
64 1719.84 667.04 478.82 465.45 36486.24 3333.87
128 3344.45 868.66 - - 70077.18 3771.19
256 6159.03 - - - 138600.70 4335.86
512
- - - 237511.15 -
Batch size overfeat inceptionv3 inceptionv4 resnet50 resnet101 resnet152
32 346.22 809.19 459.10 1116.42 760.83 513.37
64 676.99 928.26 - 1418.60 937.95 -
128 1322.01 - - 1504.64 - -
256 2387.97 - - - - -
512 - - - - - -

AWS EC2 tests (NVIDIA® Tesla® K80)

Testing environment:

  • Instance type:p2.8xlarge
  • GPU:8x NVIDIA® Tesla® K80
  • OS:Ubuntu 16.04 LTS
  • CUDA / cuDNN:8.0 / 5.1
  • TensorFlow GitHubb1e174e
  • Benchmark GitHub hash:9165a70
  • Date of testing:May 2017
Options InceptionV3 ResNet-50 ResNet-152 Alexnet VGG16
Batch size on GPU 64 64 32 512 32
Optimization sgd sgd sgd sgd sgd

p2.8xlarge from AWS (1, 2, 4, 8 GPUs)

Testing synthetic data (images/sec)

GPUs InceptionV3 ResNet-50 ResNet-152 Alexnet VGG16
1 30.8 51.5 19.7 684 36.3
2 58.7 98.0 37.6 1244 69.4
4 117 195 74.9 2479 141
8 230 384 149 4853 260

Other results (images/sec)

GPUs InceptionV3 (batch size 32) ResNet-50 (batch size 32)
1 29.9 49.0
2 57.5 94.1
4 114 184
8 216 355

The test results originated from https://www.tensorflow.org/lite/performance/measurement#details_for_amazon_ec2_nvidia_tesla_k80

Google Compute Engine tests (NVIDIA® Tesla® K80)

Testing environment:

  • Instance type:n1-standard-32-k80x8
  • GPU:8x NVIDIA® Tesla® K80
  • OS:Ubuntu 16.04 LTS
  • CUDA / cuDNN:8.0 / 5.1
  • TensorFlow GitHubb1e174e
  • Benchmark GitHub hash:9165a70
  • Date of testing:May 2017
Options InceptionV3 ResNet-50 ResNet-152 Alexnet VGG16
Batch size on GPU 64 64 32 512 32
Optimization sgd sgd sgd sgd sgd

n1-standard-32-k80x8 from Google (1, 2, 4, 8 GPUs)

Testing synthetic data (images/sec)

GPUs InceptionV3 ResNet-50 ResNet-152 Alexnet VGG16
1 30.5 51.9 20.0 656 35.4
2 57.8 99.0 38.2 1209 64.8
4 116 195 75.8 2328 120
8 227 387 148 4640 234

Other results (images/sec)

GPUs InceptionV3 (batch size 32) ResNet-50 (batch size 32)
1 29.3 49.5
2 55.0 95.4
4 109 183
8 216 362

The test results originated from https://www.tensorflow.org/lite/performance/measurement#details_for_google_compute_engine_nvidia_tesla_k80

Now let’s make an assessment of image processing calculating costs.

In the table below, we will calculate the cost and processing time of 500,000 images using the Inception V3, ResNet-60 and ResNet-152 models, and find the best offer. As can be seen from the table, LeaderGPU® is the most favorable market offer among other considered suppliers.

Model GPU Platform Number of images Time Price (per minute) Total cost
Inception V3 8x K80 Google cloud 500000 36m 43sec € 0,0825* € 3,02*
Inception V3 8x K80 AWS 500000 36m 14sec € 0,107* € 3,87*
Inception V3 8x 1080 LeaderGPU 500000 12m 9sec € 0,10 € 1,22
ResNet-50 8x K80 Google cloud 500000 21m 32sec € 0,0825* € 1,77*
ResNet-50 8x K80 AWS 500000 21m 42 sec € 0,107* € 2,32*
ResNet-50 8x 1080 LeaderGPU 500000 8m 46sec € 0,10 € 0,88
ResNet-152 8x K80 Google cloud 500000 56m 18sec € 0,0825* € 4,64*
ResNet-152 8x K80 AWS 500000 55m 55sec € 0,107* € 5,98*
ResNet-152 8x 1080 LeaderGPU 500000 22m 35sec € 0,10 € 2,26

* The Google cloud service is not provided per minute. Minute cost is calculated based on the hourly price ($ 5,645)